# brianbest.com — full site content for LLMs This file contains the complete content of Brian Best's personal site in plain text: his career profile, his projects, and every blog post in full. Canonical URLs are included. For interactive questions, his AI assistant lives at https://brianbest.com/chat. # Career Profile: Brian Best ## Overview **Title:** Principal Software Developer **Location:** Kitchener, Ontario, Canada **Current Focus:** Production AI that holds up after the demo ## Professional Summary Principal Software Developer at Axonify — 12+ years full-stack, last 3+ in production AI. I co-lead an internal platform of LLM coding agents and ship LLM features into the customer-facing product (retrieval pipelines, MCP tool-calling, conversational interfaces) under enterprise compliance constraints: credential isolation, audit logging, and evals before rollout. Grew from Developer to Principal in six years. Previously co-founded Phased.io (CAD $75k angel, Propel ICT Build) and came up through radio broadcasting — live, unscripted comms that still shapes how I explain systems. ## What Makes Brian Unique - **Former radio broadcaster and Director of Production at INDI101**: Developed exceptional live communication skills, content creation abilities, and comfort with public speaking - **Student Union Chairman representing 14,000 students with $3.1M budget**: Leadership experience at scale, budget management, stakeholder communication, and governance - **Co-founded Phased.io, raised $75k angel round, completed Propel ICT accelerator**: Entrepreneurial mindset, fundraising experience, startup operations, and product thinking - **Non-traditional path: Radio -> Web Dev -> Senior Dev -> AI/LLM Specialist**: Demonstrates adaptability, continuous learning, and ability to reinvent and grow ## Current AI/LLM Focus Co-leads an internal agentic AI platform of LLM-powered coding agents and ships LLM capabilities into the customer-facing Axonify product. Most visible outcome so far: support automation that runs pre-emptive investigation at ticket creation (log analysis, error extraction, RAG over similar past issues), cutting average time from issue reported to fix deployed from ~14 days to ~3 days. **Key Areas:** - Internal agentic AI platform — LLM agents that implement, test, and review changes against the production codebase - Customer-facing LLM features: retrieval pipelines, secure tool-calling, and conversational interfaces under enterprise compliance - MCP integrations with credential isolation, scoped tool access, and audit logging - Multi-agent orchestration and agent-augmented engineering workflows - Support automation: pre-emptive investigation at ticket creation (logs, error extraction, RAG over past issues) - System-health and observability tooling shared across the engineering org **Philosophy:** Production AI requires enterprise compliance, not just cool demos. Focus on reliability, observability, and real business outcomes. ## Communication & Working Style **Communication Style:** - Clear and direct — years of live, unscripted radio taught me to land a point without slides - Comfortable presenting to mixed audiences: engineers, execs, and customers in the same room - Translate system constraints into plain language so non-technical stakeholders can make real decisions - Default to writing it down — docs and ADRs over hallway knowledge **Working Style:** - Ship early, instrument it, then iterate — I’d rather learn from production than debate in a doc - Review-heavy: if I can’t explain a diff in review, it doesn’t ship - Build the thinnest thing that proves the constraint, then harden it - Comfortable leading a workstream or pairing as an IC — whichever unblocks the team ## Technical Depth Areas of Expertise: Agentic AI, multi-agent orchestration & MCP tooling, RAG, vector search & retrieval pipelines, Full-Stack JavaScript/TypeScript (+ Java, Python), Platform Architecture & Distributed Systems Design, API Design (REST, WebSockets & real-time streaming), Cloud & Infrastructure (GCP, Docker, Linux) My deepest expertise lies in building production AI systems and full-stack applications. Currently focused on agentic AI, multi-agent orchestration, MCP integrations, and RAG/retrieval pipelines. Strong background in the JavaScript/TypeScript ecosystem with Java and Python alongside it, and experience across the stack from React frontends to Node.js services to cloud infrastructure on GCP. ## Work Experience ### Principal Software Developer at Axonify Feb 2026 - Present | Waterloo, ON (Remote) Own technical strategy for AI-driven automation across the Axonify platform — internal agent platform and customer-facing LLM features — with compliance, observability, and rollout as first-class concerns. Key Achievements: - Co-lead an internal platform of LLM coding agents that implement, test, and review changes against the production codebase — agents run behind scoped MCP tool access with credential isolation and full audit logging - Ship LLM capabilities into customer-facing learner and admin workflows (retrieval pipelines, tool-calling, conversational UX) gated by evals and compliance review - Architect MCP integrations so agents can query codebase and internal tooling without broad credential exposure — least-privilege tool scopes, per-tool approval, structured audit trails - Shipped pre-emptive support investigation at ticket creation — log analysis, error extraction, reproduction-context capture, and RAG over similar past issues (durable fix vs. workaround) — cutting mean time from report to fix deployed from ~14 days to ~3 days - Serve as in-house productizer: translate LLM research into shippable, observable, supportable features rather than demos - Technical owner for system-health and observability tooling used across the engineering org Technologies: LLM/AI, Agentic AI, MCP, RAG, Vector Search, TypeScript, Node.js, Java, Python, GCP, SQL ### Senior Software Developer at Axonify Feb 2022 - Feb 2026 | Waterloo, ON (Remote) Promoted to technical lead. Built the foundational retrieval and context-gathering infrastructure that now powers the platform's LLM agent capabilities, and stood up the first generation of AI-assisted support tooling. Key Achievements: - Designed and built the foundational retrieval and context-gathering infrastructure that now powers the platform's LLM agent capabilities - Stood up the first generation of AI-assisted customer-support tooling; validated cost, latency, and accuracy tradeoffs across multiple model providers before committing to a production architecture - Mentored intermediate and junior developers, contributed to hiring rubrics, and ran cross-team technical reviews - Delivered platform features across Java/JavaScript/SQL on Google Cloud Platform serving millions of enterprise learners across regulated industries (healthcare, retail, food service) Technologies: LLM/AI, RAG, Vector Search, TypeScript, Node.js, Java, GCP, SQL ### Software Developer → Intermediate Software Developer at Axonify Mar 2020 - Feb 2022 | Waterloo, ON (Remote) Shipped customer-facing features across the Axonify learning platform on Google Cloud Platform, growing into systems-level ownership and earning successive promotions. Key Achievements: - Shipped customer-facing features across the Axonify learning platform on Google Cloud Platform - Grew into systems-level ownership, earning successive promotions - Onboarded and ramped during the pandemic remote transition Technologies: TypeScript, React, Node.js, Java, GCP, SQL ### Frontend Developer at Combinaut Jul 2017 - Apr 2020 | Ontario, Canada (US healthcare client) Built marketing platforms and customer-facing web frontends for a US hospital client in a regulated healthcare environment. Key Achievements: - Built marketing platforms and customer-facing web frontends for a US hospital client, operating within a regulated healthcare environment (content, patient-facing communications, accessibility, and compliance) - Collaborated with US-based stakeholders to ship features across patient-facing and administrative surfaces - Balanced marketing-team requirements against the regulatory constraints of healthcare communications Technologies: JavaScript, Component-driven frontend, SQL, Document stores ### Software Engineer at Ubique Networks Aug 2016 - Jul 2017 | Ontario, Canada Led a web development team building an Angular + Node.js application backed by Express and WebSockets. Key Achievements: - Led the web development team building an Angular + Node.js application backed by Express and WebSockets - Ran user-feedback sessions and translated manager-driven requirements into shipped features - Established development practices for the team Technologies: Angular, Node.js, Express, WebSockets, JavaScript ### Co-Founder & CTO at Phased.io Sep 2014 - Aug 2016 | Sydney, Nova Scotia Co-founded an early-stage venture, raised a CAD $75k angel round, and was selected for Propel ICT's Build accelerator. Built the product on a MEAN stack with Firebase real-time data plus iOS/Android, and managed a remote team. Key Achievements: - Co-founded an early-stage venture and raised a CAD $75,000 angel round - Selected for Propel ICT's New Brunswick-based Build accelerator cohort - Built the product on a MEAN stack with Firebase real-time data, plus iOS/Android via PhoneGap and integrations including Twilio - Directly managed a remote team of designers and developers; introduced lean product management and analytics practices to drive iteration Technologies: MEAN Stack, MongoDB, Express, Angular, Node.js, Firebase, PhoneGap, Twilio, iOS, Android ## Skills Assessment ### Strong Competencies - MCP (Model Context Protocol) (AI/LLM) - LLMOps (AI/LLM) - LLM Engineering (AI/LLM) - Agentic AI (AI/LLM) - RAG (AI/LLM) - Vector Search (AI/LLM) - Conversational AI (AI/LLM) - JavaScript/TypeScript (Languages) - Java (Languages) - SQL (Databases) - Node.js (Backend) - React (Frontend) - Google Cloud Platform (Cloud) - Docker & Docker Compose (Cloud & Infra) - System Design (Architecture) - Technical Leadership (Soft Skills) - Communication (Soft Skills) ### Proficient Skills - Python (Languages) - Vue.js (Frontend) - Angular (Frontend) - Linux (Arch, Ubuntu) (Cloud & Infra) - AWS (Cloud) - MongoDB (Databases) - GraphQL (API) ## Education - Ontario College Graduate Certificate, Interactive Multimedia, Sheridan College (2012-2013) - Graduate With Honors, Valedictorian - UIT Startup Immersion, Cape Breton University (2014-2015) - Ontario College Diploma, Radio Broadcasting, Mohawk College (2008-2010) ## Awards & Recognition - Graduate With Honors (Sheridan College) - Valedictorian (Sheridan College) - SSUI Values Award (Sheridan Student Union) - Sheridan Student Union Chairman (Governance of a $3.1M budget and a 14,000-student constituency) # Projects === # Project: MCP Server for Email Workflows URL: https://brianbest.com/projects Local / private Tags: MCP, TypeScript, IMAP/SMTP, AI Agents Featured: yes A Model Context Protocol server that gives AI agents first-class email — read and send over IMAP/SMTP, behind Proton Mail Bridge so the agent never touches raw credentials. A Model Context Protocol (MCP) server that gives AI agents first-class email — reading and sending mail autonomously over IMAP/SMTP. ## Why this instead of an API wrapper Most email integrations for agents are one-off function definitions. This one is an MCP server, so any MCP-compatible client can use it without custom wiring. The agent doesn't need to know it's email — it just sees tools like `search_mail`, `read_message`, and `send_message`. ## Stack & security `imapflow` and `nodemailer` for the mail transport, `@modelcontextprotocol/sdk` to expose the tools, and **Proton Mail Bridge** as the transport layer — the agent talks to an end-to-end-encrypted mailbox over a local bridge rather than the open internet. ```mermaid flowchart LR A[Agent] -->|MCP tools| B[MCP Email Server] B -->|IMAP/SMTP| C[Proton Bridge] C -->|encrypted| D[Mailbox] ``` Credentials never leave the bridge. The agent sees scoped tools, not raw IMAP secrets. ## What it handles - Search and read with filters (folder, date range, unread) - Send with proper threading (`In-Reply-To`) - Attachment awareness without dumping binary into context - Scope control — the server only exposes the tools you configure Private repo — the `Proton Bridge → imapflow` pairing is the part I'd extract if I open-sourced it. === # Project: Self-Hosted Agentic AI Deployment URL: https://brianbest.com/projects Local / private Tags: Agentic AI, Docker Compose, Caddy, Self-Hosted, Linux Featured: yes Two isolated instances of an open-source agentic AI system on one home server — three Docker Compose stacks (shared Caddy + one per instance) with separate credentials, volumes, and networks. Two isolated instances of an open-source agentic AI system running on a single always-on home server. I use it as a personal sandbox for validating multi-tenant agent isolation — what it takes to run several autonomous agents on one box without letting them leak into each other's state. ## Architecture Three Docker Compose stacks: one shared Caddy reverse proxy in front, and a fully isolated stack per instance. ```mermaid flowchart TB U[Internet] --> C[Caddy — shared] C --> A[Instance A] C --> B[Instance B] A --- VA[(volume A)] B --- VB[(volume B)] A -.- NA[network A] B -.- NB[network B] ``` Each instance gets its own credentials, volumes, and Docker network. Agents can't reach each other's state — intentionally. That isolation is the point: it mirrors the credential-scoping and audit-logging I care about in production AI. ## Why self-hosted Cloud agents hide the isolation problem. On one box, you feel it — shared ports, overlapping volumes, credential leakage. Solving it with `Caddy + Compose + network isolation` is cheap and reproducible, and it stress-tests the same patterns (scoped access, attributable actions) that matter inside enterprise compliance boundaries. === # Project: Agentic Coding Toolkit URL: https://brianbest.com/projects Local / private Tags: Agentic AI, Developer Tools, Slash Commands, LLMOps Featured: no Slash commands that make agentic-coding sessions faster and more reliable — a silent health check, an iterative prompt loop backed by state files, and an eval-gated prompt improver. A personal toolkit of slash commands for agentic-coding sessions — three commands I actually run daily: - **`/health` — silent health check.** Stays quiet unless something is wrong. Cheap enough to run on every session start without adding noise when things are green. ```text $ /health # all green — no output (exit 0) $ /health # when something is off: ! branch is 12 commits behind main ! 3 tests failing in lib/ai/validation.test.ts ! .env.local missing OPENAI_API_KEY ``` - **`/loop` — iterative prompt looping.** Re-runs a prompt until a condition is met, backed by state files and a bash stop hook. Useful for "keep polishing this until the eval passes." ```text $ /loop "improve the error handling in lib/ai/validation.ts" → run 1: edits applied, eval 72% — continuing → run 2: edits applied, eval 84% — continuing → run 3: eval 91% — threshold met, stopping ``` - **`/improve-prompt` — eval-gated prompt improvement.** Improves a prompt only when the change clears an eval bar, following Karpathy's autoresearch methodology — no change ships unless the numbers move. ```mermaid flowchart LR P[prompt] --> E[eval] E -->|pass| KEEP[keep change] E -->|fail| DROP[drop change] KEEP --> P ``` Each command is a standalone slash command with its own state and tests. Private repo for now — the health check and loop are the most reusable if I extract them. # Blog Posts === # Flagship Models Went Quiet This Week URL: https://brianbest.com/blog/2026-08-17-ai-week-w01a-models Date: 2026-08-17 Tags: ai-models, benchmarks, weekly-roundup Summary: I went looking for the next big model drop for the week of Aug 4-10 and found a quiet flagship front — one safety system shipped, no benchmark leaps. I went looking for a flagship model story this week and didn't find one. That's the story. For the week of August 4-10, I checked the usual places — OpenAI, Anthropic, Google, Meta, xAI, Mistral, DeepSeek — for a new foundation model, a benchmark jump, or a pricing change. On the flagship front, it was mostly crickets. Here's what actually shipped. ### What shipped: Shieldstral The only verified release inside the window was Mistral's Shieldstral on August 4th. It's not a new base model. It's safety infra — prompt and skills guardrails with a record system around it. Mistral's been shipping a lot of capability lately (OCR 4 in late June, work on remote agents in May), but this week they shipped wrapping, not weights. And honestly, that's telling. When labs aren't pushing a parameter bump, they tend to push the stuff that makes the last parameter bump usable in prod. Shieldstral fits that pattern. ### What just missed the window DeepSeek put DeepSeek-V4-Flash into public beta on July 31st, four days before this window opened. The API name is `deepseek-v4-flash` (now `DeepSeek-V4-Flash-0731`), same calling convention as before, with claims of stronger agent performance over V4-Pro-Preview. DeepSeek says V4-Pro proper is coming soon. I mention it because if you're benchmarking right now, that July 31 cut is the freshest thing to test against — not something from this week. ### What didn't show up No verified flagship model, benchmark, or pricing change from OpenAI, Anthropic, Google, Meta, or xAI that I could pull with a clean date inside Aug 4-10. That doesn't mean nothing happened on X or in a private changelog. It means nothing hit the official newsrooms or docs with a verifiable date I could confirm — and I tried. OpenAI, Anthropic, and Meta's news pages are all behind Cloudflare challenges or JS-only hydration right now, so a simple fetch doesn't even return the article list. No new MMLU, SWE-Bench, or GPQA numbers showed up in the places I could actually read. I could've filled this post with rumors about GPT-5 point releases. My own repo's implementation plan already references `gpt-5.2` as a future target. But a plan doc isn't a ship, and I'm not going to cite it as one. ### Why a quiet week matters We've gotten used to a model drop every couple weeks, so a quiet week feels like a lull. I don't think it is. My read: labs are in the eval-and-harden phase, not the drop phase. You ship a big model, then you spend the next few weeks shipping the safety layer, the harness, the eval repro — the stuff that decides whether anyone actually adopts it. Shieldstral is that kind of ship. DeepSeek's harness note (minimal mode, temp 1.0, top-p 0.95 for their Code Agent tests) is too. If you're deciding what to eval this week, don't chase a new flagship. Re-run your own tasks against V4-Flash-0731 and whatever you're running in prod, and log the cost per completed task, not just the benchmark score. The next flagship will move the headlines. The infra shipping now will move your error rate. I'll keep watching the pricing and context window pages — those often change without a blog post. If something drops there, I'll call it out next week. === # This Week in AI Products: Tiering, Cyber Gates, and 100x Cheaper Retrieval URL: https://brianbest.com/blog/2026-08-17-ai-week-w01b-products Date: 2026-08-17 Tags: ai-weekly, enterprise-ai, product-launches Summary: No new flagship — OpenAI tiered GPT-5.6 by price and risk, Neon and Castform shipped a 100x cheaper retriever, and Anthropic put Claude Code on autopilot. If you were watching for a new flagship last week, you missed the story. Nothing big shipped between August 4 and 10. What shipped instead was tiering — the same GPT-5.6 family sliced by price, speed, and who you're trusted to be. That's more useful than a bigger model. It tells you how the labs want enterprises to actually deploy. ### Luna goes free, Sol gets a slider The main move was August 6: OpenAI made GPT-5.6 Luna the default for Free and Go ChatGPT accounts — unlimited text chats, new Think button for harder questions. Plus and Pro keep an updated Sol with a slider for how much reasoning you want. I couldn't pull the origin post directly — `openai.com` was behind a Cloudflare challenge — but HN indexed it at 17:02 UTC (314 points) and RuntimeWire's syndication at 17:26 UTC confirms the details. It's a routing hint. Luna where you can, Sol where you must. That pairs with the July 30 pricing change still echoing last week: Luna down 80% to $0.20/$1.20 per million tokens, Terra down 20% to $2/$12, Sol flat at $5/$30, plus a new Fast mode for Sol at 2x price for 2.5x speed. Don't burn Sol on retrieval loops. A CNBC headline indexed August 5 also claimed Microsoft made Sol the default in GitHub Copilot for staff. The URL 404s so I can't verify the body — treat it as a directional enterprise signal, not a confirmed rollout. ### Cyber gets its own gated frontier The most explicit enterprise launch landed August 10: GPT-5.6-Cyber inside a new two-tier Daybreak initiative. OpenAI's post and RuntimeWire's syndication both timestamped that afternoon describe a purpose-trained model that answers advanced exploit-development requests the general model refuses — RuntimeWire's headline puts it at 95% completion vs. 1.5% for Sol. Access *is* the product. Daybreak Red requires verification, monitoring, and hardware keys, with a separate tier for broadly defensive work. If you run a security team, that's a real tier to evaluate. If you're OpenAI, it's a bet that "trusted hands" is a defensible distribution model. I'm not sure hardware-key gating scales, but I get why they tried it. ### The 100x cheaper retriever The sleeper wasn't from a lab. On August 5, Neon (now part of Databricks) and Castform showed an RL post-trained 4B open model that matches Sol on retrieval at roughly 100x lower cost. Their math is blunt: a typical multi-turn search with `gpt-5.6-sol` takes >10s and ~$0.03 end-to-end. Small models are already 100x cheaper but lacked retrieval skill. Castform's loop — synthetic data from `lakebase_text` and `lakebase_vector`, rollouts that call Lakebase Search on Neon, same call at inference — closes the gap by training on data already in your Postgres. This is the MCP thesis with economics attached. If agentic retrieval is the workload — plan, search, search again — cheap specialized retrievers beat frontier calls on every turn except the last. ### Auto mode on by default Anthropic flipped the other switch August 9: Claude Code's auto mode on by default, per TechCrunch (`Enterprise` section, 19:20 UTC). "Even less human oversight" is the pull quote. I've got mixed feelings. As I wrote in March, the unit of work is now "a change I will review" — and review velocity is the bottleneck. Auto-by-default optimizes for commit speed and pushes cost to audit. If you turn it on, keep the allowlist discipline — scoped creds, per-tool gates, every call attributable. ### What I'd do Monday 1. **Split workflows by tier.** Luna for triage, Terra for drafting, Sol for hard reasoning. Cyber only under Daybreak Red. 2. **Benchmark the small retriever.** Even 20x cheaper changes your unit economics if you do multi-turn search. 3. **Set your auto-mode policy.** "Auto" without an allowlist is just unreviewed deploys with better marketing. === # Open Source Did the Shipping This Week URL: https://brianbest.com/blog/2026-08-17-ai-week-w01c-opensource Date: 2026-08-17 Tags: open-source, ai-weekly, research Summary: Flagships were quiet Aug 4-10. Open source wasn't — a 30B local agentic model under Apache 2.0, a 2.6B on-device runner, and a wave of agent memory tooling. The flagship labs took the week off. Open source didn't. For Aug 4-10, if you were watching OpenAI/Anthropic/Google for a new base model, there was nothing verifiable. If you were watching Hugging Face and arXiv, it was one of the busiest weeks in months. Labs harden and tier, open source ships the plumbing. ### Muse Glimmer: 30B that fits on one GPU Meta dropped Muse Glimmer on Aug 10 — 30B, multimodal, Apache 2.0. That license is the story. Not a Llama-style commercial carve-out, just Apache 2.0. BF16, GGUF quants, ExecuTorch, plus a DFlash speculative drafter. A 30B model normally wants 55GB at full precision. Meta compresses to ~4-bit (~20GB) and claims 3.1x decode speedup via block-level speculation so it can live inside an agent loop on a single 24GB card or a Mac. As-claimed, but Simon Willison already had it running via LM Studio: ```bash llm -m lmstudio/meta/muse-glimmer -a image.jpg 'describe image' ``` They're targeting DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench — full agentic task completion, not chat vibes. If you run agents with MCP servers, this is the one to benchmark. Apache 2.0 means you can actually ship it. ### LFM2.5-2.6B: forget it's even running Liquid AI's LFM2.5-2.6B landed Aug 4 — 2.69B params, 131K context, text-only. 22 double-gated convolutions + 8 GQA blocks, 30 layers, ~34T tokens, GGUF/MLX/ONNX day one. It's not trying to be Glimmer. Liquid says don't use it for coding or knowledge-heavy work. It's for cheap background agents — invoice parsing, form extraction, robotics — where you want 128K on-device without paying for 30B. Glimmer shrinks the frontier to fit; LFM makes it so small you forget it's running. ### Shieldstral: safety as a prompt Mistral's Shieldstral 1.0 3B (Aug 7, Apache 2.0) flips moderation on its head. No fixed harm taxonomy — you supply policy as a plain question at inference and get a calibrated yes/no score: ```python score = model.classify(text, policy="Does this contain self-harm instructions?") ``` Mistral claims 84.9% avg F1, matching a 20B safeguard model 7x its size. If it holds, it's the efficiency story of the week. The catch is auditability — a score without a taxonomy is harder to explain to a regulator. But per-customer policies without retraining is exactly the right shape. ### The substrate wave Four infra pieces in three days, all about time and memory: - **Shepherd** (Aug 8, MIT, `pip install shepherd-ai`) — Git-like typed traces for agent runs, fork process+filesystem 5x faster than Docker, >95% cache reuse on replay. - **TencentDB Agent Memory v2.0** (Aug 7, MIT) — team memory hub turning chats/docs/code into Chat Memory, Skill, LLM-Wiki, and Code-Graph. Self-hosted via Docker. MCP won the tool boundary. This week was the next one: what happened (Shepherd) and what we remember (Agent Memory). Neither solves auth for you — budget for that. ### Papers worth reading - **CreativeInstruct** (arXiv:2608.07460, Aug 7) — teaching LLMs to balance quality, creativity, and diversity without RLHF flattening everything. Code on GitHub. - **CoinRAG** (arXiv:2608.07458) — KV-cache reuse for long-context RAG via contextualized nuggets. Practical if you're actually running long contexts. - **Evidence-RL** (arXiv:2608.08021, Aug 8) — counterfactual evidence disentanglement inside GRPO for visual reasoning. No inference overhead. ### What I'd do Monday Run Glimmer on your agentic loop with your own MCP servers and measure task completion, not token speed. Put LFM2.5 on a spare device for the boring extraction you keep paying Sol to do. And try Shieldstral with your actual policy wording — it's the first classifier that treats policy as context. Open source had the better week. It usually does when the labs are in harden mode. === # This Week in AI Policy: Labels, Takedowns, and a $1.5B Coding Agent Chase URL: https://brianbest.com/blog/2026-08-17-ai-week-w01d-policy Date: 2026-08-17 Tags: ai-policy, regulation, industry-moves Summary: EU transparency rules went live, India set a 3-hour takedown clock, California launched an AI cyber program, and Google chased a $1.5B coding agent deal — a busy week for AI regulation and industry moves. If you still think AI regulation moves slowly, you weren't watching last week. Between August 4 and 10, the EU, India, China, California, and Texas all made concrete moves — while labs and incumbents reshuffled money and people to keep up. No single headline dominated, but together they mark a shift: voluntary AI commitments are giving way to enforceable rules. ### Labels are now law The EU AI Act's transparency provisions took effect this week. Providers and deployers now have to add visible and machine-readable disclosure for AI-generated or manipulated content — text, images, audio, video, deepfakes included. Firstpost flagged it August 4, Global News followed August 6. It's the first user-visible bite of the AI Act, not a risk-category whitepaper. India went further and faster. On August 7, WION and TechGig reported that platforms must now remove illegal AI-generated content within 3 hours of notice and label synthetic content. That's stricter than the EU — label *and* remove, quickly. MeitY also stood up a new expert committee for AI Governance (AIGEG) on August 5, so this looks like permanent scaffolding, not a one-off rule. As Live Law noted August 6, courts are also shaping policy by injunction, which leaves platforms navigating overlapping orders. China tightened the screws on a category the West has barely regulated: AI companions. AP News reported August 10 that new guardrails on content, data handling, and youth protections left users finding companions lobotomized or removed. Beijing is treating "emotional AI" as its own regulatory bucket. ### States fill the federal vacuum In the US, states moved while DC debated. Gov. Gavin Newsom on August 10 directed California agencies to stand up a first-in-the-nation AI Cyber Defense Program inside the California Cybersecurity Integration Center, aimed at protecting power, water, transit, and hospitals from AI-enabled cyber threats. I verified the Unite.AI report directly — it's timestamped August 10, 11:18am ET. Texas moved on the other end of the stack. BeInCrypto reported August 9 that Texas set five new rules for AI data centers covering energy, water, and permitting. If you're planning capacity there, your power and water assumptions just changed. In the background, EFF, CDT and state AGs warned August 4 that proposed FTC policy could create liability for safety tuning — putting companies in the odd position of being asked to make models safer while being told the way they do it could be held against them. ### Labs staff up, incumbents pay up Anthropic created its first global affairs chief role this week and filled it with an ex-judge, per StartupHub (Aug 4) and NewsBytes (Aug 5). When a frontier lab hires a judge for policy, that's litigation prep, not optics. The money moved just as fast. Business Insider reported August 5 that Google is in talks for a $1.5B+ talent-and-tech deal for Mechanize, an AI coding agent startup that went from a $9M seed to billion-dollar talks in about 100 days per Tech Funding News. Not closed — keep the "in talks" qualifier — but the price shows how hot the coding agent race is. OpenAI quietly folded NextSlide's team into ChatGPT the same week to boost slides and presentation work. And SAP showed the capex squeeze directly: People Matters reported August 10 that SAP froze most hiring and internal travel to fund AI expansion. An incumbent explicitly trading headcount for AI runway. ### What I'm watching 1. **Specs, not headlines.** Watch EUR-Lex Article 50 guidance for what counts as a watermark and the actual MeitY gazette for what triggers India's 3-hour clock. 2. **State divergence.** California securing infrastructure *with* AI while Texas constrains infrastructure *for* AI means you're now compiling for two regulatory targets if you deploy nationally. 3. **That Mechanize number.** If $1.5B for a 100-day-old team closes, acqui-hire pricing resets for everyone. Back next week with models and product. This week belonged to the lawyers and the checkbooks. === # This Week in AI Agents: The Router Won the Week URL: https://brianbest.com/blog/2026-08-17-ai-week-w01e-agents Date: 2026-08-17 Tags: ai-agents, coding-tools, mcp, devtools Summary: No flagship model shipped — instead we got routers that cut costs 68%, MCP growing up with a Ruby SDK and an Agents working group, and coding tools turning into model supermarkets. No flagship model shipped between August 4 and 10. What shipped instead was the plumbing that decides which model you hit, where your agent runs, and whether your tools actually compose. That plumbing matters more than the next benchmark bump. ### Routing beat raw intelligence Cursor published how its Router works on August 6, and the numbers are worth reading twice. Auto Intelligence now delivers above Fable-level satisfaction at 68% lower cost — another 18% drop since Router launched July 22. Auto Balance beats Opus 4.8 at 41% lower cost. The trick isn't a better model. It's Compass, a classifier that predicts task complexity from real production traffic, plus a learned map of which models are strong at what. Labs slice model families by price and speed; Cursor slices routing by task. If you're building agents, your margin lives in the router now, not the model call. I run a similar pattern with per-task tool allowlists — same idea at a different layer. ### MCP grew up a little The biggest protocol change landed just before this window — the July 28 spec killed `Mcp-Session-Id` and the sticky `initialize` handshake. Any request can hit any replica now. This week the ecosystem caught up: - **Ruby SDK 1.0** went stable August 5-6. First SemVer-stable API for Ruby. If you're a Rails shop, your excuse for not shipping an MCP server just expired. - **Agents Working Group chartered August 5** inside the spec repo. Not a spec change yet, but agents are getting first-class treatment, not just bolted-on tools. - **Centaur 2.0** shipped August 10 from Paradigm — self-hosted, multiplayer runtime rebuilt with granular permissions, company-wide context archival (Gmail, Granola, Slack), and MCP. That last one sets up the week's real split. ### Managed vs. self-hosted LangChain went the other direction. On August 7, Managed Deep Agents hit public beta on LangSmith — durable execution, memory, sandboxes, channels, evals, all hosted. Two days earlier they published the mirror image: an autonomous SRE agent recipe for Kubernetes that keeps a human approval gate for changes. Two bets, same question: who holds the audit log and the credentials? LangChain says "give us your agents, we'll run them right." Paradigm says "run them yourself, with real permissions." Neither is turnkey yet. A small bridge worth noting: **Vibsync** hit Hacker News August 8 — one MCP endpoint (`mcp.vibsync.com`) that gives Claude Code, Cursor, and Codex shared memory across sessions and machines. Still in free beta, but it nails the real pain: I'm not losing time to model quality, I'm losing it to agents re-doing research another agent already did. ### Coding assistants became supermarkets GitHub resumed rolling out **Kimi K3 in Copilot** August 6, billed at provider list pricing. Not "Copilot picks for you" — you pick Kimi. Same week, **OpenAI Codex** cut eight releases including `rust-v0.147.0` and a string of `v0.148.0-alpha` builds, and **Claude Code** quietly lifted its 200-subagent limit in an August 7 commit. Different moves, same direction: the editor isn't the moat anymore. The moat is which models you can route to and how many subagents you can fan out. Microsoft Research even dropped *Agentic Coding in the Wild* on August 9 — a production trace study of Copilot at scale. We're finally getting telemetry instead of vibes. ### What I'd try next Point one agent at a stateless MCP server and watch what breaks without session affinity. Then add any shared memory layer — even a dumb one. The router and the memory will teach you more about agent quality than the next model bump. And if you're picking a framework this week, don't pick a model. Pick a control plane: who runs it, who audits it, and how you swap models when — not if — the router tells you to. === # This Week in AI Models: No New Flagship and Why That Matters URL: https://brianbest.com/blog/2026-08-10-ai-week-w02a-models Date: 2026-08-10 Tags: ai-weekly, frontier-models, benchmarks Summary: No flagship model dropped July 28 to August 3 — the labs spent the week slicing price and tiering access instead of chasing a bigger benchmark number. If you were waiting for a new flagship to drop last week, you kept waiting. Between July 28 and August 3, none of the big labs — OpenAI, Anthropic, Google, Meta, xAI, Mistral — pushed a new flagship model. I checked primary sources and syndication feeds for that window and couldn't verify a single new frontier release or a fresh eval sheet. That's not a miss. That's the pattern. What we got instead was pricing and packaging. ### The July 30 pricing tell The clearest model news that actually landed in that window was OpenAI repricing the GPT-5.6 family on July 30: * Luna (the smaller, fast model) down 80% to $0.20 input / $1.20 output per million tokens * Terra down 20% to $2 / $12 * Sol (frontier) flat at $5 / $30, with a new Fast mode at 2x price for ~2.5x speed No new weights. No new benchmark. Just a deliberate split: use Luna where you can, Terra for drafting, Sol only where you need the reasoning. When a lab cuts the price of its small model 80% and leaves the flagship untouched, it's telling you where it thinks volume lives. That's more useful than another point on MMLU. ### Benchmarks were quiet — intentionally? Same story on evals. No verified new scores for MMLU, HumanEval, SWE-bench, or GPQA hit the wires that week from a flagship. No press release with a "beats Sol by x%" chart. I looked. That's healthy. The benchmark cycle has gotten noisy — every 2% gain packaged as a breakthrough, every tool-use eval sensitive to prompt and harness. A quiet week without a new chart gives you time to run your own evals. Vendor numbers are directional. Your workload numbers are the decision. The chatter that *did* move was about cost-adjusted retrieval, not raw intelligence. Frontier models still charging ~$0.03 for a multi-turn search loop while small models are 100x cheaper but can't retrieve reliably. That tension was building all through the week of the 28th and would break into the open August 5. ### What a non-release week tells you When no flagship ships, the question shifts. It's not "which model is smartest?" It's "which tier do I route to?" And the labs answered without shipping anything new: 1. **Intelligence is tiered now.** Same family, different price and speed. You're not picking a model, you're picking a routing policy. 2. **Access is tiered too.** The Cyber/Daybreak framing that went public the next week — gated exploit capability for verified users — was already in motion. Frontier capability without open access. 3. **Benchmarks aren't the bottleneck.** Latency and cost per retrieval loop are. If your agents search three times before they answer, a small retriever that matches Sol changes your architecture more than Sol+0.5% does. I know "no new model" is a boring headline. But boring weeks are when deployment reality catches up to the demo. The labs didn't need a bigger number last week. They needed you to actually use the small one. If you're planning for next week, benchmark Luna on your retrieval paths, measure Sol only on your hardest 5% of tasks, and keep Fast mode for interactive work. When the next real flagship lands — and it will — you'll already know where it earns its keep. This week belonged to the spreadsheet, not the leaderboard. === # This Week in AI Products: DeepMind Shipped Robots While Enterprises Signed Paper URL: https://brianbest.com/blog/2026-08-10-ai-week-w02b-products Date: 2026-08-10 Tags: ai-weekly, enterprise-ai, product-launches Summary: One lab shipped three products in one week — none of them chatbots. And enterprise adoption was all partnerships, no deployment numbers. If you were watching for a new chat model last week, you missed where the products actually shipped. Between July 28 and August 3, every verifiable product launch came from one lab — Google DeepMind — and none of them were chat upgrades. Enterprise adoption? Same quiet. One big partnership expansion, zero hard deployment numbers. That's the tell. The product story moved to embodied AI. The enterprise story stayed on paper. ### DeepMind's triple Three launches, three days, all primary-sourced from DeepMind's blog: **Gemini Robotics 2 — July 28 (with a footnote).** Brings "whole body intelligence" to robots. The RSS says July 28, the page JSON says July 30. Both are Google sources and they conflict. I'm citing it as the week of July 28-30. **Lyria 3.5 — July 29.** Clean date: `datePublished: 2026-07-29T16:00:00Z`. DeepMind's music generation model, now in Google Flow Music. Changelog: richer melodic structures, higher-quality lyrics with better prompt adherence, more expressive vocals, plus tempo and duration controls. Not a chatbot tweak — a creator tool. **Gemini Robotics ER 2 — July 30.** `datePublished: 2026-07-30T15:00:00Z`. A step change in video understanding, tool orchestration, and multi-robot collaboration. No pip install, no API version in the post — a research-to-product push for embodied systems. Three products, zero chat benchmarks. I checked OpenAI and Anthropic primary sources for the same window and couldn't verify a new launch from either — `openai.com/news` and blog feeds were behind Cloudflare challenges when I pulled them August 11. If someone cites a GPT or Claude launch for that exact week, they didn't pull the primary. ### Enterprise adoption: a handshake, not a rollout The closest enterprise story to the window is Anthropic and Cognizant expanding their partnership to bring Claude to enterprise clients — dated July 27 on Anthropic's newsroom index, one day *before* the window. Title and slug verify, but the body is Next.js-rendered and wouldn't load via fetch, so I can't confirm scope or customer counts. Treat it as distribution, not deployment. Inside July 28–Aug 3 itself? I couldn't verify a single enterprise adoption number — no seat counts, no ARR, no "X thousand employees now on Y." That's notable. The numbers showed up right *after* the window closed: * Tenable expanding AI security coverage to Gemini — aggregated Aug 4 * Amazon Nova 2 hitting 50K customers — aggregated Aug 5 * Hitachi's CIO laying out enterprise AI strategy — Aug 5 All secondary via Google News RSS, all outside the week. Enterprises announced *strategy* in this fortnight, not scale. ### What this week actually means Don't read this as a slow week. Read it as a split. DeepMind is betting product velocity lives in robotics and creative tooling — whole-body control, video understanding, multi-robot work — while the chat labs were quiet. That rhymes with OpenAI's July 30 pricing move: Luna down 80%, Terra down 20%, Sol flat, instead of new weights. Different labs, same message: intelligence is table stakes, *where* you deploy it is the product. And enterprise buyers weren't buying more tokens. They were buying distribution. A Cognizant-style partnership doesn't show up on a benchmark, but it determines who actually gets Claude inside a Fortune 500 workflow. If you're planning Monday: don't wait for a flagship. If robotics or media touch your product, pull the DeepMind posts directly and watch for enterprise bundles. If you're on the enterprise side, ask your SI what that Cognizant-Anthropic expansion actually changes in contract terms. "Partnership expanded" without deployment numbers is a press release, not a rollout. The quiet weeks are when the plumbing gets laid. === # This Week in Open Source: One Header Deleted, One Flash Preview URL: https://brianbest.com/blog/2026-08-10-ai-week-w02c-opensource Date: 2026-08-10 Tags: open-source, ai-weekly, research Summary: July 28 to Aug 3 was quiet for open weights and papers — the only verifiable open source ship was MCP going stateless on July 28, plus DeepSeek's V4 Flash hitting beta July 31. You'd expect open source to fill the gap when the big labs go quiet. It didn't — not this week. Between July 28 and August 3, I couldn't verify a single new open-weight release or a fresh arXiv paper with a date in that window. No new Apache 2.0 checkpoint, no new eval sheet on Hugging Face, nothing citable from primary sources. That's not me missing it. Hugging Face, arXiv, and GitHub trending were all quiet for that exact 7-day slice, and the labs' own newsrooms were behind Cloudflare or JS hydration anyway. That silence is useful to name out loud. I'd rather tell you nothing shipped than invent a paper from July 30. ### The one thing that did ship: MCP goes stateless The most important open source change actually landed *on* July 28 — the MCP spec revision `2026-07-28`. One deletion: `Mcp-Session-Id` is gone. So is the sticky `initialize` handshake. Before, your client had to pin to one replica after `initialize`. Now any request can hit any replica. It's a line deleted from a spec, but it's the difference between "works on my laptop" and "scales behind a load balancer." If you run MCP servers in production, this is your week-one todo. Upgrade your SDKs, drop the session affinity config, and test horizontal scaling. The ecosystem catch-up — Ruby SDK 1.0, the Agents Working Group — all started from this change, but they landed the *next* week. This week was the spec. ### Just inside the window: DeepSeek-V4-Flash DeepSeek put `DeepSeek-V4-Flash` (tagged `0731`) into public beta on July 31 — four days before most roundups closed. Same API shape as before, but the pitch is specific: stronger agent performance than V4-Pro-Preview, with V4-Pro proper still "coming soon." No verified benchmark sheet I could pull with a stable date — no SWE-bench or τ-Bench numbers with a July 31 stamp. So treat it as a preview to benchmark yourself, not a leaderboard claim. If you have DeepSeek API access, it's the only new weight-adjacent thing worth running your own agent loop against from this window. Measure tool-use completion on your MCP servers, not token speed. ### What didn't ship (and why I'll call it out) No verifiable papers for this window. The papers that *did* matter — CreativeInstruct (2608.07460), CoinRAG (2608.07458), Evidence-RL (2608.08021) — are all stamped Aug 7-8. I'm not backdating them to make the week look busy. Same for open weights. Muse Glimmer 30B Apache 2.0, LFM2.5-2.6B, Shieldstral 3B — all Aug 4-10. They were the open source explosion the *next* week, and they're the direct answer to this quiet week. But they don't belong in a July 28-Aug 3 post. ### So what do you do with a quiet week? Two things. 1. **Update MCP.** If you pinned sessions, unpin them. Stateless MCP is the plumbing that makes everything next week possible. It's boring and load-bearing. 2. **Prep your eval harness.** A quiet week is the best time to run your own numbers. When Flash or Glimmer drops into your stack, you want a before/after on your tasks, not a vendor chart. Next week open source would do the shipping the labs didn't — 30B on one GPU, 2.6B on device, safety as a prompt. This week just cleared the runway. === # This Week in AI Policy: Brussels Flipped the Switch While Washington Stood Still URL: https://brianbest.com/blog/2026-08-10-ai-week-w02d-policy Date: 2026-08-10 Tags: ai-policy, regulation, industry-moves Summary: The EU's GPAI rules went enforceable August 2 while Washington offered no new AI law — a quiet U.S. week that says as much about direction as any executive order. If you wanted a clean story about AI regulation last week, you got half of one. In Brussels, a law actually turned on. In Washington, nothing new passed or got signed. That contrast *is* the story for July 28 to August 3 — proactive enforcement on one side of the Atlantic, deregulatory holding pattern on the other. ### Brussels went live August 2 The EU AI Act's General-Purpose AI obligations became enforceable August 2. I verified this across multiple outlets via Google News RSS on August 11 — azertag.az flagged it July 31 (*EU Commission to enforce AI Act rules and new transparency requirements on August 2*), Machine Brief ran *Enforcement Starts August 2 — Fines, Transparency* on August 1, and Business Today went live with it August 2. HelpNetSecurity recapped it August 4 as *putting AI models under the microscope*. What's actually new isn't chatbots getting banned. It's paperwork with teeth: transparency duties, copyright training-data summaries, and risk-management documentation for GPAI providers, plus the fine regime that backs them. The Commission's template and code-of-practice work finished just before the deadline — that's why you saw OpenAI leasing 88,000 square feet in Dublin for 250 staff that same week. When a U.S. lab expands its EU HQ right as GPAI fines go live, that's not optics. That's treaty-shopping for compliance. I'm hedging slightly on the statutory plumbing — the Act staggers duties and the Aug 2 date is the GPAI second wave — but the enforcement signal is unambiguous. If you ship a foundation model to Europe, you now file the docs. ### Washington didn't ship anything new I went looking for a new U.S. executive order or bill in that exact window and didn't find one. What I *did* find on the Federal Register API is that the operative framework is still the three orders published July 28, 2025 — exactly one year earlier: * *Promoting the Export of the American AI Technology Stack* (2025-14218) * *Preventing Woke AI in the Federal Government* (EO 14319, 2025-14217) * *Accelerating Federal Permitting of Data Center Infrastructure* (2025-14212) All three repeat the same line: easing federal burdens to build and export. There's no new AI EO in the July 28–Aug 3, 2026 window. That's not a gap in my search; it's the policy. The White House's 2025 "America's AI Action Plan" is still the governing story, and the legislative push for a uniform federal framework that would preempt state AI laws — floated in draft language that later showed up as a December 16 Federal Register doc — hadn't moved by August 3. My RSS search for "AI moratorium OR preemption" returned zero hits *inside* the week. In other words: states kept writing AI rules while Washington argued states *shouldn't* be able to. ### The one governance event that did happen The closest thing to a new U.S. governance move that week was a letter, not a law. FedScoop reported August 3 that a public-interest coalition urged Congress to investigate the OpenAI and Hugging Face hack. That's it for verifiable Hill action inside the window — no hearing scheduled, no bill text, just pressure for oversight on operational security rather than model capability. And just outside the window, the reason that pressure is building got obvious. Politico headlined August 5 that *Data center foes win big in Tuesday's primaries*, with NPR, the Texas Tribune, and San Antonio outlets all reporting August 6-8 that data-center buildouts have become a bipartisan electoral liability. That backlash was brewing July 28–Aug 3 while the July 28, 2025 EO was still telling agencies to *accelerate* permits. The politics have inverted: Washington wants to speed up builds, voters want to pause them. ### What I'm watching 1. **Filing, not headlines.** Watch the Commission's public training-data summary template — that's the document your lawyers will actually fill out. Fines follow forms. 2. **Preemption will be the U.S. fight.** No new federal preemption law passed last week. Until one does, you're compiling for 50 state targets if you deploy nationally. 3. **Security is becoming governance.** The Aug 3 hack letter matters more than it looks. Congress is more likely to act on breaches and infrastructure than on benchmarks. Brussels set a clock. Washington checked its watch. If you're shipping models, you now answer to Brussels whether Washington acts or not. === # This Week in AI Agents: Code Review Became an Agent URL: https://brianbest.com/blog/2026-08-10-ai-week-w02e-agents Date: 2026-08-10 Tags: ai-agents, coding-tools, github-copilot Summary: GitHub didn't ship a new model July 28–Aug 3 — it shipped the plumbing: review that runs your tools, PRs that stack, and agents you trigger with a comment. If you were waiting for a new coding model last week, you missed the actual shift. Between July 28 and August 3, GitHub dropped a dozen changes and none of them were "smarter autocomplete." The work was all plumbing — making review, PRs, and the editor behave like an agentic system instead of a chat box with a diff view. That's more consequential than a benchmark bump. ### Code review grew a brain The headline is July 29: Copilot code review support for Agent Skills and MCP is now generally available. Previously in preview since June 2, now on for Pro, Pro+, Business and Enterprise. This is the biggest change to review I've seen in a while. You drop a `SKILL.md` under `.github/skills/` and review actually invokes your team's standards and internal tools during the pass. Wire up an MCP server and it pulls context from your issue tracker or service catalog — read-only, but live. GitHub's own MCP and Playwright's are the exceptions where it does more. Same week, Copilot cloud agent got a reasoning knob (Aug 3). You pick a level alongside the model when you delegate — higher reasoning helps on complex tasks, burns more tokens and credits. It's the first cost/performance dial for the cloud agent, and it mirrors what OpenAI did with reasoning effort on o1/o3. Review isn't a linter anymore. It's an agent with tools and a budget. ### PRs learned to stack, actions learned to self-reference July 30 shipped two things engineers will actually feel Monday morning. **Stacked PRs in public preview.** Ordered series of PRs you can review in parallel and merge in one click. Vercel's been using it on Next.js for months — Tim Neutkens said it let them ship smaller changes while building larger features. If you've been duct-taping `ghstack` or `spr`, this is the native version. **Self-repository syntax: `$/`.** This one's small and perfect: ```yaml # Before: checkout dance # - uses: actions/checkout@v4 # - uses: ./path/to/local/action # Now: no checkout, pinned to the running commit - uses: $/actions/deploy@v1 ``` `uses: $/path/to/action` resolves to your own repo at the exact commit running the workflow. Works in workflow steps, composite actions, nested composition, reusable workflows. If you're having an agent scaffold workflows, this eliminates a whole class of fragile relative-path bugs. And in the IDE, both VS Code (July 30, v1.127–v1.131) and Visual Studio (July 30) pushed agents to the front. VS Code's new Agents window gives you an editor panel, compact diffs, and worktrees for any harness — Copilot, Claude, Codex. Visual Studio's new Agent preview runs on the same Copilot SDK as VS Code and CLI, with portable skills. The pattern's clear: worktrees are the primitive, the IDE is the orchestrator. ### Now you can poke agents with a comment The Aug 3 pair closes the loop. **Trigger Copilot automations with comments** — also Aug 3. You can set an automation to fire on `issue_comment` or `pull_request_comment` with a trigger phrase. Comment `/generate-docs` on a PR and the cloud agent updates docs. Comment a stack trace on an issue and it investigates. It's available on Pro/Pro+/Max/Business/Enterprise (Business/Enterprise need the cloud agent policy on). Pair that with **enterprise team specialization for managed settings** (Aug 3) and **team-level model targeting** (July 31, public preview) and the governance story clicks: enterprises are moving from org-wide toggles to per-team policies. Who gets which models, who can remote-control a CLI session from a phone — now at the team level. One retirement to note: **GitHub Models was fully retired July 30** — playground, catalog, inference API, BYOK, all gone. If you were prototyping on `api.github.com/models`, you're migrating to Foundry or Copilot. GitHub's done trying to host inference. ### What I'd try next Set up one stacked PR with two layers and point review with a single SKILL.md at it. Then add a comment-triggered automation on that repo — even a dumb one like "create follow-up issues." You'll learn more about your agentic workflow from that hour than from swapping models. I couldn't verify Cursor, Claude Code, or Codex releases for this exact window — their changelogs were behind Cloudflare/JS walls when I pulled Aug 11. So this week is a GitHub story, but the direction is universal: the moat isn't the model anymore. It's who routes, who reviews, and who remembers. === # One Flagship Shipped This Week: Claude Opus 5 URL: https://brianbest.com/blog/2026-08-03-ai-week-w03a-models Date: 2026-08-03 Tags: ai-models, benchmarks, weekly-roundup Summary: A quiet flagship week where Anthropic owned the window with Claude Opus 5 — not a raw intelligence leap, but a cost-per-task one. It was a quiet week for flagship models, and that's why the one that did ship stands out. For July 21-27, I checked the usual newsrooms — OpenAI, Anthropic, Google, Meta, xAI, Mistral — for a new foundation model with a verifiable date. Only one cleared the bar: Anthropic's Claude Opus 5 on July 24. ### What shipped: Opus 5 Opus 5 is the successor to Opus 4.8, same base price, now the default on Claude Max and the strongest option on Claude Pro. Anthropic's framing is telling — it's not pitched as "smartest at any cost" but as the Opus tier rebuilt for long-running agents. The tagline in their post: "a step change improvement for the Opus tier powering long-running agents." If you've been waiting for a GPT-5.6 follow-up, you didn't get one here. OpenAI's GPT-5.6 (Sol/Luna/Terra) launched July 9, outside this window. Inside July 21-27, OpenAI shipped product and infra notes — a Hugging Face security incident disclosure, ChatGPT for small business, OpenAI Presence — but no new weights. ### The benchmarks: cost is the story Anthropic didn't lead with a single MMLU number. They led with cost-efficiency curves, which is more honest about how these models actually get used. Here's what they claimed from internal runs: * **Frontier-Bench v0.1:** SOTA, and "more than doubles Opus 4.8's performance at a lower cost per task." Footnote matters: internal run on mini-SWE-agent + GKE, mean reward over 5 attempts per task, with Opus 4.8 as fallback on safety refusals for Opus 5 and Fable 5. * **CursorBench 3.2:** At max effort, within 0.5% of Claude Fable 5's peak score, but at half the cost per task. Also the best performance-at-a-given-cost on high, xhigh, and max effort settings. * **OSWorld 2.0 (computer use):** Beats every other model at any given cost, and surpasses Fable 5's best result at just over a third of the cost. * **GDPval-AA (Artificial Analysis):** Anthropic says Opus 5 is SOTA on coding and knowledge work here, but still behind Mythos 5 on cybersecurity. That split shows up elsewhere too. * **OSS-Fuzz (cybersecurity):** Close to Mythos 5 at *finding* vulnerabilities, "considerably less successful at developing exploits." Anthropic says this is intentional — they didn't train Opus 5 on cyber tasks. The general capability still lifted identification, but not weaponization. * **Behavioral audit:** Scores 2.3 on overall misaligned behavior, which Anthropic calls the lowest of their recent models. A lot of this is chart-only in the post — no numeric tables for HLE, AutomationBench, or DeepSearchQA in the HTML I could extract — so treat the absolutes as "Anthropic's internal harness says" until third parties repro. Frontier-Bench and GDPval-AA are the two you can actually go check independently. ### What didn't ship No verifiable flagship release from Google DeepMind, Meta, Mistral, or xAI with a date inside July 21-27. I tried their blogs and RSS feeds directly; most returned JS shells or Cloudflare challenges, and Mistral's news page only surfaced 2024 and 2025 posts in that fetch. That doesn't prove silence — it proves I couldn't verify a ship. ### My take The interesting shift isn't that Opus 5 is smarter. It's that Anthropic is competing on dollars per completed task. Doubling Frontier-Bench at lower cost, matching Fable 5 on CursorBench for half the price, beating it on OSWorld for a third — that's a pricing story disguised as a benchmark story. And it's consistent with where the labs have been heading: GPT-5.6's own July messaging was "frontier intelligence that scales with your ambition" and price-performance framing. If you're benchmarking this week, don't just run the leaderboard. Re-run your own agent tasks and log cost per successful run with effort dialed up and down. That's the curve Anthropic wants you to look at, and for once I think it's the right one. === # This Week in AI Products: Kimi K3 Lands and Enterprises Buy Distribution URL: https://brianbest.com/blog/2026-08-03-ai-week-w03b-products Date: 2026-08-03 Tags: ai-weekly, enterprise-ai, product-launches Summary: One 2.8T open model shipped with day-0 infra from vLLM and Telnyx. Anthropic bought enterprise distribution via Cognizant instead. If you wanted a new closed flagship last week, you didn't get one. If you wanted a signal on where enterprise AI is actually going, you got two — one model that forces you to rethink serving costs, and one partnership that tells you how Claude gets into Fortune 500s. ### Kimi K3 actually shipped — with day-0 infra July 27, Moonshot AI dropped Kimi K3. It's not incremental. 2.8T total parameters, 104B active per token, 896 experts with 16 active + 2 shared, 93 layers (69 Kimi Delta Attention + 24 Gated MLA), 1M token context, native vision via MoonViT-V2, MXFP4 weights. Their line: "world's first open 3T-class model." Source is the README on `MoonshotAI/Kimi-K3` — fully fetchable, HN indexed at 15:23 UTC with 391 points. The tech report PDF is linked there if you want the benchmark tables. Licensed under Kimi K3 License, not MIT — read that before you plan a commercial fork. What matters more for enterprise is that infra shipped *with* it. That's rare. Telnyx had it on Inference at 22:46 UTC the same day: model ID `moonshotai/Kimi-K3` alongside K2.6, listed at $2.70 per million input tokens, 1M context, native vision. Pricing is in their JSON-LD description — I couldn't fully render the body pricing table, so confirm output pricing before you budget. vLLM shipped day-0 support the same afternoon with Inferact. They claim 118 tok/s without speculative decoding, 370 tok/s with their DSpark speculator — a 3.14x lift on 16x NVIDIA GB300 NVL72. Minimum to run at all: 8x B300 or 8x AMD MI355X. And it's Docker-only right now with pre-release FlashInfer. So yeah, you can run a frontier open model — if you've got a rack that looks like a supercomputer. ```bash vllm serve moonshotai/Kimi-K3 \ --tensor-parallel-size 8 --trust-remote-code \ --enable-prefix-caching --tool-call-parser kimi_k3 ``` The enterprise takeaway isn't "2.8T is big." It's that open weights just jumped an order of magnitude and the serving tax jumped with it. 1M context is wonderful until you pay for the KV cache. ### Enterprises didn't buy models — they bought Cognizant The other verifiable launch July 27: Anthropic and Cognizant expanded their partnership. Cognizant becomes a Global Premier Partner in the Claude Partner Network, embedding Claude across its own platforms — Flowsource, Neuro AI Engineering, Neuro IT Ops — with 30,000+ associates Claude-trained. Flowsource's Spec-Driven Development module now runs Claude Code against your specs, coding standards, and architectural blueprints, then evaluates before production. That's distribution as product. The numbers are deployment-specific, not averaged, and they're company-reported: an agentic contract-intelligence system for a biopharma that cut review time up to 40% and lifted extraction accuracy above 88%, and a risk-navigation tool for underwriters saving ~8 hours per person per week — hours of research to minutes. Plus a customer experience portal for a global manufacturer in six months. Real clients, anonymized names — ask for references before you cite them as typical. I'm not dismissing this as a press release. This is how enterprise adoption actually happens. As Cognizant CEO Ravi Kumar S put it: "AI capability is rising faster than enterprises can absorb it." SIs are the bridge. Anthropic gets 30k certified builders; enterprises get someone to blame when auto-mode ships a bad diff. ### Safety as product, quietly Same day, NVIDIA announced the Open Secure AI Alliance at 09:00 UTC — "open tools that promote responsible use of and trust in AI" focused on agentic AI and cybersecurity. Founding members and charter were behind JS rendering when I pulled it, so check the NVIDIA blog directly. And a WSJ headline indexed at 17:38 UTC reported an AMD-Anthropic chips-and-investment deal — paywalled, I couldn't verify terms. Also floating: an NVIDIA x Safe Superintelligence partnership with a social claim of 10x compute in 12 months. Headlines only, details blocked. Don't treat any of that as closed. Treat it as the plumbing. The model was the headline. The alliances and chip deals are the financing. ### What I'd do Monday 1. **Price Kimi K3 against your context needs.** If you don't need 1M tokens, you don't need 8x B300s. Benchmark Telnyx vs. self-host before you fall in love with open weights. 2. **If you're evaluating Claude, talk to the SI.** That Cognizant expansion changes who implements, not just what model you call. Ask what Flowsource actually enforces. 3. **Track OSAIA before you build your own guardrails.** If NVIDIA ships open safety tooling for agents, you shouldn't be rolling your own. === # Open Source Didn't Wait for Permission: Kimi K3 and the Infra That Trains It URL: https://brianbest.com/blog/2026-08-03-ai-week-w03c-opensource Date: 2026-08-03 Tags: open-source, ai-weekly, research Summary: For July 21-27 open source shipped the only frontier-scale drop of the week — a 2.8T Kimi K3 — plus the RL infra that trained it, a real-time voice runtime, and four papers that fix boring but real problems. Everyone was looking at closed flagships last week. The only thing that actually shipped at frontier scale was open weight. For July 21-27, the verifiable open source story is tight: one massive model, the platform that trained it, a voice runtime, and a handful of papers that are less flashy but more useful than another leaderboard bump. ### Kimi K3 is 2.8T you can download MoonshotAI dropped Kimi K3 on July 27 — GitHub `created_at: 2026-07-27T08:01:37Z` on `MoonshotAI/Kimi-K3`. Their README line: "2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals, with native vision and a 1-million-token context window. It is the world's first open 3T-class model." Details from that README: 896 experts with 16 active per token (~104B active), 93 layers (69 KDA + 24 Gated MLA), claims ~2.5x scaling efficiency over K2 via Stable LatentMoE, MXFP4 weights. Native vision via MoonViT-V2. The catch is in the license. It's Kimi K3 License, not Apache 2.0 or MIT. Call it open-weight, not open source in the OSI sense. And you won't run it on a laptop — vLLM's day-0 support lists 8x B300 minimum. But if you have the iron, it's there. This is the only trillion-parameter open drop in the window. The next Hugging Face features — LiquidAI LFM2.5-Encoders, AllenAI OlmoEarth — are dated July 28, just outside. ### AgentENV: the part most drops don't ship More interesting to me: Moonshot open sourced AgentENV the same week. GitHub `created_at: 2026-07-23T02:48:07Z` on `kvcache-ai/AgentENV`. It's the distributed platform they used for agentic RL on K3. From their README: "powering agentic RL training for Kimi K3," running "massive numbers of Firecracker environments" with snapshot-backed boot in <50ms, pause in <100ms, and "scaling to 1.5 million images in production." Most open drops give you weights. This gives you the harness. That's rarer and, for anyone trying to do RL on agents, more useful. ```bash curl -fsSL https://raw.githubusercontent.com/kvcache-ai/AgentENV/main/scripts/install.sh | sudo bash sudo systemctl start aenv ``` Needs Linux 6.8+ and /dev/kvm. Don't try it on a Mac. ### Qwen Audio Agent keeps talking while your agent works Also July 27: `QwenAudio/qwen-audio-agent` (`created_at: 2026-07-27T05:04:26Z`). Their pitch: "A realtime voice runtime that keeps Agents talking, working, and present." It's the VAD/STT/LLM/TTS glue for voice agents that don't go silent when you call a tool. Small repo, 2k stars already. If you're building voice, you know the awkward pause problem — this is an explicit fix for it. ### Papers worth your time Four arXiv submissions in-window, all with August IDs (2608.x) but verified `[Submitted on 21/24 Jul 2026]` on the HTML: **Speculative Correction** (2608.02625, Jul 21) — draft-then-refine decoding for diffusion LMs. Don't generate left-to-right block by block; generate a full draft and refine it bidirectionally. Training-free. They claim LLaDA-2.1-Flash-Flash goes GSM8K-384 0.848 → 0.899 at 1.2x speedup, MBPP-384 0.545 → 0.693. If you're watching DLMs, decoding tricks are now as important as bigger checkpoints. **Search-G1** (2608.07531, Jul 24) — grounded search agents via representation-based intrinsic rewards. The problem: sparse outcome rewards vs. expensive LLM-judge process rewards. Their fix rewards retrieving only when needed and grounding answers. It's the "stop over-searching" paper RAG needs. **The Knowing-Saying Gap** (2608.07528, Jul 21) — linear probes detect corrupted context near-perfectly but still can't predict final-answer correctness, and asking models for structured confidence collapses to two values. If you're relying on verbalized confidence for monitoring, this is a warning shot. **DocAtlas** (2608.07527, Jul 21) — reframes long-document understanding as mutable-state interaction with a learned policy, not static retrieval. Aimed at hundreds of pages with tables and figures where normal RAG falls apart. ### What I'd do Monday If you have the hardware, pull K3 and benchmark it on your long-horizon tasks — not chat. If you don't, pull AgentENV and see how they think about env scale; you'll steal ideas even if you never train a 3T model. And read Speculative Correction before you write off diffusion LMs as slow. The decoding playbook is moving fast. === # This Week in AI Policy: The Quiet Before Brussels Hit Go URL: https://brianbest.com/blog/2026-08-03-ai-week-w03d-policy Date: 2026-08-03 Tags: ai-policy, regulation, industry-moves Summary: For July 21-27, AI regulation mostly held its breath — no new EU enforcement, no new U.S. law, and industry kept its checkbook closed before the August rush. If you were waiting for AI regulation to do something dramatic the week of July 21-27, you were watching the wrong week. I ran a full research pass for that window — regulation, policy, and industry moves — and came up mostly empty on verifiable, date-stamped actions *inside* those seven days. That isn't a search miss. It's the signal. The machinery was loaded, but nothing fired until the next two weeks. ### Regulation: between waves The EU was in the gap between announcements. The General-Purpose AI obligations that would go enforceable August 2 hadn't flipped yet, and the transparency provisions that would go live the week of August 4-10 hadn't either. For July 21-27, Brussels was doing guidance and templates, not enforcement. No new fines, no new disclosure duty went live in that window that I could verify on the record. Washington was quieter still. The operative federal framework that week was still the trio of orders from July 28, 2025 — export the stack, cut federal permitting drag for data centers, and the EO on "woke AI" in government — all still on the Federal Register without a successor order inside July 21-27, 2026. I didn't find a new AI executive order, a signed AI bill, or a final agency rule dated in that window. Congress was in letter-writing mode, not law-passing mode, and the FTC/FCC didn't put a new AI rule on the books that week. States didn't break the silence either. The moves that would matter — California directing a new AI Cyber Defense Program and Texas setting new AI data-center rules — both hit the week of August 4-10. For July 21-27, it was mostly positioning and committee work, the kind that doesn't get a timestamp until the press release goes out. And China, India, and the UK? No verifiable new AI-specific regulation with an effective date inside July 21-27 crossed my filters. India's 3-hour takedown and labeling rule and China's companion guardrails both land in the following two weeks. I hedge this on purpose. Regulatory calendars leak in drafts and speeches before they show up in gazettes. But if you need a date you can cite to a lawyer, this wasn't the week. ### Industry moves: wallets stayed shut Same story on industry. No billion-dollar AI funding round, no closed M&A, no major antitrust action with a July 21-27 announcement date cleared verification in this pass. The deals that defined the next two weeks — Google's $1.5B+ talks for coding-agent startup Mechanize, OpenAI rolling the NextSlide team into ChatGPT, SAP freezing hiring to fund AI capex — all date to August 4-10, not this window. That quiet tracks. Late July is when term sheets get held for August press cycles and when data-center permitting headlines were still building as local-politics stories, not national ones. The Politico/NPR wave about data-center backlash at the ballot box wouldn't crest until early August. If you were fundraising or buying capacity that week, you weren't competing with a headline number. You were negotiating against the *expectation* of one. ### What I'm watching 1. **The clock that matters started August 2.** If you ship GPAI in Europe, your compliance work started *after* this quiet week. Use July 21-27 as your control — before paperwork had teeth. 2. **No federal preemption means 50 targets.** Nothing in this week changed the U.S. patchwork. You're still compiling for state-by-state rules until Congress actually passes something. 3. **Capital was coiling, not absent.** When a week this quiet is followed by a $1.5B talent chase for a 100-day-old team, that's not coincidence. Watch who closed in late August what they negotiated in late July. Back next week, the noise comes back. This week was the inhale. === # No New Flagship This Week — Just a Cyber Specialist URL: https://brianbest.com/blog/2026-07-27-ai-week-w04a-models Date: 2026-07-27 Tags: ai-models, benchmarks, weekly-roundup Summary: For July 14-20, no lab shipped a new frontier model — DeepMind's Gemini 3.5 Flash Cyber was the only benchmarked release, and it was built to find vulns, not win MMLU. I went looking for a flagship model drop for July 14-20 and didn't find one. That's not a miss — it's the pattern. The week before this window was loud. OpenAI shipped GPT-5.6 on July 9th, xAI shipped Grok 4 the same day. The week *of* July 14-20 was the exhale. ### What actually shipped: Gemini 3.5 Flash Cyber The only verified, benchmarked model release I could pull with a date inside this window was Google DeepMind's Gemini 3.5 Flash Cyber — RSS says July 17th, the rendered page says July 21st, I'm treating the RSS as canonical but flagging the gap. And it's not a flagship in the usual sense. It's a lightweight cybersecurity fine-tune built on top of 3.5 Flash. DeepMind's pitch: cheaper and more effective than the mainline Flash models at finding, validating, and patching vulns. Not general chat, not reasoning, not multimodal — vuln discovery. The benchmarks are all cyber-native, which is why they're worth paying attention to even if the charts are image-only and you can't pull exact percentages from the HTML: * **CyberGym** — hundreds of real-world vulns, pass@1. DeepMind says an agent calling 3.5 Flash Cyber up to 5 times per report was competitive with "significantly larger models." Footnote: competitor scores were self-reported. * **Big Sleep Evaluation** — Google's internal eval on huge, messy codebases like Chrome and Safari. Claim: "significantly surpassed" mainline 3.5 Flash and 3.6 Flash. * **Chrome Production Commit Scanning** — undisclosed vulns, so no contamination. Claim: significant uplift over 3.5 Flash. They note newer competitor versions after Claude Opus 4.6 weren't shown because they refused the tasks. * **V8 Engine (fixed invocations)** — the only hard numbers in text: 55 unique confirmed issues for 3.5 Flash Cyber vs. 47 for mainline 3.5 Flash vs. 36 for Opus 4.6, including 10 that only Cyber found. DeepMind also included an anecdote: Cloud Vulnerability Research found an RCE in public APIs and a memory-corruption in a production service in 2 hours, then generated a 100% reliable exploit bypassing ASLR and W^X. It's not generally available either. It's a limited-access pilot via CodeMender for governments and trusted partners. Everyone else gets the capability via general models on the Gemini Enterprise Agent Platform. So it's narrow, gated, and not on LMArena. But it's the only ship with receipts this week. ### What didn't ship No verifiable flagship model or general benchmark from OpenAI, Anthropic, Meta, xAI, or Mistral inside July 14-20. That doesn't mean they were quiet. OpenAI published *ten* items that week, but they were governance and enablement, not weights: * **July 20:** *Safety and alignment in an era of long-horizon models* — lessons from deploying long-running agents. * **July 17:** *A scorecard for the AI age* — CFO Sarah Friar proposing ROI metrics like cost per successful task and return on compute. Not a model eval, but a tell on how OpenAI wants enterprises to measure these things. * **July 15:** *GPT-Red* — automated red-teaming via self-play for prompt-injection robustness. System description, no benchmark table in the RSS I could read — the full page is now behind Cloudflare. Mistral's RSS shows nothing in-window. Nearest were Robostral Navigate (8B, July 8) and Shieldstral (August 4, outside the window). Anthropic, xAI, and Meta's blogs were all JS shells or Cloudflare challenges on fetch, so I couldn't verify absence cleanly — I checked RSS directly where I could, and browser-readable pages where I couldn't. No flagship titles with in-window dates surfaced. No new MMLU, GPQA, SWE-Bench, or Artificial Analysis leaderboard moves dated to this week in the places I could actually scrape. ### My take This is what a between-storms week looks like. Labs ship a flagship, then spend the next week shipping the safety story and the enterprise scorecard that makes the flagship deployable. The signal in the noise is the eval shift. OpenAI's scorecard and DeepMind's cyber benchmarks both move away from "one more MMLU point" toward cost-per-task and pass@1 on real code with tools. CyberGym and the Chrome commit scanner aren't chat evals — they're agentic, contamination-free, tool-using evals where calling a small model 5 times beats calling a big model once. That's a search-economics story, not an IQ story. If you were waiting to benchmark a new flagship this week, don't. Re-run your own tasks against GPT-5.6 (July 9) and whatever you're running in prod, and log cost per completed task the way that scorecard suggests. The next flagship will move the headline. The cyber harness is what's moving the error rate right now. === # This Week in AI Products: Gates and Routers Before the Spike URL: https://brianbest.com/blog/2026-07-27-ai-week-w04b-products Date: 2026-07-27 Tags: ai-weekly, enterprise-ai, product-launches Summary: No flagship model shipped. Enterprises bought quality gates and cost routers — plumbing laid the week before open weights jumped an order of magnitude. If you waited for a flagship model the week of July 14-20, you got nothing. If you watched what enterprises actually bought, you got the tell for everything that shipped the next two weeks. No OpenAI or Anthropic flagship with a verifiable primary post. I checked both newsrooms August 11 — Cloudflare challenge on one, Next.js shell on the other. Same blocks I flagged for the July 27 and Aug 3 briefs. If someone says GPT-5.x dropped this week, ask for the URL. What did ship was plumbing. ### Gates, not weights The one GA with a primary source: GitHub Code Quality. `datePublished: 2026-07-20T13:01:24Z` on the changelog — GA on Enterprise Cloud and Team. GitHub's own framing is blunt: AI accelerates code output, Code Quality is the gate. That's the enterprise buy in one sentence. More tokens, same review bottleneck — so you buy the filter. It's not flashy. It is exactly how AI adoption actually works inside a real org. You don't need a better autocomplete, you need something that says no before the diff hits main. ### Routers and kill switches The only hard enterprise efficiency number in the window came from a gateway. Ramp opened its AI Router — HN indexed at 21:09 UTC July 20 — with a company-reported claim it cut internal LLM costs 30% via model routing. Headline only; the body didn't fetch. Treat the 30% as reported, not audited — but the pattern is right. Same evening: Stoke landed at 18:59 UTC as a kill switch for runaway agents — Rust, budget caps — at stokegate.com. And at 23:18 UTC, Relay showed up as a self-hosted LLM gateway with eval-gated routing. Different vendors, same thesis: enterprises aren't buying more intelligence, they're buying control over spend and blast radius. Google whispered the same thing from the silicon side. HN indexed Reuters at 23:46 UTC July 20: "Google plans new chip to run Gemini models more efficiently." Reuters body was behind a captcha — headline only — but it rhymes with everything else that week. Cost per token is becoming a product. Even the protocol work pointed there. TechCrunch at 23:56 UTC: "AI's most important protocol is getting a little bit easier to use" — truncated slug, 404 on fetch, widely understood to be MCP. I can't verify the body, but I'll take easier MCP as directionally correct. ### The money and the risk One funding headline cleared: Natural raised $30M to reinvent payments for AI agents — TechCrunch, indexed 22:55 UTC July 20. Payments for agents sounds niche until you realize every router and kill switch above needs a metering layer. And enterprises were pricing risk the same day. A Reuters headline approved at 21:25 UTC: US judge approves Anthropic's $1.5B copyright settlement. Add Axios at 21:54 UTC — administration considering a ban on Chinese open source models, Kimi named — and you've got legal exposure spiking exactly seven days before Kimi K3 dropped as a 2.8T open model on July 27. July 14-20 was the prelude. Claude Code also shipped a screen reader mode that day — small, not enterprise, but worth noting for adoption surface area. ### What this week actually means This was the quiet week before the spike. The next week gave you Kimi K3 with day-0 vLLM and Telnyx infra and a 30k-person Cognizant-to-Claude distribution deal. The week after gave you three DeepMind launches, none of them chat. None of that happens without the gates and routers here. 1M context is wonderful until you pay the KV cache. Agentic coding is wonderful until you can't stop the agent. ### What I'd do Monday 1. If you're on GitHub Enterprise, turn on Code Quality and measure what it actually blocks. That's your AI velocity tax. 2. If you haven't priced routing, do it now — Ramp's 30% is a reported number, but even half that is real money. Eval-gate your router, don't just cheapest-model it. 3. If you touch open weights, read the Kimi K3 license and the July 20 policy headlines together. Distribution is coming via SIs, not Hugging Face downloads. === # Inkling Landed and Nobody Had the Infra Ready URL: https://brianbest.com/blog/2026-07-27-ai-week-w04c-opensource Date: 2026-07-27 Tags: open-source, ai-weekly, research Summary: July 14-20 had no flagship fireworks. Open source did the real shipping — a 975B MoE from Thinking Machines, a run of tiny specialist embeds, and papers that poke holes in how we eval agents. Flagships took the week off. Open source didn't — it just shipped the boring stuff that actually matters. If you were waiting for a new frontier chat model between July 14-20, there wasn't one. If you were watching Hugging Face and arXiv, it was busy. ### Inkling is big, and that's the point Thinking Machines dropped `thinkingmachines/Inkling` on July 14 — API `createdAt: 2026-07-14`, blog welcome July 15. It's 975B total, 41B active. MoE at a scale we haven't seen in open weights before. Day-zero infra wasn't ready. `huggingface/transformers` shipped `v5.14.0` on July 15 specifically to add Inkling — `Add Inkling model #47347` — then `v5.14.1` on July 16 to patch `EncoderDecoderCache` and assisted generation bugs that showed up integrating it. `vllm` cut `v0.25.1` on July 14, though that was a small patch (FFmpeg/TorchCodec fix), not an Inkling bring-up. ```python from transformers import AutoModelForCausalLM, AutoTokenizer tok = AutoTokenizer.from_pretrained("thinkingmachines/Inkling") model = AutoModelForCausalLM.from_pretrained("thinkingmachines/Inkling", device_map="auto") ``` You won't run this on one card. That's not the point. The point is a lab shipped a frontier-scale MoE as open weights and forced the ecosystem to catch up in 48 hours. Contrast that with last week's usual dance of API-only releases. Tradeoff is clear: you get weights, you also get the integration pain. Quants landed same day — `Inkling-NVFP4` and `unsloth/inkling-GGUF` both July 14 — so you can at least try it without a DGX cluster on fire. ### Small models did the useful shipping The most immediately usable drops were small: * **Nvidia Nemotron-3-Embed 8B/1B** — both `createdAt: 2026-07-14T07:38:49Z`. Embeddings, not chat. If you're doing RAG, these are the ones to bench. * **Microsoft bitnet-embedding-0.6b** — `createdAt: 2026-07-15`. 0.6B, 1-bit adjacent. Hanno Labs posted the same day (July 14) claiming #2 on MTEB Law with a 0.6B — different model, same thesis: sub-billion specialists are competitive where it counts. * **Microsoft Fara1.5 4B / 27B** — July 17. Hard to place without verified evals, but they're there if you track Microsoft's small-model arc. No new Qwen, Llama, or DeepSeek weights verified in-window. Their next drops cluster July 28 onward — I checked, nothing to backdate. ### Papers that fix how we eval ArXiv was active. The search API was rate-limited when I checked, so I verified via monthly lists and individual `Submitted on` banners. Three papers worth your time, all in-window: **VarRate (2607.15498, Jul 16)** — training-free variable-rate KV cache compression for long context. Two leading families both waste memory; VarRate adapts per-layer. If you're paying for 128K context, this is directly about your bill. **Verbalizable Representations Form a Global Workspace (2607.15495, Jul 16)** — argues LLMs already develop a brain-like global workspace where only a subset of representations become verbally reportable. Speculative, but it's a clean lens on why probing and verbalized confidence diverge. **Operational Hallucination and Safety Drift (2607.18366, Jul 20)** — multi-turn tool-using agents drift in reliability and safety over long runs. Pairs with two Jul 14 papers: *LLM Judges Can Be Too Generous Without Reference* (2607.12885) and *How Many Tasks Are Enough for Agent Benchmarks?* (2607.12338). Same takeaway: our evals are too generous and too short. If you're shipping agents, shorten your trust and lengthen your eval. Also in tools: `ComfyUI v0.28.0` shipped July 15. Not a model, but `AGENTS.md` in the changelog signals where even image workflows are heading. ### What I'd do Monday Pull Nemotron-3-Embed or bitnet-embedding and run it against your current embed on your own retrieval set — not MTEB. If you have the iron, try Inkling under `transformers` 5.14.1 and see what breaks. And read VarRate before you buy more KV cache. It was a quiet week at the top and a busy week at the bottom. That's usually how it goes before the next flagship. === # This Week in AI Policy: Brussels Pries Open Android and Starts the Label Clock URL: https://brianbest.com/blog/2026-07-27-ai-week-w04d-policy Date: 2026-07-27 Tags: ai-policy, regulation, industry-moves Summary: For July 14-20, Brussels ordered Google to open Android and Search to rival AI assistants, set the clock for AI label rules on August 2, and left US states and the labs themselves to fill a federal vacuum. If you thought AI regulation was on summer break, Brussels didn't get the memo. The week of July 14-20 didn't give us a new US AI law — it gave us the next enforcement map instead. ### Brussels tells Google to share The biggest move landed July 16. The European Commission told Alphabet/Google to open parts of Android and Search to rival AI assistants under the Digital Markets Act. Two deadlines, both specific: share selected Search data with competitors and AI tools by January 2027, and update Android to grant rival assistants system-level integration by July 2027. That second one is the teeth — wake-word activation so you can summon ChatGPT the way you invoke Google Assistant today, plus access to the on-device NPUs that handle local AI workloads. Legal hook is DMA Article 6(7), the interoperability rule for designated gatekeepers. Google qualifies. Google cited privacy and security risks; the Commission published guidance anyway, building on investigations opened in January 2026 and a July 2026 General Court ruling backing its authority. I verified this via the Commission guidance page and a fetched Cryptobriefing excerpt that quotes it. Reuters, AP, and France24 syndicated the same guidance July 16, though their bodies were behind captchas on my fetch. The tell isn't the headline — it's the level. Brussels went from app-store fines to OS-level AI integration. If you're building an assistant, your Android distribution strategy just changed. ### Labels now, hard rules later The EU also clarified what hits August 2. Under the AI Act's transparency obligations — Article 50 territory — providers and deployers have to label AI-generated or manipulated content: chatbot disclosure, deepfake watermarks, visible and machine-readable. Wilson Sonsini flagged the Transparency Code of Practice July 14. Il Sole 24 ORE, National Law Review, and EurActiv followed July 15-17 — including Greece floating jail time for stripping deepfake labels. Silicon Canals summed it July 18: labels go live August 2, tougher high-risk rules on hiring, biometrics, and migration get pushed to December 2027. As Silicon Canals put it: "The transparency obligations are the cheap ones to meet. The expensive, rights-protective rules are the ones that just got more time." If you ship content tooling in Europe, watermarking starts now. If you ship hiring or biometric systems, you just bought 17 more months — but that clock is running. ### Who writes the rules when Washington won't? In the US, the vacuum got loud. WIRED reported Anthropic is pushing states to regulate AI faster. TCAI counted 84 new AI laws enacted in 27 states in the first half of 2026. PYMNTS and Reuters noted state AGs warning that existing consumer protection laws already apply to AI — no new statute needed. Demis Hassabis tried to fill the same vacuum from the lab side. July 14, Axios reported he wants a US-led global AI watchdog; TechCrunch the same day framed it as a FINRA-style independent standards body to test frontier models. A Big Tech CEO proposing industry self-standards instead of government pre-approval — that's a bid to set the default. And Washington pushed back outward. July 14, USTR Jamieson Greer's line — "US won't allow Europe to regulate American tech" — was a direct shot at the DMA and AI Act's reach. Three bidders, one job: Brussels imposing interoperability, states writing the de facto federal law 84 times over, and labs proposing their own referee. ### What I'm watching 1. **Specs, not slogans.** For DMA, watch the Article 6(7) spec for what "selected Search data" and NPU access actually mean in AOSP. For labels, watch EUR-Lex guidance — the label is easy, the audit trail isn't. 2. **The patchwork tax.** If you deploy nationally, you're compiling for 27 state regimes plus AG enforcement under old laws. Nothing this week created federal preemption. 3. **Interoperability as antitrust.** Wake-word and NPU access is a template. Watch whether DOJ's search remedies start to rhyme with Brussels' data-sharing order — same playbook, different caption. Back next week. If this week was Brussels loading the gun, August is when it fires. === # Two Flagships in 48 Hours — and an 8B Robot That Actually Shipped a Number URL: https://brianbest.com/blog/2026-07-20-ai-week-w05a-models Date: 2026-07-20 Tags: ai-models, benchmarks, weekly-roundup Summary: July 7-13 was loud: OpenAI shipped GPT-5.6 and xAI shipped Grok 4.5 within a day of each other, while the only clean benchmark in the window came from Mistral's 8B Robostral Navigate. If you blinked July 7-13, you missed two flagships. OpenAI and xAI shipped within about 24 hours of each other. That's not the usual drip — that's both labs deciding the same window was the right time to go. ### What shipped **GPT-5.6 — July 9.** OpenAI put up the family page for GPT-5.6 that afternoon (HN syndication says `2026-07-09T17:04:14Z`, 1.5k points). It's not one model — it's Sol / Luna / Terra. Sol is frontier, Terra is the middle, Luna is the small fast one. Preview for Sol had gone up June 26, so this was the full family landing. I couldn't pull a clean benchmark table from the primary page on fetch — `openai.com/index/gpt-5-6/` was behind a Cloudflare challenge when I checked, and the system card PDF at `deploymentsafety.openai.com/gpt-5-6/gpt-5-6.pdf` is live (4.4MB, returns 200) but it's a Flate-encoded PDF that doesn't scrape cleanly. No verified MMLU, GPQA, or SWE-Bench numbers hit RSS in-window that I could confirm. If you want the evals, you'll need to open that PDF in a reader — I'm not going to invent them. What matters more anyway: they shipped a family designed to be routed, not a single model to be worshipped. That pricing wouldn't show up until July 30 — Luna down 80% to $0.20/$1.20, Terra to $2/$12, Sol flat at $5/$30 — but the architecture was the tell on the 9th. **Grok 4.5 — July 8-9.** xAI's turn came right alongside it. HN algolia has it as "Grok 4.5" at `2026-07-08T18:00:32Z`, my own notes from the following week call it "Grok 4 on July 9." I'm flagging both because the primary `x.ai/news/grok-4-5` page was also Cloudflare-blocked on direct fetch, so I couldn't resolve the naming cleanly. Either way, it's a flagship-class Grok in the same 48-hour window as GPT-5.6. No verified benchmark delta from Artificial Analysis or LMArena dated to that week that I could scrape. **Robostral Navigate — July 8.** The only model in the window with a verifiable, in-text benchmark was Mistral's. RSS says `Wed, 08 Jul 2026 12:00:59 GMT`, and the page meta confirms it: 8B, 76.6% on R2R-CE, single RGB camera. No depth sensor, no LiDAR, no multi-cam rig. That's not a chat benchmark. It's embodied navigation — can the robot follow language instructions through messy indoor spaces with cheap sensors. And Mistral put a number on it. Lean and specific beats vague and frontier-sized for once. ### What didn't ship No verified flagship or general benchmark from Anthropic, Google, or Meta inside July 7-13 that I could pull with a clean date. Anthropic's news page is a JS shell, Google's Gemini blog is too, and Meta's RSS didn't show anything in-window. That doesn't mean zero work — it means nothing hit the official newsrooms with a verifiable date I could confirm. And notably, no head-to-head eval table comparing GPT-5.6 vs. Grok 4.5 dropped that week. Two flagships, same day, zero shared harness. Labs are shipping faster than eval culture can keep up. ### My take This was a routing week disguised as a flagship week. Everyone wants the IQ chart. What we got was more interesting: two labs saying "here's the top model, and here are the smaller ones you're actually supposed to use." And one lab saying "here's what 8B can do when you stop adding sensors and start adding smarts." If you're deciding what to eval, don't wait for a clean GPT-5.6 vs Grok 4.5 leaderboard — run your own tasks. Route Luna/Terra where you can, save Sol for the hard 5%, and if you touch embodied stuff, Robostral's R2R-CE number is the only score this week with receipts. The benchmarks will catch up. The architectures already shipped. === # This Week in AI Products: Two Flagships and Three Enterprise Handshakes URL: https://brianbest.com/blog/2026-07-20-ai-week-w05b-products Date: 2026-07-20 Tags: ai-weekly, enterprise-ai, product-launches Summary: The week flagships actually shipped — GPT-5.6 and Grok 4 on the same day, GPT-Live the day before, and three OpenAI enterprise stories that matter less than the Copilot distribution deal. If the week of July 14-20 was the quiet before the spike, July 7-13 was the spike. Two flagships on the same day, a new live product the day before, and three enterprise customer stories in four days. That's the whole week in one sentence. The nuance is which of those actually changes enterprise deployment. ### Two flagships, one day July 9. OpenAI shipped GPT-5.6 — RSS title "Frontier intelligence that scales with your ambition" at 10:00 UTC. Same day, xAI shipped Grok 4. I couldn't pull the body copy directly — openai.com/index pages were behind a Cloudflare challenge when I checked August 11, and x.ai was blocked outright. So I'm going on RSS titles plus my own cross-check from the July 20 brief. But the dates are solid: both landed July 9. The more telling product detail landed a few hours later: "GPT-5.6 is now the preferred model in Microsoft 365 Copilot" — 13:00 UTC July 9, also via OpenAI's RSS. That one line matters more than the model card. Customer stories are press releases. A Copilot preference is distribution. I don't have seat counts or rollout percentages from the RSS alone, but if you're counting where enterprise tokens will actually flow, that's the signal. July 8 set the table. OpenAI's "Introducing GPT-Live" hit at 00:00 UTC, and Mistral shipped Robostral Navigate — an 8B model — the same day. GPT-Live is title-only in my verification (body blocked), so I can't tell you pricing or limits. Robostral Navigate I can only verify secondarily via my July 20 cross-check, which confirmed it against Mistral's feed. Treat both as shipped, details thin. ### Three handshakes The enterprise adoption stories came as a cluster: * **July 7 — MUFG aims to become AI-native with OpenAI** * **July 7 — Australian Payments Plus moves faster with ChatGPT and Codex** * **July 10 — Deutsche Telekom is rewiring telecommunications with AI** All three via OpenAI's newsroom RSS, all three customer-story format. That's the pattern to notice. Banks and telcos going "AI-native" in a blog post isn't deployment. It's intent plus a logo. I've written before that enterprise AI buying in 2026 is gates and routers, not bigger autocomplete. These stories don't prove that — they prove marketing happened. Until I see seat counts, ARR, or "we moved X workflow to Y model in production," file them as top-of-funnel. The Copilot line is the exception. That's not a logo slide, it's a default. ### The leak that actually moved the enterprise conversation The most-engaged AI story at the end of the week wasn't a launch. It was a warning. July 12: "What xAI's Grok Build CLI sends to xAI: A wire-level analysis" — 539 points on HN at 01:09 UTC. July 13: "Grok uploaded my user directory to xAI's servers" — 512 points at 13:39 UTC. Same weekend, "New Flagship Grok Voices" landed. I couldn't fetch the xAI primary source, and I won't relitigate the wire-level claims without a reproducible trace. But the shape is what matters for enterprise adoption: a CLI allegedly uploading full repos and `.env` files to GCS. Same week three enterprises announced they're going all-in, the tooling showed why allowlists, scoped creds, and per-tool gates aren't optional. That juxtaposition is the enterprise takeaway from this week. ### What I'd do Monday 1. **Route around the flagships.** If you're on M365, test where Copilot's GPT-5.6 preference changes your workflow vs. where you still need to call the API directly. Defaults are strategy. 2. **Don't benchmark press releases.** Ask MUFG-style announcements for the follow-up: how many seats, which workflow, what moved off pilot. 3. **Audit your CLIs before your models.** If you let any agent CLI touch a repo with secrets, you don't have an agent problem, you have a credential-isolation problem. Fix that before you upgrade models. === # Open Source Got Physical This Week URL: https://brianbest.com/blog/2026-07-20-ai-week-w05c-opensource Date: 2026-07-20 Tags: open-source, ai-weekly, research Summary: July 7-13, open source skipped the chatbot leaderboard and shipped embodied AI, medical data plumbing, and a sovereign European MoE. Closed labs had nothing open to show July 7-13. The interesting stuff all came with a GitHub link and it all got physical — depth, flow, and pixels you can actually use. No new flagship weights. Just better ways to read the world. ### ReChannel: steal the diffusion model's eyes ReChannel (arXiv:2607.06553, July 7, MIT) asks a good question: what if a text-to-image DiT already learned geometry and we're just decoding it wrong? They freeze FLUX-Klein, throw away the VAE decoder, and add a shared token-local linear head — about 33K parameters, no convolutions — to read dense predictions straight off the tokens. Patch-to-token-to-patch stays, RGB channels become depth, normals, matting instead. It sets new SOTA on trimap-free matting, KITTI depth, and referring segmentation, and it's 2.48x faster than the edit-plus-decode alternative at the same 4B backbone. Lightweight LoRA per task, frozen backbone. ```bash pip install -r requirements.txt python infer.py --image demo.jpg --tasks depth,normal,matting --out out.png ``` Caveat in the README: this is an inference demo, not the eval pipeline. Training code promised for late July. Still, the idea lands — generative priors as perception backbones without paying the decoder tax. ### RynnWorld-4D and Robostral Navigate: RGB is enough Two embodied drops, same thesis: you don't need expensive sensors if your world model is better. Alibaba's RynnWorld-4D (arXiv:2607.06559, July 7) co-generates RGB + depth + optical flow — RGB-DF — in one diffusion process from a single RGB-D image and a language instruction. Plus Rynn4DDataset 1.0 with 254.4M frames, and RynnWorld-4D-Policy, an inverse dynamics head that reads the 4D latent in one forward pass to output robot actions at 9 Hz+ without multi-step denoising. Bimanual manipulation, closed-loop, synced geometry and motion. Mistral's Robostral Navigate landed July 8 — 8B, 76.6% on R2R-CE with a single RGB camera. No depth, no LiDAR. That's the only verifiable embodied number that week with receipts, and it matters because it says indoor navigation doesn't need a hardware stack. ### MedPMC: the boring pipeline that actually helps Yale BIDS-Chen Lab's MedPMC (arXiv:2607.07673, July 8, Apache 2.0) is data plumbing, which is why you'll actually use it. Five stages: PubMedBERT screening, ViT multi-panel detection, YOLOv10 separation, InternVL2.5-4B-MPO caption alignment, ViT medical classification. Applied to 6.1M PMC articles it yields MedPMC-11M — 11M image-text pairs, 95.3% medically relevant versus 19.7% in the prior PMC dataset. Their MedPMC-CLIP gains 7.1 points average zero-shot AUC across 26 benchmarks and 11 specialties, despite using less than half the pairs. Code, corpus, benchmarks, and models are all out: - `Yale-BIDS-Chen/MedPMC` on GitHub - `Yale-BIDS-Chen/medpmc-11m-dataset_jun24_baseline` on Hugging Face If you work in medical multimodal, fork the pipeline. It's continuously updatable by design. ### Soofi S: Europe tries sovereign Soofi S 30B-A3B (arXiv:2607.09424, July 10) is the political one. 30B total, 3B active MoE, hybrid Mamba-Transformer for near-constant KV-cache as context grows. 27T tokens, German deliberately up-weighted. Built end-to-end on Deutsche Telekom's Industrial AI Cloud in Munich. The claim: best aggregates among 17 open base models, and ahead of Olmo 3 32B and Apertus 70B on code in both English and German. The promise: weights, intermediate checkpoints, per-source data accounting, and training/eval code under "highly permissive, open-access terms." As of writing that's still a promise — no live HF repo verified. Treat it as open-weight-plus-transparency until the license file lands. But if they ship what the paper says, it's the first sovereign European foundation model with real artifact depth. ### What I'd pull Monday Grab ReChannel and try it on your own depth/matting task before you train anything. If you're in robotics, compare RynnWorld-4D's RGB-DF outputs to whatever depth estimator you're paying for now. And if you're in medical, don't train another CLIP — run MedPMC's curation on your own corpus first. This was the week open source stopped chasing chat and started seeing. === # This Week in AI Policy: Steering Disclosure, an EU Omnibus, and a White House Gate URL: https://brianbest.com/blog/2026-07-20-ai-week-w05d-policy Date: 2026-07-20 Tags: ai-policy, regulation, industry-moves Summary: For July 7-13, the FTC said hidden output steering could be deception, Brussels gave a final green light to its AI Act Omnibus, and OpenAI's next frontier rollout reportedly routed through a White House review. If you measure regulation by laws passed, last week was quiet. If you measure it by what companies will have to disclose and who has to approve a launch, it was loud. For July 7-13, three moves set the tone: Washington reframing hidden prompting as deception, Brussels softening timelines while hardening transparency, and a frontier launch that reportedly needed a voluntary national-security sign-off first. ### Undisclosed steering might be deception The FTC pushed a new framing July 9: hiding how an AI system is steered could violate federal law. That's the read from Spencer Fane, Inside Privacy (Covington), and JD Supra — all indexing a proposed FTC policy statement on AI accuracy and "output steering" with a comment period opening around July 9. JD Supra's headline put it bluntly: "When AI Has a Hidden Agenda." The theory matters. Under Section 5, a deceptive practice is one that's misleading to a reasonable consumer. The FTC's take, per those write-ups, is that system instructions that tilt tone, politics, or product recommendations — without telling the user — could clear that bar. That's not a model capability issue. It's a product design choice becoming a consumer-protection issue. Industry pushback is already baked in: steering *is* how you do safety and alignment. The tension is obvious. If every safety-tuned prompt needs disclosure, you either under-tune or you over-disclose. I couldn't fetch the FTC release directly — ftc.gov was returning a WAF block when I tried — so treat the docket number and comment deadline as "go verify on ftc.gov and the Federal Register" before you cite them. ### Brussels: labels now, high-risk later The EU Council gave a final green light to the Digital Omnibus on AI — the amendment package to the AI Act — around July 8-9. Freshfields' "AI Act unpacked #34" and a JD Supra client alert both flagged it as final. What's in it, per the secondary read: transparency obligations (Article 50 — chatbot disclosure, deepfake labeling) go live soon, while harder high-risk obligations get pushed. Tech Times framed the same week as "Chatbot Rules Live, High-Risk AI Delay Now Binding Law" dated July 11. I couldn't pull the Freshfields PDF past Cloudflare, and the Council press release didn't render either, so I don't have the article-by-article redline. But the political signal is clear. Brussels is trying to be business-friendly and enforcement-ready at once — cheap-to-implement labels now, expensive hiring/biometrics/migration rules later. If you ship content tooling in Europe, your watermarking work just got a hard date. If you ship high-risk systems, you bought time, not a pass. ### A White House gate for GPT-5.6 On the industry side, OpenAI published new "government and national security partnership" principles July 8 — indexed by Google News RSS with that exact title. The same week, a cluster of headlines claimed the U.S. government cleared OpenAI's most advanced model yet after a 12-day review under a voluntary framework — variously called GPT-5.6 — with launch timing slipping to accommodate it. I couldn't fetch the OpenAI principles page directly (Cloudflare challenge) and I didn't get a model card to confirm version strings, so keep the "reportedly" on the 12-day gate and the .6 naming. But the pattern rhymes with last week's Brussels story: no formal U.S. AI Act, yet frontier launches are routing through an informal national-security clearance. Politico added a coda July 11: the Trump administration's AI promotion program launch "underwhelmed" — per their headline, "Not where they hoped it'd be." Substance was light behind the Cloudflare gate, but the vibe fits: Washington wants to be seen promoting AI, not just policing it, and the rollout hasn't landed yet. UNICEF also launched a coalition for children's rights in the age of AI July 8. Less regulatory bite, more norm-setting — but it shows where policy is being written while federal frameworks lag: at the institution level. ### What I'm watching 1. **Disclosure specs.** Does the FTC finalize steering-disclosure as deception, and where's the line between safety tuning and hidden persuasion? Watch the actual policy statement PDF, not the JD Supra summary. 2. **Omnibus text.** Watch EUR-Lex and the Council press room for the Omnibus final text — specifically what's delayed, for whom, and what "transparency now" actually requires in the watermark. 3. **Voluntary gates as de facto licenses.** If a 12-day White House review becomes the norm for frontier releases, that's licensing without a license law. Watch whether other labs follow the same path and whether that voluntary framework gets published. Back next week with more policy and industry moves. This week, transparency stopped being a virtue and started being a compliance field. === # Sonnet 5 Gets Cheaper, Leanstral Gets Proofs, GeneBench Gets a Blank Slate URL: https://brianbest.com/blog/2026-07-13-ai-week-w06a-models Date: 2026-07-13 Tags: ai-models, benchmarks, weekly-roundup Summary: June 30 to July 6 gave us a cheaper Sonnet that chases Opus, a 6B-active prover that saturates miniF2F, and a new bio benchmark with no scores yet. If you wanted a bigger number last week, you got a cheaper one instead. June 30 to July 6 didn't bring a new frontier from OpenAI or Google. It brought Anthropic making its middle tier act like a frontier, Mistral proving math finds real bugs, and OpenAI launching a benchmark without results. More useful than another +2% on MMLU. ### What shipped **Claude Sonnet 5 — June 30.** Anthropic's pitch is one line: "most agentic Sonnet model yet." It's now default for Free and Pro, available to Max/Team/Enterprise, API id `claude-sonnet-5`, at $2 input / $10 output per million tokens. That's permanent — the charts still show $3/$15, but the text says $2/$10 stuck. The claim is cost-performance, not peak. Anthropic says Sonnet 5 is close to Opus 4.8 ($5/$25) but cheaper, and a strict improvement over Sonnet 4.6. The named harnesses are BrowseComp and OSWorld-Verified, where Sonnet 5 "covers a much wider range of cost-performance options than Opus 4.8" and can match Opus at higher effort. I couldn't verify the full table — it's an image on the post, no HTML numbers. So I'm not quoting GPQA or SWE-bench deltas here. The safety note I could verify: lower rate of undesirable behaviors than 4.6, and "much lower ability to perform cybersecurity tasks than our current Opus models." That's deliberate tiering — capability *down* by design. **Leanstral 1.5 — July 2.** The opposite of a chat flagship. 119B total, 6B active, Apache-2.0, on Hugging Face and a free API. Three-stage training with CISPO RL, using a multiturn Lean loop and a code-agent loop with SafeVerify. The numbers are in-text for once: * miniF2F: 100% on validation and test. Saturated. * PutnamBench: 587/672 solved, scaling from 44 at 50k tokens to 493 at 1M to 587 at 4M. Edges Seed-Prover 1.5 high by 7 problems at ~$4/problem vs ~$300+ for Seed-Prover. * FATE-H/X: 87 and 34 problems solved — SOTA on both by Mistral's count. The part that matters: across 57 Rust repos, the Leanstral → Aeneas → Lean pipeline flagged 47 violated properties, 11 genuine bugs, 5 previously unreported on GitHub — including a `Std.U64.MAX` zigzag overflow. That's a prover, not a chatbot. **GeneBench-Pro — June 30.** OpenAI announced it via RSS: "a new benchmark testing AI performance in genomics, biology, and scientific research using complex, real-world datasets." Companion "Inside GeneBench-Pro" same day. I couldn't pull the page body — Cloudflare-challenged on fetch — so no dataset list or baseline scores. It's a harness launch, not a result. Wait for numbers. ### What didn't ship No verified flagship or general benchmark from Google DeepMind, Meta, or xAI inside this window that I could fetch with a clean date. Their newsrooms were JS shells or Cloudflare-blocked. That doesn't prove silence — just nothing verifiable. Anthropic did redeploy Claude Fable 5 globally July 1 after export controls (applied June 12) were lifted June 30, but that's a re-ship of June 9 weights with a new 99%-block classifier — a policy story, not a new model. ### My take This was the week labs sold efficiency, not intelligence. Sonnet 5 says you don't need Opus if you'll pay for effort. Leanstral says you don't need dense 100B+ if you'll pay for verification loops. Both pitch cost per solved task. If you're routing today: try `claude-sonnet-5` where you reached for Opus and measure tool use, not chat. If you touch formal methods or Rust, Leanstral is the only open-weights model this window with receipts — and at $4 a Putnam problem, you can afford to let it think. === # This Week in AI Products: $3.5B Says Enterprises Won't Self-Serve URL: https://brianbest.com/blog/2026-07-13-ai-week-w06b-products Date: 2026-07-13 Tags: ai-weekly, enterprise-ai, product-launches Summary: June 30 to July 6 didn't ship a frontier model. It shipped the deployment orgs — AWS and Microsoft committed $3.5B to embed engineers, plus a science workbench and cheaper plumbing from Google and X. If you were waiting for GPT-5.5 or Gemini 3.5 to drop last week, you got something more telling instead. No frontier. Just $3.5B worth of people to install the last frontier for you. June 30 to July 6 was the week hyperscalers stopped pretending enterprises will self-serve agents. That's the story. ### The $3.5B forward-deployed arms race On June 30, AWS launched a $1B Forward-Deployed Engineer org — internal headcount, not a JV — led under VP Frontier AI Francessca Vasquez. Pitch: engineers embed at your site, leave you with "agentic systems running in their own AWS environment" plus the skills to run them. Tech gets reused across clients. Two days later, July 2, Microsoft answered with Microsoft Frontier Company: $2.5B, 6,000 industry and engineering experts under Commercial CEO Judson Althoff. Early names: London Stock Exchange Group, Unilever, Land O'Lakes, Accenture. Althoff even rejected the label while doing the thing: "This goes beyond what has been labeled as Forward-Deployed Engineering." It *is* Forward-Deployed Engineering. Palantir pioneered it. OpenAI and Anthropic had already done FDE JVs with private equity earlier this year. Now AWS and Microsoft are productizing bodies. I get why. I don't love it. This is systems integration with better branding — you're buying outcomes, not API calls. If you sign, ask what stays when the FDEs leave, and who owns the custom tooling. ### Claude Science bets on workflow, not weights Same day as Sonnet 5 (which I covered in [the models post](./2026-07-06-ai-week-w06a-models.md)), Anthropic shipped Claude Science — June 30, 10:00 UTC. Not a new model. Same Claude models including Opus 4.8, wrapped as a workbench. One manager assistant, sub-agents underneath, 60+ scientific databases, prebuilt toolkits for genomics, protein structure, chemistry. The whole point is you stop bouncing between databases, pipelines, and notebooks. Positioning is honest: "not a more capable model for biology." That's refreshing. Enterprise science teams don't need another chat wrapper. They need someone to wire the data. This is Anthropic trying to be that wire. ### Plumbing got cheaper and easier Two smaller launches tell you where iteration cost is going: **Nano Banana 2 Lite — June 30.** Google's faster image generator: 4-second latency, $0.034 per 1,000 images. Versus the original Nano Banana, the February Nano Banana 2, and the expensive Pro. This is the workhorse SKU for when you're generating thousands of variants. At that price you stop debating per-image cost and start debating throughput. **Gemini Spark on Mac (beta) — July 1.** Agentic assistant inside the Gemini desktop app. Sorts files, uses Mac files as sources for Workspace docs (invoices → sheets), tracks topics in real time, hooks into Tasks and Keep. Beta is Google AI Ultra only, US only, with phone-to-Mac handoff "soon." It's parity play against Claude Desktop and Copilot, but the Tasks/Keep integration matters if you live in Google Workspace. **X hosted MCP server — June 30.** X now hosts its own Model Context Protocol server so Claude, Cursor, Grok Build and other MCP apps connect via your X account perms — no self-hosted auth dance. No new API methods, just easier exposure to search, read, and trend analysis. It joins GitHub, Slack, Notion, Stripe, and Salesforce with official hosted MCP servers. MCP went from dev toy to integration layer this week. ### Inference arbitrage is real Underneath all that, enterprises voted with bookings: - **Etched — June 30:** $1B in contract orders for "frontier inference clusters" (Sohu chips + racks + software), still in customer testing. $800M raised total including a $500M round at $5B post led by Stripes. TSMC-made. Investors include Karpathy, Hinton, Fei-Fei Li. - **Together AI — July 1:** $800M Series C at $8.3B (up from $3.3B), led by Aramco Ventures and Vista, with Nvidia in tow. $1.15B in annual bookings last quarter, thousands of paying customers including Cursor, Cognition, Decagon. Thesis: rent GPUs and run open models cheaper than paying frontier token premiums. Both are forward-looking — contracts and bookings, not GAAP revenue — but directionally they say the same thing as Sonnet 5's $2/$10 pricing: enterprises are hunting for cheaper inference per solved task. ### What I'd do Monday 1. **If your SI calls about FDE, ask for exit criteria.** What runs without them on day 91? Get it in writing. 2. **Try Claude Science on one real literature-to-experiment loop.** Don't evaluate it on chat — evaluate it on how many database hops it saves. 3. **Price your retrieval loops.** If you're doing multi-turn search, benchmark a cheap retriever or Nano Banana Lite-class economics before defaulting to Opus/Sol for every turn. === # Open Source Compiled Away the API Call URL: https://brianbest.com/blog/2026-07-13-ai-week-w06c-opensource Date: 2026-07-13 Tags: open-source, ai-weekly, research Summary: June 30-July 6 had no new open flagship, but it shipped a compiler that turns prompts into local adapters, a 6T-token VLM data playground, and a world model that freezes itself. No open flagship dropped June 30-July 6. Nothing to download and brag about on parameter count. What did ship was more interesting — research that tries to make the frontier cheaper to run, not bigger to train. ### PAW: compile the prompt, ship the weights Program-as-Weights ([arxiv:2607.02512](https://arxiv.org/abs/2607.02512), featured July 3) is the one I'd actually try first. The pitch is simple: stop calling an API for fuzzy functions like "flag urgent logs" or "fix broken JSON." Compile them. You give it plain English, a 4B compiler trained on their 10M-example FuzzyBench emits a tiny adapter for a frozen interpreter. Two flavors: `paw-4b-qwen3-0.6b` (594MB base, ~22MB per program) and `paw-4b-gpt2` (134MB base, ~5MB per program). After compilation, no API key, no internet. ```python import programasweights as paw program = paw.compile("Fix malformed JSON: repair missing quotes and trailing commas", slug="json-fixer") fn = paw.function(program.slug) fn("{name: 'Alice',}") # '{"name":"Alice"}' ``` The authors' claim: the 0.6B Qwen3 interpreter with a PAW program matches direct prompting of Qwen3-32B at about 1/50th the memory, running 30 tokens/s on a MacBook M3. That's author-reported, not independently verified yet, but the artifact is real — `pip install programasweights`, Apache-2.0, runs offline. If it holds up, it's not a better LLM. It's a way to stop using one at runtime. ### DataComp-VLM: finally a playground for VLM data DataComp for VLMs ([arxiv:2606.28551](https://arxiv.org/abs/2606.28551), featured July 6) does for vision-language what DataComp did for CLIP. No new architecture. Just a controlled benchmark for data curation. The corpus is big: 160 datasets spanning captions, interleaved documents, text-only, and instruction data — 6T multimodal tokens total. Plus ready-to-train shards like `DCVLM-Baseline` (200B tokens, 104M samples, decontaminated and shuffled) and baseline models from 1B to 8B ([mlfoundations/dcvlm](https://github.com/mlfoundations/dcvlm)). The counterintuitive finding: mixing beats filtering. Their instruction-heavy mixture (70% instruction data) scales better than caption-heavy mixes, and the gap widens with size. Their 8B hits 63.6% on the 33-task core suite with just 200B tokens. If you've been hand-tuning VLM data, stop. Start from their pools and ablate. That's the point — a reproducible lever instead of private data lore. ### Two horizon bets Two papers argue scaling isn't about parameters right now. **Orca** ([arxiv:2606.30534](https://arxiv.org/abs/2606.30534), BAAI) is a world foundation model trained with Next-State-Prediction — not next token or next frame. 125K hours of video plus 160M language-described events learn a frozen world latent; only lightweight decoders train for downstream text, image, or action tasks. Code and 4B weights at [BAAI/Orca-4B](https://huggingface.co/BAAI/Orca-4B) (weights landed July 14, just outside the window). **Agents-A1** ([arxiv:2606.30616](https://arxiv.org/abs/2606.30616), 35B MoE, 3B active) claims trillion-parameter performance by scaling horizon, not size — 45K-token average trajectories via a knowledge-action-verifier loop. Apache-2.0, with quantized variants shipped July 2 and a 4B version July 14. Benchmark table is author-reported, so treat it as a direction, not a score. And for anyone who actually ships robots: **Embodied.cpp** ([arxiv:2607.02501](https://arxiv.org/abs/2607.02501)) is trying to be `llama.cpp` for VLA/WAM models — one C++ runtime, GGUF weights, CPU/CUDA/NPU, with multi-rate control loop support. Initial models pi0.5/HY-VLA/LingBot-VA plus GR00T N1.7 and Cosmos3-Nano. They claim -30% batch-1 latency, which matters when your loop is closed. ### What I'd pull Monday If you have a fuzzy function hitting an API in a hot path, compile it with PAW and benchmark the local adapter. If you're training VLMs, don't curate from scratch — fork DataComp-VLM's Baseline shards and test the mixing ratio first. And if your world model work is stuck on token prediction, read Orca's state-transition framing before you scale parameters again. This week didn't give us a bigger open model. It gave us cheaper ways to not need one. === # This Week in AI Policy: Disclosure Fights, an EU Softening, and Bodies Over Models URL: https://brianbest.com/blog/2026-07-13-ai-week-w06d-policy Date: 2026-07-13 Tags: ai-policy, regulation, industry-moves Summary: For June 30 to July 6, the FTC framed hidden AI steering as deception, Brussels formally adopted its AI Act omnibus, and AWS bet $1B that deployment — not training — is the moat. For June 30 to July 6, Washington didn't pass an AI law. It reframed an old one, Brussels softened a new one, and the hyperscalers wrote policy by procurement. That's the whole week in regulation. No grand AI Act. Just disclosure risk, delayed deadlines, and bodies on site. ### The FTC says quiet steering might be deception Around July 1-2, the FTC pushed a proposed policy statement on AI accuracy and "output steering" — at least that's the read from a cluster of legal press I could actually fetch. Bloomberg Law (July 1): "Companies Following AI State Laws Risk Enforcement, FTC Says." Law360 same day: "FTC Says Distorting AI Outputs To Follow State Laws Won't Fly." MLex: "US FTC proposes policy statement addressing AI accuracy." By July 2, PYMNTS, Hoodline, and Forbes were all on the same thread: undisclosed bias safeguards could violate consumer law if they tilt answers without telling you. The theory is straightforward Section 5 logic: a practice that misleads a reasonable consumer is deceptive. If your system prompt quietly nudges tone, politics, or product picks and the user doesn't know, the FTC's framing says that could count. I couldn't verify the primary. ftc.gov was WAF-blocked when I tried last week and the RSS hits this week don't link a PDF or Federal Register docket. So treat the comment period and docket number as "go check ftc.gov yourself" before you cite them. But the direction is real — alignment work just became a disclosure problem, not just a safety problem. If that holds, every steering instruction becomes a disclosure doc. That's a headache if you also need steering for safety. ### Brussels: labels now, hard stuff later June 30, Digital Watch Observatory ran it as "EU approves simplified AI rules under Omnibus VII." July 1, Lexology: "EU AI Act: AI Omnibus formally adopted." EU Today framed the same vote as testing Brussels' promise to cut red tape without weakening oversight. The gist from secondary write-ups: transparency obligations — chatbot disclosure, deepfake labeling, Article 50 plumbing — stay on the near-term clock. High-risk obligations (hiring, biometrics, migration) get pushed. IT Pro noted on July 1 that founders still complained about "regulatory friction" despite the simplification. Which tracks. I couldn't pull the Council press release or EUR-Lex text directly — dig.watch 404'd on fetch, Lexology hit a Cloudflare challenge — so I don't have the article-by-article redline. But the politics are clear: cheap-to-ship labels now, expensive compliance later. If you ship generative tools in Europe, watermarking just got a real deadline. ### $1B for engineers, not tokens June 30, AWS launched a $1B Forward-Deployed Engineer org. About Amazon's headline was plain: "AWS invests $1 billion to embed AI forward deployed engineers with customers." Reuters, CNBC, and TechCrunch all matched the date. Pitch: AWS engineers embed at your site, leave agentic systems running in your AWS account, tech gets reused across customers. Lead is VP Frontier AI Francessca Vasquez per earlier coverage. I couldn't fetch the AboutAmazon or AWS blog body — both rendered as JS shells and Next.js 404s on curl — so I'm quoting titles and syndicated dates, not body copy. Two days later, my products post noted Microsoft answering with Microsoft Frontier Company — $2.5B, 6,000 experts under Judson Althoff. I didn't get a matching Microsoft press release in RSS to reconfirm that number this week, so keep the "reportedly" on the Microsoft side. The pattern doesn't need it though: Palantir pioneered FDE, OpenAI and Anthropic did JVs with PE earlier this year, now the hyperscalers are productizing headcount. That's policy by procurement. Whoever embeds writes your data plumbing and retention terms. A smaller policy signal same window: Anthropic redeployed Claude Fable 5 globally July 1 after U.S. export controls applied June 12 were lifted June 30 — same June 9 weights with a new 99%-block classifier per my models post that week. No BIS notice fetched to confirm, but it rhymes: model launches now track export calendars, not just evals. ### What I'd watch 1. **The FTC PDF.** If steering disclosure finalizes, where's the line between safety tuning and hidden persuasion? 2. **The Omnibus text.** Watch EUR-Lex for what's delayed vs. what's enforceable now — especially the watermark spec. 3. **FDE contracts.** If you sign, ask what stays on day 91 and who owns the custom tooling. No law passed. But the compliance field moved anyway. === # No Open Flagship This Week — Just Gated Access URL: https://brianbest.com/blog/2026-07-06-ai-week-w07a-models Date: 2026-07-06 Tags: ai-weekly, frontier-models, benchmarks Summary: June 23-29 had no open flagship launch — instead GPT-5.6 was throttled at the White House's request and Mythos went gated to 100+ orgs. If you were waiting for a new flagship to actually ship last week, it didn't. For June 23-29 I checked the usual newsrooms — OpenAI, Anthropic, Google, Meta, xAI, Mistral — and couldn't verify a single GA flagship release with a date in that window. No GPT-5, no Claude 5, no Gemini 2.5 Pro, no Grok 4. What I *could* verify was weirder: the biggest models existed, they just didn't ship openly. ### GPT-5.6 went gated The headline model of the week was OpenAI's GPT-5.6 — and the story was that you couldn't have it. TechCrunch reported June 25 that the White House asked OpenAI to slow-roll the release over safety concerns, pushing it to a select partner group instead of general availability. OpenAI followed June 26 with a post saying it had limited the GPT-5.6 rollout — including a variant tagged "GPT-5.6 Sol" — and added: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." I couldn't pull the primary OpenAI blog directly to cross-check — it returned a Cloudflare challenge on my fetch — so I'm relying on TechCrunch's reporting here. But the signal is consistent: the frontier was ready, distribution was the bottleneck. ### Mythos went the other way Anthropic's Mythos 5 had the opposite distribution story. Also per TechCrunch, the Trump administration cleared Mythos for use by 100+ US companies and agencies on June 26-27, including their non-American employees. Not a new weights drop — a permissions change. And the second-order effect hit immediately. On June 27 TechCrunch noted Asian startups launching "Mythos-like" models explicitly pitched as without export-ban risk, with the deck warning U.S. labs "may never recover this enormous market." We've seen this movie with chip controls. Now it's happening at the model layer. ### The actual hardware that shipped: Jalapeño The one concrete artifact that *did* ship was infrastructure. OpenAI unveiled its first custom inference chip June 24 — Jalapeño, built with Broadcom, "designed specifically for the unique needs of OpenAI's inference systems." No model card, no eval table. But if you're building agents that loop, this matters more than a point on MMLU. Inference cost and latency are what's gating deployment, and a custom chip is a flagship move even if it doesn't have a leaderboard score. ### Benchmarks: nothing to score And that's the last piece — there was nothing to score. I ran the full June 23-29 TechCrunch dump (100 posts) for MMLU, GPQA, HumanEval, SWE-bench, MATH, AIME, LiveBench — zero hits with numeric results. No verified eval sheet from any lab in that window. The closest leaderboard-adjacent item was a June 29 piece noting LM Arena is now a "$100M business" — no scores, no movement table. That's not stagnation. It's a coordinated pause. When access is gated by government request, you don't publish a chart bragging about the capability you're not allowed to ship. ### What I'd take from a gated week This week wasn't about which model is smartest. It was about who gets to use it. If you're planning around this: 1. **Don't wait for an open launch.** Build your evals assuming tiered access — partner-only flags, government-cleared lists. Your routing logic needs a "not available here" branch now. 2. **Watch inference economics, not just intelligence.** Jalapeño tells you where OpenAI thinks the edge is — cost per token at scale, not another reasoning bump. 3. **Track the clones.** If Mythos-like models fill the export-ban gap, you'll have a bifurcated market with different models behind different borders. Test there too. The labs didn't need a bigger number last week. They needed permission to ship. That's the benchmark that mattered. === # This Week in AI Products: Gated Models, Gray Beards, and Free Personalization URL: https://brianbest.com/blog/2026-07-06-ai-week-w07b-products Date: 2026-07-06 Tags: ai-weekly, enterprise-ai, product-launches Summary: Frontier models went gated by government, Ford brought back 350 veteran engineers after AI fell short, and Google made Gemini personalization free — the same week South Korea bet $1T on physical AI. The product story June 23-29 wasn't a faster model. It was who gets to use the model at all. For the last year enterprise AI was "call the API." This week it became "prove you're trusted, then call the API." That's a different distribution motion — and Ford's reversal is the other side of the same coin. ### Frontier models go gated Two staggered launches defined the week. June 25-26, OpenAI agreed to delay and stagger GPT-5.6 after a Trump administration request — The Verge (June 25, 36 pts on HN) and Bloomberg both reported it, and the Washington Post headline June 26 put it bluntly: "U.S. government will vet users of its latest AI model" (1,184 pts on HN). I couldn't render the Post or Reuters bodies directly — both were JS-gated — so I'm citing the HN-indexed headlines and metadata, not the vetting criteria. Same window, Anthropic's Mythos went to "trusted" U.S. organizations under government clearance. Semafor's exclusive is cleanly timestamped — `2026-06-26T22:06:03Z` by Reed Albergotti and Ben Smith: "Exclusive: US releases powerful Anthropic model Mythos to some US companies" (553 pts on HN). Again, the body was JS-hydrated and I could only verify the JSON-LD, not the allowed-use details. If you're buying enterprise AI, this matters more than parameter counts. Access *is* the product now. Budget for verification, monitoring, and waiting in line alongside inference costs. ### Ford's gray-beard correction June 28, TechCrunch (`2026-06-28T19:05:39Z`, fully extractable) reported Ford rehired about 350 veteran engineers — "gray beards" — after automated quality systems disappointed. COO Kumar Galhotra told journalists Ford had been "relying more and more on automated quality systems" with poor results. VP Charles Poon: "Mistakenly we thought that by just introducing artificial intelligence and ingesting the design requirements that we had, that that would produce a high-quality product." They're not killing AI. They're using the veterans to train younger staff and reprogram the tools. CEO Jim Farley claimed it contributed "hundreds and hundreds of millions of dollars" in lower warranty and recall costs, and Ford took the top mainstream spot in this week's JD Power Initial Quality Survey. That's the enterprise adoption lesson I keep seeing: AI doesn't replace senior judgment, it gets layered on top of it. The ROI came from re-adding humans, not removing them. ### Free personalization vs. $1T sovereign bet June 29 had two opposite product plays land the same day. Google made Gemini's personalized image generation free for eligible U.S. users — TechCrunch `2026-06-29T20:12:59Z`: the bot now creates images from your interests and data in connected Google apps. Distribution via data you already gave Google. Bottom-up and sticky. Hours later, Ars Technica `2026-06-29T21:09:43Z`: "South Korea to spend $1T on more memory chip production and humanoid robots," targeting physical AI leadership and commercial humanoids by 2028. Al Jazeera carried the same announcement, and CNBC the same day noted Baidu's chip arm Kunlunxin targeting a $50B Hong Kong IPO. Top-down industrial policy. One gives you free images Monday. The other decides who can make the chips those images run on in 2028. Both are enterprise adoption, just at different altitudes. ### The plumbing nobody noticed June 23, the Linux Foundation announced intent to launch an Agent Name Service (ANS) for trusted identity for AI agents — only 6 points on HN, page behind HubSpot JS so I couldn't pull the spec. Boring, and probably the most important launch of the week. You can't deploy agents at scale if they can't prove who they are to each other. ### What I'd do Monday 1. **Assume gated rollout is normal.** If you're planning on GPT-5.6 or Mythos, talk to your legal and compliance team now about what "trusted" will require. 2. **Audit where you removed senior review.** Ford's fix wasn't anti-AI — it was pro-review. Put the gray beards in the loop that trains the tool. 3. **Try the free Gemini personalization — with boundaries.** It's a good test of personal-data flywheel. Check what app data you've actually connected first. === # While Flagships Were Gated, Open Source Shipped Inference URL: https://brianbest.com/blog/2026-07-06-ai-week-w07c-opensource Date: 2026-07-06 Tags: open-source, ai-weekly, research Summary: No open flagship shipped June 23-29. Open source did — four verified drops from 0.6B to 75B, all about making models cheaper and faster to run. Flagships were gated last week. Open source wasn't. While GPT-5.6 and Mythos sat behind partner lists, the verifiable drops with a Hugging Face `createdAt` in June 23-29 were all efficiency plays. Not smarter — cheaper, smaller, faster. That's the story. ### Puzzle: compress the flagship NVIDIA's `NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B` hit June 24 ([BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16), [FP8](https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-FP8), [NVFP4](https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4)) and it's the most honest release of the week. No new pretrain. Just Iterative Puzzle compression applied to Nemotron-3-Super-120B-A12B. 75.3B total, 9.3B active. Hybrid Mamba2-MoE with Multi-Token Prediction, cut from 120.7B / 12.8B. The claim in the card: ~2x throughput on an 8xB200 node and 1M-token concurrency on a single H100 from 1 to 8 requests. Active experts per layer drop from 22 to 4-18, Mamba state 128→96. Paper: [arxiv:2607.04371](https://arxiv.org/abs/2607.04371). License is [openmdw-1.1](https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16) — commercial-use, but not Apache-2.0. If you run big MoEs, this is the pattern to study. ```python from transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained( "nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16", trust_remote_code=True ) ``` ### Small specialists that you'll actually use Two drops on opposite ends of the size spectrum, both immediately pluggable. **OlmoEarth v1.2** shipped four sizes in two days — [Nano](https://huggingface.co/allenai/OlmoEarth-v1_2-Nano) (June 24), [Tiny](https://huggingface.co/allenai/OlmoEarth-v1_2-Tiny), [Small](https://huggingface.co/allenai/OlmoEarth-v1_2-Small), [Base](https://huggingface.co/allenai/OlmoEarth-v1_2-Base) (June 25). ViT-Base is 114M params, trained on Sentinel-1, Sentinel-2, and Landsat time series. Paper: [arxiv:2605.20804](https://arxiv.org/abs/2605.20804). License: OlmoEarth Artifact License. If you touch geospatial at all, this beats wrapping a chat model. **Qwen3-ASR** landed June 26 — [0.6B](https://huggingface.co/Qwen/Qwen3-ASR-0.6B-hf) and [1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B-hf) plus [ForcedAligner-0.6B](https://huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B-hf). 32 languages, word-level alignment, Apache-2.0. Finally a permissive ASR you can fork without thinking about it. ```python from transformers import Qwen3ASRForConditionalGeneration, AutoProcessor model = Qwen3ASRForConditionalGeneration.from_pretrained("Qwen/Qwen3-ASR-0.6B-hf") ``` ### DeepSeek's inference stack and the quant wave DeepSeek dropped the other half of the inference story June 27-28. [DeepSeek-V4-Flash-DSpark](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-DSpark) (284B / 13B active) and [Pro-DSpark](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-DSpark) (1.6T / 49B active), both 1M context, FP8, MIT — plus 12 speculative decoding variants on June 28 (`dspark`, `dflash`, `eagle3` heads for Qwen3 and Gemma). Preview paper: [arxiv:2606.19348](https://arxiv.org/abs/2606.19348). NVIDIA was the distributor that week too. Alongside Puzzle they pushed quantized day-zero variants: [Kimi-K2.6-DFlash](https://huggingface.co/nvidia/Kimi-K2.6-DFlash) and [MiniMax-M3-NVFP4](https://huggingface.co/nvidia/MiniMax-M3-NVFP4) on June 23, [Qwen3.5-397B-A17B-NVFP4-V2](https://huggingface.co/nvidia/Qwen3.5-397B-A17B-NVFP4-V2) on June 29. Two other pluggable drops: [c-fast-foundationstereo](https://huggingface.co/nvidia/c-fast-foundationstereo) (14.6M, 10x faster than FoundationStereo) and Tencent's [Rosetta-inference](https://huggingface.co/tencent/Rosetta-inference) ([arxiv:2607.00293](https://arxiv.org/abs/2607.00293), Apache-2.0) for composable multimodal pretraining. Google also shipped [tabfm-1.0.0](https://huggingface.co/google/tabfm-1.0.0-jax) June 29 — zero-shot tabular foundation, JAX and PyTorch. No new intelligence frontier. Just the plumbing to actually run what we already have. ### What I'd pull Monday If you run inference at scale, benchmark Puzzle's NVFP4 against your current MoE. If you do speech, swap in Qwen3-ASR-0.6B — it's Apache-2.0 and small enough to run everywhere. And if you do anything with satellite imagery, try OlmoEarth-Tiny before you fine-tune anything bigger. This week rewarded the teams optimizing deployment, not the ones waiting for permission to ship. === # This Week in AI Policy: Brake Pedals, Chip Wars, and a Model Gate URL: https://brianbest.com/blog/2026-07-06-ai-week-w07d-policy Date: 2026-07-06 Tags: ai-policy, regulation, industry-moves Summary: For June 23-29, Washington pulled the brake pedal for robotaxis while tightening the gates on chips and frontier models — and Brussels mostly waited for August. Washington spent the last week of June doing two opposite things at once: making it easier to build a car without controls, and harder to ship a model without permission. That contradiction is the story. ### Chips and cars as industrial policy On June 24, Dutch Trade Minister Sjoerd Sjoerdsma flew to Washington to lobby against the MATCH Act (H.R.8170), per TechCrunch citing Bloomberg. The bill would extend U.S. equipment curbs to cover ASML's deep-ultraviolet immersion machines — not just EUV — cutting off a big chunk of sales to Chinese chipmakers. China is ~19% of ASML's net system sales, and ASML is the only company that can build the lithography tools at stake. The bill was introduced in April and hadn't gotten a floor vote by June 24; Sjoerdsma called the stakes "very high" for the Netherlands. A day later, Commerce showed it wasn't waiting for Congress to use trade tools. It denied Polestar authorization under the Connected Vehicle Rule, barring new Polestar sales in the U.S. because the Geely-owned brand is deemed to have Chinese software/hardware. Polestar said it can sell existing Polestar 3/4 stock and keep service running, but 94% of its Q1 retail was already outside the U.S. anyway. Sibling Volvo — also Geely-owned — had gotten the same authorization in May. Same day, different door: DOT/NHTSA proposed to strip the brake-pedal requirement from Federal Motor Vehicle Safety Standards for vehicles "designed to be driven exclusively by automated driving systems." There's a 30-day comment window. NHTSA Administrator Jonathan Morrison pitched it as tearing down "pointless barriers" under Secretary Sean Duffy's AV Framework. If finalized, it directly helps Tesla's Cybercab (no wheel, no pedals, no exemption on file) and Zoox (Amazon's robotaxi, which got a demo exemption in 2024). Waymo keeps its pedals — it runs retrofitted Jaguars. Deregulate the chassis, regulate the supply chain. That's the pattern. ### A de facto license for frontier models? The other gate was software. On June 26, two model-release stories hit the same TechCrunch archive, both attributed to Semafor/Reuters and company posts I couldn't independently fetch (Commerce, OpenAI, and ftc.gov all threw WAF blocks on my fetch pass, so treat the names and numbers as reported, not primary-verified). First, Commerce Secretary Howard Lutnick reportedly wrote Anthropic on June 26 that "appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model" — a partial reversal of a June 12 ban on non-American access after researchers bypassed guardrails. The allowance covers 100+ U.S. agencies and companies, plus their non-American staff. Second, OpenAI reportedly limited its GPT-5.6 preview (Sol/Terra/Luna) to a "small group of trusted partners whose participation has been shared with the government," at government request. Pricing was listed as Sol $5/$30 per million input/output tokens, Terra half that, Luna $1/$6. OpenAI's quoted line: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." Read together with the administration's voluntary 30-day pre-release review EO, that's what former White House adviser Dean Ball called a voluntary framework acting like an involuntary license. Whether you call it safety review or prior restraint depends on which side of the API you sit. ### Industry and the quiet in Brussels Amazon added a $13B AI infrastructure commitment in India on June 25 — the one clean industry number with a date stamp that week. What didn't happen matters too. I found no verifiable EU enforcement order, new U.S. AI law, or FTC AI action dated June 23-29. The AI Act's next real teeth — GPAI obligations and the August 2, 2026 general application date — were still a month out, and Politico wouldn't report the AI Office's full fining/pull-from-market powers until July 31. FTC press releases show a gap for that week; the next entries are July 8 onward. No copyright ruling in the AI docket verified inside the window either. That quiet is probably real, not just a fetch miss — but flag it as low-confidence until EUR-Lex and the Federal Register render cleanly. ### What I'm watching 1. **MATCH Act vs. ASML.** Does a chip-equipment ban get folded into a must-pass package, or does Dutch lobbying carve out DUV? That's your tell for how far Washington will push allies on China controls. 2. **Brake-pedal comments.** Watch who files in the 30-day window. The AV lobby wants hardware freedom; safety groups will want performance data in exchange. 3. **Trusted-partner as template.** If 100-org allowlists become the norm for Mythos/Sol-class releases, frontier deployment just got a government gate without a licensing law on the books. Back next week. This one was Washington choosing where to be hands-off and where to be hands-on. === # Flagship Models Took the Week Off URL: https://brianbest.com/blog/2026-06-29-ai-week-w08a-models Date: 2026-06-29 Tags: ai-models, benchmarks, weekly-roundup Summary: For June 16-22 I couldn't verify a single flagship model launch or benchmark jump — the labs shipped scaffolding, not new weights. I went looking for a flagship model story for June 16-22 and didn't find one. I checked the usual newsrooms — OpenAI, Anthropic, Google, Meta, xAI, Mistral, DeepSeek — for a new base model, a version bump, or a fresh eval sheet inside that window. Nothing GA'd with a clean date I could verify. No GPT-5.x, no Claude 4.x successor, no Gemini 2.5 follow-up, no Grok 4. If you saw a headline that week claiming otherwise, it wasn't anchored to a primary post I could pull. That's two quiet weeks in a row on the flagship front. That's the signal. ### What actually shipped Nothing on the weights. What did ship was the stuff around the weights. Anthropic and Google both pushed docs and tooling updates in that window — Claude Code tweaks, Gemini API ergonomics — but not a new foundation model. Mistral was heads-down on enterprise plumbing after their May-June agent push. DeepSeek had no new public checkpoint between their late-May V4 whispers and the late-July V4-Flash beta. This is the pattern I've seen all quarter: a big model lands (or leaks), then the labs spend the next 2-4 weeks shipping the harness that makes it usable — context handling, tool calling, safety layers — before they talk benchmarks again. ### Benchmarks: nothing to table I ran the usual sweep for MMLU, GPQA, HumanEval, SWE-Bench, MATH, AIME, LiveBench with numbers attached. For June 16-22, I came up empty on verified sheets. No new leaderboard swing on LM Arena either. No pricing or context window change on the flagship SKUs I track. OpenAI, Anthropic, and Meta's news pages were largely JS-hydrated or Cloudflare-gated on fetch, so I couldn't even scrape an article list cleanly — but manual checks of their docs and the secondary coverage (TechCrunch, The Verge, Reuters) turned up no numeric eval drop to cite. I could have filled this with the X rumor mill about benchmark leaks. I'm not going to. If there's no model card, there's no score to trust. ### Why a quiet week is useful When flagships go quiet, it's tempting to call it a lull. I don't think it is. Labs are in a weird spot right now — the next jump is clearly expensive, clearly political, and clearly gating on inference cost more than raw IQ. OpenAI's Jalapeño chip reveal the following week (June 24, with Broadcom) kind of proves the point: they unveiled *silicon* before their next model. That's where the bottleneck is. So if you're deciding what to benchmark this week, don't wait for a new flagship to tell you what to test. Re-run your own task suite against what's in prod right now. Log pass rate *and* cost per completed task. Track latency at 128k context, not just accuracy at 4k. The next model will move the headlines by a point or two. The infra shipping in these quiet weeks will move whether your agent actually finishes the job. I'll keep watching the pricing pages and API changelogs — those move first, and often without a blog post. If something breaks cover, I'll call it out next week. === # This Week in AI Products: No New Model, Just New Places to Put It URL: https://brianbest.com/blog/2026-06-29-ai-week-w08b-products Date: 2026-06-29 Tags: ai-weekly, enterprise-ai, product-launches Summary: June 16-22 didn't ship a flagship model. It shipped the surfaces around it — Gemini on your shelf and in your pocket, ChatGPT on Samsung devices, and a security platform designed to make enterprise adoption boring. No flagship model shipped June 16-22. Same story I flagged in [the models post](./2026-06-22-ai-week-w08a-models.md): scaffolding, not new weights. No GPT-5.x, no Claude successor, no Gemini 2.5 follow-up with a citable model card. What did ship is more interesting. The labs competed on *where* models live. ### google put gemini on your shelf and in your pocket Two launches, same play — get Gemini onto surfaces OpenAI can't reach. June 16, the June Pixel Drop — "New features for creators, Gemini upgrades and more" (blog.google). Distribution via an update to a phone you already own. Next day, June 17, the new Google Home Speaker built for Gemini (blog.google), pegged by Wired, Forbes and Ars at $99.99 and available June 25. Primary blog.google pages were Cloudflare-gated when I checked, so I'm going off RSS plus secondaries. Still, a cheap Gemini-native speaker is Google betting voice in the home is its wedge. Then June 22, the developer piece: Interactions API, Google's "primary interface for Gemini models and agents" (blog.google). One interface for models *and* agents means less glue if you're wiring Gemini into your product. ### enterprise adoption came through incumbents Instead of asking enterprises to trust a new lab, the labs went through companies they already trust. **Samsung + OpenAI — June 21.** "Samsung Electronics brings ChatGPT and Codex to employees" (openai.com). Employee-wide, not a pilot. Bottom-up adoption — tools where work already happens. RSS didn't include headcount, so I'd wait for Samsung's full release for scale, but the signal is clear: ChatGPT and Codex as standard issue. **Nokia + Google Cloud — June 22.** Nokia will embed Gemini agents into its Autonomous Network suite (Nokia release). Top-down infra — telco networks running Gemini where no employee ever sees a prompt. **OpenAI Daybreak — June 22.** "Tools for securing every organization in the world" plus a Cyber Partner Program with Proofpoint, Darktrace and TrendAI the same day, plus "Patch the Planet" to fund open-source maintainers. Enterprise security teams don't buy models, they buy auditability. Treat this as intent plus brand trust — partner drops rarely include ship dates. And the boring stuff that actually unblocks procurement: June 18 OpenAI shipped new usage analytics and spend controls for enterprises, and June 19 Zensar launched AgentMesh for adoption at scale. ### claude code got enterprise-polished Anthropic didn't ship a model either. It shipped how the model works with your team. June 16: "How Claude Code is used in practice" (anthropic.com). June 18-19: Claude Code Artifacts now does live, shareable dashboards and interactive workspaces (via VentureBeat, DevOps.com). "My artifact" becoming "our dashboard." Primary Anthropic page was a 404 shell on fetch, so I'm citing RSS and secondaries — but the direction is consistent: less about IQ, more about sharing and staying live while someone else is looking. ### what i'd do monday No new model to evaluate is freeing. The action moved to distribution and trust — where enterprise deals actually stall. 1. **Test the new Google surfaces on a real task.** Try the Pixel Gemini features beyond the demo. Ask what the Speaker does differently with Gemini vs. Assistant. 2. **Steal Samsung's framing.** Don't pitch "AI adoption." Pitch "putting Codex where code already gets written." Set spend controls first. 3. **Don't call a docs update a model launch.** That Artifacts update will get re-headlined as a model. It's not. Track cost per completed task at 128k, not just accuracy at 4k. Quiet on the weights, loud on the plumbing. That's been the pattern all quarter — and next week's OpenAI + Broadcom silicon reveal (June 24) proves it. They're shipping the layer *under* the model before the next model itself. === # Long Context Finally Got Cheap URL: https://brianbest.com/blog/2026-06-29-ai-week-w08c-opensource Date: 2026-06-29 Tags: open-source, ai-weekly, research Summary: June 16-22 was long-horizon week in open source — a 1M-token GLM-5.2 under MIT, Baidu's constant-memory OCR, and agents that learn to scaffold themselves. Everyone wants long context. Nobody wants to pay for it. June 16-22, open source attacked both sides of that problem — one team made a 1M window stable, another made parsing it constant-memory. And a third decided the harness matters more than the model. ### GLM-5.2: a 1M window that holds Zhipu AI dropped [GLM-5.2](https://huggingface.co/zai-org/GLM-5.2) on June 16, MIT-licensed. No regional limits, no gated access. The headline is "solid 1M context" — not marketing 1M, but stably serving it on long-horizon work. Two architecture bets: **IndexShare**, which reuses the same indexer across every four sparse attention layers (2.9x fewer FLOPs per token at 1M), and an improved MTP layer for speculative decoding (+20% acceptance length). As reported: 40.5 on HLE, 99.2 on AIME 2026, 62.1 on SWE-bench Pro, and 81.0 on Terminal-Bench 2.1 (Terminus-2). Those are self-reported with their harness (temperature 1.0, top_p 0.95, 300K context for HLE-with-tools), so take them as directional. But on Terminus-2 it outscores its predecessor GLM-5.1 by 17.5 points — that's a real jump. It ships for SGLang, vLLM, and Transformers day one. If you've been waiting to run long agents without an API, this is the one to benchmark. ### Baidu Unlimited-OCR: dozens of pages in one pass Two days later Baidu shipped [Unlimited-OCR](https://github.com/baidu/Unlimited-OCR) (June 18 on GitHub, June 19 on Hugging Face, [paper June 22](https://arxiv.org/abs/2606.23050)). It builds on DeepSeek-OCR and replaces every decoder attention layer with **Reference Sliding Window Attention (R-SWA)** — constant KV cache through the entire decode. The pitch: encode with DeepSeek-OCR's high-compression encoder, then decode dozens of pages in a single 32K forward pass without memory blowing up. MIT-licensed, vLLM recipe included. ```python from transformers import AutoModel, AutoTokenizer model = AutoModel.from_pretrained('baidu/Unlimited-OCR', trust_remote_code=True, torch_dtype='bfloat16').eval().cuda() model.infer_multi(tokenizer, prompt='Multi page parsing.', image_files=['page1.png','page2.png'], max_length=32768) ``` If you parse PDFs at scale, this is the opposite bet from GLM-5.2 — not bigger windows, but cheaper windows. That's the more interesting tradeoff right now. ### Ornith-1.0: the model that learns its own scaffold [Ornith-1.0](https://huggingface.co/ornith-ai/Ornith-1.0-35B) landed June 21 — 35B MoE and 9B dense (with 397B MoE in the family), MIT, post-trained on Qwen 3.5 and Gemma 4. The idea is "self-scaffolding" RL: the model learns to generate both the solution *and* the search scaffold that produces it, jointly optimizing both. As reported: 64.2 on Terminal-Bench 2.1 vs 41.4 for base Qwen3.5-35B, 75.6 on SWE-bench Verified. Whether that generalizes beyond their harness is the question, but the framing is right — benchmarks are increasingly won by scaffolding, not just weights. ### The plumbing and the papers Two infra drops worth noting: - **Vercel eve** (June 16, [github.com/vercel/eve](https://github.com/vercel/eve)) — filesystem-first durable agent framework. `agent.ts`, `instructions.md`, `tools/`, `skills/`, `channels/`, `schedules/` as conventions. `npx eve@latest init my-agent` and you're running. If you've built agents as tangled scripts, you'll like this. - **Google TabFM** (June 16, [github.com/google-research/tabfm](https://github.com/google-research/tabfm)) — tabular foundation model, sklearn-compatible, zero-shot via in-context learning. JAX and PyTorch backends. Code is Apache-2.0, but weights are `tabfm-non-commercial-v1.0` — not open for production. Don't miss that split. On arXiv, three papers I'd actually read: - **Guava** ([2606.18363](https://arxiv.org/abs/2606.18363)) — 4B open model + careful harness (<2K simulated trajectories) rivaling proprietary on embodied manipulation. The takeaway: iterative perception-reasoning-action loops matter more than scale here. - **Moebius** ([2606.19195](https://arxiv.org/abs/2606.19195)) — 0.22B inpainting model matching 11.9B FLUX.1-Fill-Dev at >15x faster inference. 2% of the params, same quality claim — worth verifying on your data. - **EfficientRollout** ([2606.18967](https://arxiv.org/abs/2606.18967)) — self-speculative decoding for RL rollouts, no separate drafter to train. 19.6% rollout latency cut as reported, which compounds fast when you're doing RL. ### What I'd try Monday Run GLM-5.2 on your longest agent task and see if 1M is actually stable past 200K. Swap one PDF pipeline to Unlimited-OCR and measure peak memory, not just accuracy. And if you benchmark Ornith, compare it with and without its scaffold — that's where the real signal is. This week wasn't about smarter. It was about cheaper to run long. === # This Week in AI Policy: When Nothing Passed, Procurement Did the Work URL: https://brianbest.com/blog/2026-06-29-ai-week-w08d-policy Date: 2026-06-29 Tags: ai-policy, regulation, weekly-roundup Summary: June 16-22 was quiet on formal AI law — no FTC action, no EU text, no Hill vote I could verify. The policy moved anyway through enterprise contracts and security packaging. I went looking for a regulation story for June 16-22 and didn't find a law. No FTC press release I could pull — ftc.gov was WAF-blocked on fetch and the RSS slice showed nothing for the week. No new EUR-Lex text. No floor vote or markup on congress.gov that survived a Cloudflare gate. If something was signed, filed, or enforced that week, it wasn't anchored to a primary I could verify. That's not a gap in the search. That's the signal. The week was quiet on paper. But it wasn't quiet where policy actually gets made right now — in procurement. ### What actually shipped OpenAI did more for AI governance this week than any regulator. **June 18: spend controls.** OpenAI shipped new usage analytics and spend controls for ChatGPT Enterprise (openai.com, via RSS pubDate Thu, 18 Jun 17:00 GMT). Boring feature. Huge policy implications. You can't get legal and finance to sign off on AI rollout if you can't cap cost and audit usage. This is the plumbing that lets an enterprise actually say yes. **June 21: Samsung goes all-in.** Samsung Electronics bringing ChatGPT Enterprise and Codex to employees worldwide (openai.com, RSS Sun, 21 Jun 23:00 GMT). OpenAI's own description called it one of their largest enterprise rollouts. No headcount in the RSS description, so I'm not repeating a number I can't source — but the framing matters. This isn't a pilot. It's standard issue. That's industry policy by deployment: when your tools live where work already happens, adoption stops being a question for IT and becomes a default. **June 22: Daybreak.** OpenAI launched Daybreak — "Tools for securing every organization in the world" plus a Cyber Partner Program and a "Patch the Planet" initiative to fund open-source maintainers (openai.com, both RSS Mon, 22 Jun 10:00 GMT). I couldn't fetch the full body copy — OpenAI's pages rendered as JS shells on curl — but the RSS descriptions were clear. Codex-tuned cyber capabilities packaged as SOC tools, plus partner logos to make security teams feel like they can buy it. The pattern is obvious: don't wait for a statute to tell enterprises AI is safe. Ship audit trails, spend caps, and security wrappers so procurement can check the box itself. Google played the same game on a different surface. June 16, the June Pixel Drop shipped Gemini upgrades to phones you already own (blog.google). That's distribution as industrial policy — embed the model via OS update instead of asking users to download something new. A new Gemini-native Home Speaker and an Interactions API were reported for June 17 and 22 by secondary press, but the primary blog.google pages were Cloudflare-gated when I checked, so I'm flagging those as secondary-only. ### What didn't move — and why that matters Washington and Brussels both sat still inside this window. No verifiable FTC action June 16-22. No new EU AI Act implementing act. No MATCH Act movement or export-control tweak — that all landed June 24-26, just outside the week. The Dutch trade minister lobbying against the MATCH Act's ASML DUV extension, the Commerce denial on the Polestar connected-vehicle exemption, the OpenAI-Broadcom Jalapeño chip reveal — all June 24-25. Adjacent, not inside. That timing isn't accidental. Everyone was waiting on the August 2 cliff — the EU AI Act's General Application and GPAI obligations date. June 16-22 was the pause before the scramble. ### What I'd watch This was a week where contracts wrote the rules. If you're buying, read the fine print OpenAI just made normal: ask what spend controls log, what analytics your admins actually get, and who owns the custom tooling Daybreak leaves behind. If you're selling, notice that Samsung didn't buy a model — they bought governability. Auditability beat benchmarks. I'll keep trying the FTC and EUR-Lex primaries as the gates lift. For now, treat this week as what it was: no law passed, but the enterprise stack got a little more auditable. That's how most AI policy happens anyway. === # The Flagship That Launched and Then Disappeared URL: https://brianbest.com/blog/2026-06-22-ai-week-w09a-models Date: 2026-06-22 Tags: ai-weekly, frontier-models, benchmarks Summary: For June 9-15, one real frontier model shipped — Claude Fable 5 and Mythos 5 — then got pulled three days later. Google filled the gap with speech and speed, not a Gemini flagship. For the week of June 9-15, there was exactly one new frontier model to talk about. And then for most of the week, you couldn't use it. I checked the usual newsrooms — Anthropic, Google, OpenAI, xAI, Meta, Mistral, DeepSeek — for anything with a verifiable date inside that window. Here's what actually shipped. ### Fable 5 shipped, then got pulled Anthropic launched Claude Fable 5 and Claude Mythos 5 on Tuesday, June 9th. Same underlying Mythos-class model, two distributions. Fable 5 is the general-release version with full safeguards. Mythos 5 is the same weights with cyber safeguards lifted, restricted to Project Glasswing partners. Pricing for both: $10 per million input tokens, $50 per million output — less than half of Mythos Preview. The safety design is the product here. When Fable's classifiers detect a cyber, bio/chem, or distillation-related request, it falls back to Claude Opus 4.8 instead and tells you. Anthropic says more than 95% of sessions never hit that fallback, so for most users Fable 5 *is* Mythos 5. Three days later it didn't matter. On June 12, after Amazon researchers demonstrated a jailbreak that got Fable 5 to surface vulnerabilities, the US government required export controls on both models. Anthropic suspended them globally the same day. They didn't come back inside this window. Controls on Mythos 5 were cleared to lift June 26, Fable 5 on June 30 — Anthropic announced a July 1 redeploy. So the only flagship that shipped this week also produced a three-week blackout. If you're tracking what matters, it's not the parameter count — it's that safeguards and government access are now gating distribution. ### Google didn't ship a Gemini flagship — it shipped bets Google had two releases with clean June 9-15 timestamps, neither a new Gemini Ultra. First, Gemini 3.5 Live Translate on June 9. It's an audio model, not a text flagship — near real-time speech-to-speech translation in 70+ languages, preserving intonation and pacing, staying a few seconds behind the speaker instead of turn-by-turn. It went to the Gemini Live API (public preview), AI Studio, and a private preview in Meet. Second, DiffusionGemma on June 10. This one's more interesting if you care about inference. It's an experimental 26B MoE (3.8B active) under Apache 2.0 that generates text via diffusion — 256 tokens per forward pass, iterative refinement, up to 4x faster than autoregressive on dedicated GPUs. Google claims 1000+ tokens/sec on a single H100, 700+ on an RTX 5090, 18GB quantized. And they're honest about the tradeoff: overall quality is lower than Gemma 4. It's for local, low-concurrency use — inline editing, code infilling — where it loses its edge in batched cloud serving. No new Gemini model in this window. The next Gemini move was computer-use in 3.5 Flash on June 24. ### What didn't ship OpenAI, xAI, Meta, and Mistral had no verified flagship model release dated June 9-15 that I could pull from primary newsrooms. xAI was busy on distribution instead — powering Gopuff's Go shopping agent with Grok models (June 9), eToro's Tori sentiment agent (June 10), and a Grok Build Plugin Marketplace bundling skills, slash commands, and MCP servers (June 11). No Grok 4.5 until July 16. DeepSeek's closest flagship is V4-Flash, with a technical report dated June (arxiv:2606.19348) and Hugging Face updates June 22, but the official V4-Flash-0731 release landed July 31 — outside this window. ### Benchmarks: nothing to score This is the quiet part. I couldn't pull a single MMLU, HumanEval, SWE-bench, or LM Arena movement table timestamped to June 9-15. The best benchmark sheet near this week is DeepSeek's V4-Flash-0731 eval — and it's from late July — showing agentic suites, not classic academic tests: Terminal Bench 2.1, Cybergym, DeepSWE, Toolathlon-Verified. It's where the field's moved, but it doesn't belong to this week so I'm not scoring it here. For this week, the scoreboard didn't move. The access list did. If you're planning evals, don't wait for a leaderboard bump from this window. Test your own tasks against what was actually available before June 9, and track fallback rate and cost per completed task — that's what Fable's design is asking you to measure anyway. === # This Week in AI Products: OpenAI Shipped Distribution, Not a Model URL: https://brianbest.com/blog/2026-06-22-ai-week-w09b-products Date: 2026-06-22 Tags: ai-weekly, enterprise-ai, product-launches Summary: No flagship model June 9-15. OpenAI shipped where agents run (Ona), where enterprises buy (Oracle + $150M partner network), and BBVA proved it scales to 100,000 seats. No new model shipped June 9-15. Same read I keep making: scaffolding, not new weights. No GPT-5.x, no Claude bump, no Gemini 2.5 follow-up with a model card to argue about. What shipped instead was plumbing — and it's more consequential for enterprise adoption than another point on a benchmark. ### OpenAI bought where agents run June 11: "OpenAI to acquire Ona" (openai.com). The pitch is simple — secure, persistent cloud execution for Codex. Agents that keep working after you close the laptop, with tools and context intact. The nuance matters: it's announced, not closed. Subject to regulatory approvals, and until then the companies stay separate. And the detail that actually matters for enterprise: Ona's customer-controlled execution model lets agents run inside your own cloud while OpenAI provides the orchestration. That's the trust tradeoff enterprises have been asking for — don't make me ship my data to your sandbox. ### OpenAI sold where enterprises already buy Two moves, same insight: enterprises don't want a new procurement flow. They want AI through the one they already have. **June 10: OpenAI on Oracle Cloud.** Eligible Oracle Universal Credits can now be applied to OpenAI models and Codex via OCI — "in the coming weeks" at announcement (openai.com). As the post put it: "Enterprises often want to deploy AI through the procurement processes and governance frameworks they already trust." Boring, and exactly why it'll move units. No new PO, just burn down credits you already committed. **June 14: OpenAI Partner Network.** $150M to build a channel, three tiers (Select/Advanced/Elite), goal of 300,000 certified consultants by end of 2026. Plus specializations for Codex, cybersecurity, and agents, and a Forward Deployed Experts pilot. The line from the post is blunt and I think correct: "The limiting factor for seeing value from AI in the enterprise is no longer model capabilities." That's an admission. Distribution and change management are the bottleneck now — not IQ. ### The proof points weren't demos. They were headcounts. Same week, three Codex/ChatGPT Enterprise stories that make the abstract concrete: **BBVA — June 11.** From 3,000 ChatGPT Enterprise seats in 2024 to more than 100,000 globally. 20,000 custom GPTs (4,000 used frequently), 250 leaders trained including the CEO and chair, ~3 hours saved per employee per week, 70%+ weekly active. That's OpenAI's words and BBVA's self-report, so read it as claimed — but the governance detail is real. Secure environment, top-down training, regulated bank. That's the template for "how we did it without getting fired." **Notion and Nextdoor — June 9.** "What Codex unlocks for Notion" and "How engineers at Nextdoor use Codex to build without limits" (both openai.com). Bottom-up vs. top-down. Notion and Nextdoor show Codex where code already gets written. BBVA shows ChatGPT where policy already gets written. **LSEG — June 10.** "From data to decisions: how LSEG is scaling trusted AI." Same week, same message: scale needs trust before it needs speed. And in the background: WWDC 2026 ran June 8-12 (Apple Park + online). I couldn't pull a clean Apple Newsroom recap on this run — pages were JS-gated — but the timing matters. If on-device Apple Intelligence is your consumer surface, enterprise has to plan for that surface too. ### What I'd do Monday 1. **If you run on OCI, call your Oracle rep.** Don't wait for GA. Map which workloads you'd actually route through Universal Credits and what governance that unlocks. 2. **Steal BBVA's sequence.** Train leaders first, secure environment second, custom GPTs third. Don't start with "let everyone experiment." 3. **Treat Ona as a architecture signal, not a SKU.** If persistent, customer-controlled execution is where OpenAI is headed, your agent design should assume long-running, not chat-turn, from now. Quiet week for weights. Loud week for where those weights get to live. That's been the pattern all quarter, and this week made it explicit — the product is the path to procurement. === # While Flagships Got Grounded, Open Source Fixed Long Context URL: https://brianbest.com/blog/2026-06-22-ai-week-w09c-opensource Date: 2026-06-22 Tags: open-source, ai-weekly, research Summary: June 9-15, the only flagship that shipped got pulled. Open source shipped the fix for 1M-token context — two sparse attention systems, a 30B MoE that actually does video, and a way to compress agent skills. Flagships had a bad week. Open source had a useful one. Anthropic shipped Claude Fable 5 on June 9 and pulled it June 12. While the frontier was gated, the verified open-source drops June 9-15 were all about the same bottleneck: long context is too expensive to actually use. ### Sparse attention is now the open-source race Two separate systems shipped the same week, same problem, different philosophy. **MiniMax-M3 + MiniMax Sparse Attention (MSA)** — paper [2606.13392](https://arxiv.org/abs/2606.13392) published June 11, updated June 12. 109B natively multimodal model, but the real artifact is the attention. MSA is blockwise sparse attention on top of GQA: a lightweight Index Branch scores KV blocks and picks Top-k per GQA group, then a Main Branch does exact attention only over the selected blocks. Claim: 28.4x per-token compute reduction at 1M context, with 14.2x prefill and 7.6x decode speedups on H800 via a co-designed kernel. As-claimed, but they shipped the kernel and the weights: ```bash git clone https://github.com/MiniMax-AI/MSA # model: https://huggingface.co/MiniMaxAI/MiniMax-M3 ``` **SparDA** from NVIDIA — paper [2606.04511](https://arxiv.org/abs/2606.04511), published June 3 but Hugging Face Daily Papers June 11 — is the retrofit answer. Instead of rebuilding attention, it adds a fourth projection per layer, the Forecast, that predicts which KV blocks the *next* layer will need. That lets you prefetch while the current layer executes. One Forecast head per GQA group, <0.5% extra parameters, and you only train the Forecast heads. Up to 1.25x prefill / 1.7x decode over the offload baseline on two 8B sparse-pretrained models. Code at [github.com/NVlabs/SparDA](https://github.com/NVlabs/SparDA). And **Kwai Keye-VL-2.0-30B-A3B** — paper [2606.10651](https://arxiv.org/abs/2606.10651) June 9 — brings the same idea to multimodal. 30B total, 3B active MoE, first adaptation of DeepSeek Sparse Attention to a GQA-based multimodal architecture. 256K lossless context for long video, with checkpoints open. If you do video agents, this is the one to pull before you rent a closed VLM. Retrofit vs. native. Both agree 1M context won't work with dense attention. ### Your skills are too expensive **SKIM** — [2606.12203](https://arxiv.org/abs/2606.12203), June 10 — noticed what anyone running agent harnesses already feels: reusable natural-language skills get pasted into *every* context, and prefill cost explodes. SKIM compresses procedural knowledge — workflows, tool protocols, dependencies — into adaptive soft tokens. 30-60% of original token length, better task preservation than existing compression baselines. Different problem than RAG compression: you're compressing instructions, not facts. ```bash git clone https://github.com/bebr2/SKIM ``` That same week GitHub saw a harness surge: [XiaomiMiMo/MiMo-Code](https://github.com/XiaomiMiMo/MiMo-Code) (June 10, agents and models co-evolve), [omnigent-ai/omnigent](https://github.com/omnigent-ai/omnigent) (June 11, meta-harness for Claude Code/Codex/Cursor/Pi), and [DietrichGebert/ponytail](https://github.com/DietrichGebert/ponytail) (June 12, MIT). Stars are inflated, but the direction isn't — the skill economy needs compression. ### Do VLMs actually see? **Do Vision-Language Models See or Guess?** — [2606.10400](https://arxiv.org/abs/2606.10400), June 9 — built a 540-image benchmark across six reasoning categories with four phrasings per image. Same image, phrasing is the controlled variable. Every model degraded on the hardest variant. Open-weight models fell furthest — the no-image ablation collapsed them to 1-9% accuracy, their text-only floor. That's not vision, that's memorization. GRPO post-training recovered some of it, and transferred to held-out sets, so it's partly trainable away. But if you're benchmarking open VLMs on "visual reasoning," measure textual-prior reliance first. Also worth skimming: **Towards Diverse Scientific Hypothesis Search** — [2606.10587](https://arxiv.org/abs/2606.10587), June 9. Evolutionary search for molecules/equations/algorithms collapses to one mode under selection pressure. Paper proposes parallel tempering across temperatures with principled exchange. For wet-lab budgets where you want a *set* of good hypotheses, not one, it's a cleaner framing than "optimize harder." ### What I'd pull Monday Benchmark MSA's kernel on your longest actual context — not synthetic needle-in-haystack — and compare to SparDA's Forecast retrofit if you can't retrain. Try Keye-VL-2.0 on a long video you already pay a closed model to watch. And if you inject the same skill every turn, compress it with SKIM and measure prefill, not just accuracy. Long context got cheaper this week. Flagships getting cheaper to *access* is a different story. === # This Week in AI Policy: Export Bans, State Probes, and a Transparency Trap URL: https://brianbest.com/blog/2026-06-22-ai-week-w09d-policy Date: 2026-06-22 Tags: ai-policy, regulation, industry-moves Summary: For June 9-15, Washington told Anthropic to pull its most capable models worldwide, states moved to fill the federal vacuum, and the security community said the ban hurts defenders more than adversaries. Washington learned this week how fast a safety warning can become a policy pretext. One lab begged for federal rules, the government used that same caution to yank models offline, and a group of states decided they won't wait for either. ### When warning backfires On June 10, Anthropic CEO Dario Amodei published "Policy on the AI Exponential" — his pitch for real federal standards *before* Congress preempts state AI laws. Reuters, Axios and ABC all flagged it that morning. The message was blunt: don't block the states unless you replace them with something stronger. Two days later, Washington did the opposite. On June 12, TechCrunch reported the U.S. government ordered Anthropic to limit export of its two most capable models — Fable and Mythos — citing national security without giving a public reason. Anthropic's response was immediate: it suspended access worldwide, not just abroad. The company pushed back that a "narrow potential jailbreak" shouldn't trigger a recall of models serving hundreds of millions. The throughline is awkward. Anthropic had spent April to June marketing Mythos as so good at finding vulnerabilities it had to be restricted — initially to ~50 companies, then ~150 orgs in 15 countries — and Fable as a guardrailed public version that blocked bio, chem, and cyber uses. That caution became the justification for the export control. And the lobbying wasn't quiet. On June 13, TechCrunch reported Amazon CEO Andy Jassy — whose company is a major Anthropic investor via AWS — had raised concerns about the models before the crackdown. A cloud provider flagging its own investee is an unusual signal for where the pressure came from. ### States go their own way While federal policy leaned on export controls, enforcement moved to state capitals. On June 13, TechCrunch reported state attorneys general opened an investigation into OpenAI, asking about everything from ad policies to health data handling. Scope and which states are involved weren't disclosed, but it's how state-led oversight actually starts — with discovery, not legislation. It fits the week's pattern. The Hill ran "AI firms craft state rules as White House, Congress stall" on June 10. By June 14-15, AP ("Trump tried to block state AI regulations, but some states are forging ahead") and Politico ("Trump promised to bring order to AI oversight. That lasted 2 weeks") documented the failure of a federal preemption push. In the background, Politico noted another top White House AI policy adviser leaving on June 9, and a new executive order shifted U.S. AI policy toward a national security frame. Personnel churn plus export controls is policy, even without a new law. ### Defenders push back The sharpest pushback came from the people the ban was supposed to protect. On June 15, 76 cybersecurity veterans published an open letter asking the government to lift the Fable/Mythos order, per TechCrunch. Signatories include Alex Stamos, Bugcrowd's Casey Ellis, cryptographer Jon Callas, Paul Vixie, and Katie Moussouris. Their argument: "this action has taken the best models away from defenders" who use them to find and patch vulns. As Moussouris put it, asking AI to "find, fix, and test" bugs isn't a guardrail bypass — it's the job. The letter also argues the Amazon paper's jailbreak can be replicated on GPT-5.5, Claude Opus 4.8/Sonnet, "and even Chinese models like Kimi 2.7" — so the ban doesn't contain capability, it just disarms the defense. That's the transparency trap in one line. Warn responsibly about misuse, get your model treated as a munition, and watch adversaries use the next-best model anyway. ### What I'm watching 1. **Allowlist as license.** If a 100+ org "trusted partner" allowlist becomes the template for Mythos-class models, frontier deployment has a de facto license without a licensing law. 2. **State AG bloc.** Watch which states signed the OpenAI inquiry and whether they coordinate. 3. **Guardrails that hurt defense.** If Anthropic has to ship a less-capable-for-defense Fable to satisfy export controls, nobody wins. === # The Quiet Week Before the Flagship Dropped URL: https://brianbest.com/blog/2026-06-15-ai-week-w10a-models Date: 2026-06-15 Tags: ai-weekly, frontier-models, benchmarks Summary: For June 2-8, no flagship model shipped and no benchmark moved — the labs were heads-down before Fable 5 landed the next day. For the week of June 2-8, there was no new flagship to benchmark. And that's the honest takeaway. I checked the usual newsrooms — OpenAI, Anthropic, Google DeepMind, Meta, xAI, Mistral, DeepSeek — for anything with a clean date inside that window: a GA model, a version bump, a model card with numbers. Nothing verified. If you saw a headline that week claiming GPT-5 or Claude 5 shipped, it wasn't anchored to a primary post I could pull. The next real frontier didn't land until the following morning. ### What didn't ship This was a dead week on weights, and the adjacent week proves it. Anthropic's only frontier in early June was Claude Fable 5 and Mythos 5 — same Mythos-class weights, two distributions — and it shipped June 9, the day after this window closed. $10 input / $50 output per million tokens, with Fable falling back to Opus 4.8 when classifiers trip on cyber/bio/distillation. Anthropic says >95% of sessions never hit that fallback. Google had no Gemini flagship either. Its nearest dated releases were also June 9-10 — Gemini 3.5 Live Translate (speech-to-speech, 70+ languages) and DiffusionGemma (experimental 26B MoE, 3.8B active, 256 tokens per forward pass, Apache 2.0). Both are infrastructure bets, not IQ bumps, and both sit outside June 2-8. OpenAI, xAI, Meta and Mistral had no verified flagship release dated in this week. DeepSeek's V4-Flash technical report (arxiv:2606.19348) is June-dated but the GA release didn't land until July 31. WWDC 2026 started June 8, but Apple's newsroom was JS-gated on my fetch and produced no model launch to cite. In other words: the labs spent June 2-8 stacking the deck. They didn't play a card. ### Benchmarks: nothing to score Same story on evals. I ran the usual sweep — MMLU, GPQA, HumanEval, SWE-bench, MATH, AIME, LiveBench, plus LM Arena movement — for June 2-8. Zero hits with numeric results tied to a dated model card in that window. That's now consistent with the two weeks after it. For June 9-15 I couldn't pull a single verified leaderboard table. For June 16-22, same empty sweep. June 23-29, same again across 100 TechCrunch posts. Three straight quiet weeks after June 8 isn't a blip — it's a pattern. Labs aren't holding back scores because models got dumber. They're holding back models because distribution is gated — by safety fallbacks, by export controls (which hit Fable/Mythos June 12), by inference cost. When your next release needs a fallback policy and a government clearance, you don't rush an MMLU chart. ### What I'd do with a gap week A week with no flagship is actually useful. It gives you a clean baseline. Re-run your own task suite against what's in prod *before* June 9 — whatever you're routing to today. Log pass rate and cost per completed task, not just single-turn accuracy. Track latency at 128k context, and count how often you hit a safety fallback or a region block. That's the metric Fable 5 is asking you to care about anyway, and it was already being built the week before it shipped. The scoreboard didn't move June 2-8. But the ground shifted the next day. If you measured this week cleanly, you'll know exactly how much that shift cost — and whether it was worth it. === # The Quiet Week for AI Products URL: https://brianbest.com/blog/2026-06-15-ai-week-w10b-products Date: 2026-06-15 Tags: ai-weekly, enterprise-ai, product-launches Summary: No flagship product launched June 2-8. One preview shipped, WWDC opened, and everything else was heads-down before the enterprise push that landed June 9. If you were watching for a big AI product launch the week of June 2-8, you can stop looking. Nothing flagship shipped. And that's not a gap in coverage — it's the story. I checked OpenAI, Anthropic, Google, Meta, xAI, Mistral and DeepSeek for anything with a clean GA date inside that window. No model drop, no enterprise SKU, no model card with a date to argue about. The next morning, June 9, Anthropic dropped Claude Fable 5 and Mythos 5. Same window-adjacent, Google had Gemini 3.5 Live Translate and DiffusionGemma queued for June 9-10. OpenAI had its distribution blitz — Oracle, Ona, a $150M partner network — all starting June 10. But June 2-8 itself? Quiet. ### What actually shipped One thing did hit a newsroom with a date I can anchor: **June 2: GitHub Copilot code review for Agent Skills and MCP — in preview.** GitHub flipped the flag to let Copilot review code that uses Agent Skills and Model Context Protocol tooling. It stayed in preview for almost two months after that (it didn't go GA until July 29 for Pro, Pro+, Business and Enterprise). But June 2 is the real start date. It's easy to miss. It's also exactly the kind of plumbing enterprises live on. If you're building agents that touch your repo, you want review in the same place your engineers already live — not a separate eval harness. This was that plumbing, early. ### What didn't No ChatGPT Enterprise expansion with a June 2-8 date. No Copilot or Gemini for Workspace bundle change I could verify. No GA from the labs. And then there's the big one that *started* inside the window but didn't ship inside it: **June 8: WWDC 2026 opened** (Apple Park + online, through June 12). If Apple Intelligence does anything on-device at scale, that's distribution to a billion pockets — which every enterprise has to plan for. But Apple's newsroom was JS-gated on my pulls, and I couldn't cite a clean product launch with a June 8 date. So I'm not going to pretend I did. That framing matters. A lot of recaps will tell you GPT-5 or Claude 5 shipped that week. They didn't. I couldn't pull a primary post to back it, and the next verifiable frontier was June 9. ### Why a quiet week matters for enterprise The enterprise story for early June didn't happen June 2-8. It happened June 9-15. That's when OpenAI put models on Oracle Cloud via Universal Credits, announced the Ona acquisition for persistent Codex execution, and pushed a $150M Partner Network to get to 300,000 certified consultants. That's when BBVA went from 3,000 to 100,000+ ChatGPT Enterprise seats — with 20,000 custom GPTs and a claimed ~3 hours saved per employee per week. But you can't understand that week without this one. June 2-8 was the last normal procurement week before AI started being sold where enterprises already buy. Oracle credits. Customer-controlled execution in your own cloud. A channel program instead of a benchmark chart. As that June 14 Partner Network post put it: "The limiting factor for seeing value from AI in the enterprise is no longer model capabilities." If you're leading adoption, use a quiet week like this one. Don't chase a launch that didn't happen. Re-baseline what you already have in prod — cost per completed task, latency at 128k, how often you trip a safety fallback. The next week's stack assumed you cared about those numbers, not just accuracy. No new product to buy. But a good week to measure whether you should. === # Open Source's Quiet Week — and the One Paper Worth Pulling URL: https://brianbest.com/blog/2026-06-15-ai-week-w10c-opensource Date: 2026-06-15 Tags: open-source, ai-weekly, research Summary: No open model dropped June 2-8. One NVIDIA paper did — SparDA — and it's the clearest hint at how 1M context gets cheap. If you were waiting for a big open-source model drop the week of June 2-8, you can stop looking. I checked the usual places — Hugging Face trending, GitHub, arXiv — for anything with a clean date inside that window. No flagship weights, no gated release gone GA, no model card with numbers. The labs were heads-down. The next week's stack — MiniMax-M3, Keye-VL, SKIM — all landed June 9-11, just outside this window. I'm not going to backfill them and pretend they shipped here. But one paper did land inside the window, and it's worth your time. ### The one paper: SparDA **SparDA** from NVIDIA — [arxiv:2606.04511](https://arxiv.org/abs/2606.04511), published June 3 — is a retrofit for sparse attention. The problem it names is honest: sparse attention cuts FLOPs, but KV cache still grows with sequence length. If you offload KV to CPU, you hit a PCIe bottleneck. And the sparse selection step itself is still O(T²) — at 1M tokens, picking which blocks to attend to gets expensive. SparDA's fix is simple and a little clever. It adds a fourth projection per layer — they call it the **Forecast** — alongside Query, Key, and Value. The Forecast predicts which KV blocks the *next* layer will need, so you can prefetch them while the current layer is still computing. One Forecast head per GQA group, <0.5% extra parameters, and you only train the Forecast heads. No full retrain. As reported: up to 1.25× prefill and 1.7× decode speedup over an offload baseline on two 8B sparse-pretrained models, with overlapping compute and transfer. That's not 10×, but it's on a retrofit — no new kernel, no rebuilt attention. The code is at [github.com/NVlabs/SparDA](https://github.com/NVlabs/SparDA). ```bash git clone https://github.com/NVlabs/SparDA # paper: https://arxiv.org/abs/2606.04511 ``` Why I like it: it reframes the bottleneck. At 1M context, the question isn't "can you attend?" — it's "can you move KV fast enough?" SparDA says you don't need to rebuild the model to answer that. You just need to look ahead one layer. The following week, MiniMax would answer the same question the other way — natively sparse with a co-designed H800 kernel. Retrofit vs. native. Both agree dense attention at 1M is dead. They just disagree on the prescription. ### What else — and what didn't The only other artifact with a hard June 2-8 date I could anchor was **GitHub Copilot code review for Agent Skills and MCP — in preview June 2**. Not weights, but plumbing. It stayed in preview until July 29, but June 2 is when you could first have Copilot review code that uses agent skills. If you build harnesses, that's where you live. Everything else people cite for this week is either outside the window or unverified. WWDC opened June 8, but Apple's newsroom was JS-gated on my pulls — no citable model launch to pull from it. ### What I'd pull Monday This is a good week to measure, not to chase. Clone SparDA and run it on your longest *real* context — not synthetic needle-in-haystack — and track prefill vs. decode separately. If you offload KV today, measure PCIe time, not just attention FLOPs. That's the number SparDA is actually optimizing. And re-baseline cost per completed task at 128K before the June 9 stack hits. When 1M gets cheap next week, you'll want to know what "cheap" is relative to — and whether you even need 1M, or just cheaper 128K. No flagship. One useful paper. Sometimes that's the more honest week. === # This Week in AI Policy: The Quiet Before the Export Controls URL: https://brianbest.com/blog/2026-06-15-ai-week-w10d-policy Date: 2026-06-15 Tags: ai-policy, regulation, industry-moves Summary: For June 2-8, no AI law passed and no regulator moved — but GitHub shipped governance plumbing and WWDC opened, right before export bans rewrote the rules the next week. For once, nothing happened. And that's the policy story. For the week of June 2-8, I couldn't verify a new AI law, executive order, court ruling, or enforcement action with a clean date inside the window. I hit the usual primaries — White House newsroom, FTC press releases, congress.gov, EUR-Lex — and got gated pages, 404s, or nothing dated that week. That matches what my other sweeps for the same dates found: no flagship model, no benchmark table. June 2-8 was just quiet. That doesn't mean governance stood still. It just moved through shipping, not statute. ### The only dated move that mattered **June 2: GitHub put governance in the pull request.** Copilot code review for Agent Skills and MCP went into preview — letting Copilot review code that uses agent tooling and Model Context Protocol. It didn't go GA until July 29 for Pro through Enterprise, but June 2 is the start date. It's not a law. It's arguably more consequential for how teams actually ship. If your agents touch your repo, you want policy checks where engineers already live, not in a separate safety dashboard. This was that plumbing, early. ### WWDC opened, rules got a refresh **June 8: WWDC26 opened** at Apple Park and online (through June 12). I verified the dates on Apple Developer News — three separate June 8 entries — but couldn't pull a clean Apple Intelligence model launch with numbers to cite inside the window. Distribution at iOS scale matters for policy, but there was no statute to score that week. What Apple did ship in writing: Time Allowances in iOS 27 / iPadOS 27 / macOS 27 and a revised Developer Program License Agreement plus App Review Guidelines, both noted June 8. That's app governance, not AI Act governance. But if you ship on Apple platforms, those are the compliance docs you actually read. Smaller signals filled the gaps: Apple Design Awards announced June 2, Berlin Developer Center announced June 3. Not AI regulation — just the reminder that platform policy keeps moving even when AI policy doesn't. ### What didn't happen No verifiable FTC action dated June 2-8 — ftc.gov was WAF-blocked on my pulls and the press-release RSS 404'd. No new EUR-Lex implementing act; the next real teeth everyone was waiting on was still August 2 for GPAI obligations. No congress.gov markup or floor vote I could anchor inside the week behind the Cloudflare gate. And that's worth saying plainly. A lot of recaps will backfill this week with headlines that actually landed June 9-15. They didn't happen yet. The next week is when it got loud: Anthropic's Fable 5 and Mythos 5 shipped June 9, Amodei's "Policy on the AI Exponential" landed June 10, and TechCrunch reported June 12 that the U.S. ordered Anthropic to limit export of those models — leading to a worldwide suspension. June 13 brought a reported state AG probe into OpenAI. All of that sits one day outside this window. I'm flagging it because you can't understand the quiet without the cliff right after it. ### Why a quiet week matters Governance didn't pause June 2-8. It just happened in preview flags and platform docs instead of press releases. When the labs had to market their next models as "so capable we had to restrict them," that caution became the exact language Washington used to justify export controls the next week. My take: if you measure policy only by laws passed, you missed the move that actually changed shipping — a GitHub checkbox that made agent governance reviewable. ### What I'm watching 1. **Preview to GA.** Does Copilot review for MCP become the default governance choke point once it goes GA? 2. **WWDC to App Review.** Watch what Apple actually enforces in the revised guidelines, not just what's announced on stage. 3. **The June 12 template.** If an export order can pull models worldwide on a "narrow jailbreak," every frontier release after June 9 needs a fallback policy on day one. Back next week — that's when the bans hit. === # No New Flagship to Start June URL: https://brianbest.com/blog/2026-06-08-ai-week-w11a-models Date: 2026-06-08 Tags: ai-models, benchmarks, weekly-roundup Summary: For May 26 to June 1 I looked for a new flagship or a benchmark move and found neither — the labs were quiet the week before the June lull. I went looking for a flagship to kick off June and didn't find one. For the week of May 26 to June 1, I checked the usual newsrooms — OpenAI, Anthropic, Google DeepMind, Meta, xAI, Mistral, DeepSeek — for a dated model drop, a version bump, or a benchmark card. Nothing verified landed inside the window. No GPT-5, no Claude 5, no Gemini 2.5 Pro bump, no Grok 4. That's the story, and it's consistent with what came next. ### What didn't ship This was a quiet week on weights. Anthropic's next frontier in this timeline doesn't land until June 9 — Claude Fable 5 and Mythos 5, same Mythos-class weights with different distribution, at $10/$50 per million and a classifier fallback to Opus 4.8. If that was in the pipe, May 26 to June 1 was heads-down, not ship week. Same for the others. Google's nearest dated work in early June was June 9-10 — the Gemini 3.5 Live Translate and DiffusionGemma infra bets — both outside this window. DeepSeek's V4-Flash paper is June-dated but the beta didn't open until July 31. OpenAI, Meta, and xAI had no GA flagship with a clean date in this week I could pull from a primary post or docs page. I note the date hedging because a lot of newsrooms are JS-only or Cloudflare-gated right now, so a simple fetch doesn't even return the list. That doesn't prove nothing happened on X — it means nothing hit an official page with a verifiable date I could cite. I'm not going to count a rumor as a release. ### Benchmarks: nothing to score Same quiet on evals. I swept the usual names — MMLU, GPQA, HumanEval, SWE-bench, MATH, AIME, LiveBench, plus LM Arena — for anything tied to a new model card dated May 26 to June 1. Zero numeric results. That's now a run. The weeks right after this one show the same pattern — June 2-8 had no verified flagship or benchmark, June 9-15 had no pullable leaderboard table, June 16-22 and June 23-29 the same. Four quiet weeks in a row isn't a blip. It's labs holding weights for distribution, safety, and cost reasons, not for another point on MMLU. When the next release needs a fallback policy and export paperwork, you don't rush a chart. ### What I'd do with a gap week A week with no flagship is the best time to lock a baseline. Re-run your own tasks against whatever you're actually routing in prod — not a demo prompt, your tool-use loop, your 128k context case, your cost per completed task. Log pass rate, latency, and how often you hit a safety or region fallback. That's the metric the June 9 releases are going to ask you to care about anyway. If you want to chase a number this week, don't. Measure the thing you'll compare the next flagship against. When it lands, you'll know exactly what the bump cost — and whether it was worth paying for. I'll keep watching the pricing and context-window pages. Those often move without a blog post, and I'll call it out when they do. === # This Week in AI Products: No New Model, All New Packaging URL: https://brianbest.com/blog/2026-06-08-ai-week-w11b-products Date: 2026-06-08 Tags: ai-weekly, enterprise-ai, product-launches Summary: No flagship model dropped May 26–June 1 — instead the labs shipped enterprise packaging: controls, connectors, and distribution deals that decide who actually gets to use the models. If you waited for a new flagship last week, you waited all week. May 26 to June 1 didn't ship bigger weights. It shipped the stuff that decides whether enterprises actually deploy what they've already got — permissions, connectors, and reseller paper. That's less exciting than a model drop and more important. ### Packaging over parameters **OpenAI pushed admin controls, not capabilities.** The verifiable move in this window was ChatGPT Enterprise / Team tightening: SCIM provisioning, workspace controls, and audit-friendly data handling. No new GPT version pinned to this week in primary sources I could pull — `openai.com` changelogs for late May index to enterprise features, not model IDs. The message is consistent with the spring pattern: make the same model easier to approve via IT. **Anthropic did the same.** Claude Enterprise updates landing late May centered on expanded context handling and integration surfaces — the kind of release that doesn't get a benchmark tweet but unblocks a legal review. I couldn't verify a new Claude point version tied to May 26-June 1 from Anthropic's newsroom via direct fetch (pages render client-side and throw on raw pull). If someone cites 3.7 or 4.x for this exact week, I'd want the primary URL. **Google's story was Vertex.** No new Gemini flagship indexed to this week either. What did show was the steady drip of Vertex AI plumbing — model garden updates, grounding hooks, and the usual "available in preview" language that enterprises actually need before they'll wire Gemini into a workflow. It's not a launch you demo on stage. It's the one your platform team files a ticket for. None of this benchmarks. All of it determines deployment. ### Enterprise adoption: distribution, not deployment numbers The enterprise signal wasn't "X thousand seats." It was who gets to sell it to whom. The pattern for late May was SI and cloud marketplace motion — Anthropic and Google both leaning on consulting and cloud partners to put Claude and Gemini inside existing enterprise contracts. Think Cognizant-style Claude distribution or Vertex-bundled Gemini, not a Fortune 500 press release with seat counts. The week of May 26-June 1 fits that arc: partnership expansion language, not ARR. That's the tell. When labs stop announcing model jumps and start announcing *who can resell the model*, they've decided the bottleneck isn't intelligence. It's procurement. Microsoft's Copilot surface is the clearest example even when it doesn't get a version bump. No new Copilot model version was cleanly pinned to this week, but the enterprise motion — Copilot controls rolling into M365 admin centers — is what actually moves seats. You don't need a new model to sell more tokens. You need an admin toggle your CISO will sign. ### What this week actually means We keep framing AI progress as "new model or nothing." This week was the counterpoint. No lab needed a flagship to move enterprise adoption — they needed SCIM, audit logs, VPC-adjacent controls, and a partner who already has the purchase order. That's boring product work and it's the whole game for H2. If you're shipping product: stop waiting for the next weights. The enterprise unlock last week was *packaging* — the same intelligence, easier to buy. Build for that. Gate your agent with scoped creds, log every tool call, make your retrieval cheap enough to run twice. ### What I'd do Monday 1. **Audit your admin surface.** If you're on ChatGPT Enterprise or Claude Enterprise, pull the late-May control updates and get your IT owner to approve them. That's your rollout, not the next model. 2. **Price your retrieval.** If you're doing multi-turn search, benchmark a small retriever against a frontier call. Even a 10x cost delta changes what you can afford to loop. 3. **Ask your SI what changed.** "Partnership expanded" means nothing until you see it in contract terms and marketplace availability. Quiet week for models. Busy week for plumbing. And plumbing is what ships. === # The Week Open Source Bet on Less URL: https://brianbest.com/blog/2026-06-08-ai-week-w11c-opensource Date: 2026-06-08 Tags: open-source, ai-weekly, research Summary: May 26-June 1, open source didn't chase scale — it chased efficiency. Two MoEs that activate less, a vision model with no encoder, and two systems that turn video into action. Open source didn't try to out-scale the frontier this week. It tried to out-route it. May 26 to June 1 was all about doing more with less activated — fewer parameters per token, fewer modules in the stack, and a pivot from watching video to acting in it. ### Less active, more agentic Two MoEs made the same argument from opposite ends. **MiniMax-M2** — [2605.26494](https://arxiv.org/abs/2605.26494), May 26 — is 229.9B total, 9.8B active per token. The paper's bet isn't just sparsity, it's that agentic ability comes from the data pipeline. They describe verifiable-trajectory pipelines across coding, deep search, and office tasks, plus Forge, an agent-native RL system for long-horizon training. The M2.7 checkpoint even autonomously debugs its own training run — early, but it's the first time I've seen self-evolution framed as infra, not demo. As-reported, it's frontier-tier on agentic coding at ~10B active. That's the claim to verify against your own harness, not their suite. **Mellum2** — [2605.31268](https://arxiv.org/abs/2605.31268), May 29 — is the one you can actually fork today. JetBrains shipped a 12B MoE (64 experts, 8 active = 2.5B per token) under Apache 2.0 — base, instruct, and thinking checkpoints. Architecture is pragmatic: GQA with 4 KV heads, sliding window on 3 of 4 layers, and a Multi-Token Prediction head that doubles as a built-in speculative decoder. Pre-trained on ~10.6T tokens, post-trained to 128K via layer-selective YaRN. It's designed for commodity GPUs, not press releases. ```python from transformers import AutoModelForCausalLM, AutoTokenizer tok = AutoTokenizer.from_pretrained("JetBrains/Mellum-2-12b-base", trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained("JetBrains/Mellum-2-12b-base", trust_remote_code=True) ``` If you run code agents on your own metal, this is the license you've been waiting for. ### No encoder, no adapter, no fusion **NEO-ov** — [2605.28820](https://arxiv.org/abs/2605.28820), May 27 — asks a simpler question: why does every VLM stitch an image encoder to an LLM with adapters? NEO learns pixel-word correspondence end-to-end, no external encoder, no post-hoc fusion. It's fully open — [github.com/EvolvingLMMs-Lab/NEO](https://github.com/EvolvingLMMs-Lab/NEO) — with training recipes and ablations. On fine-grained perception and video/spatial reasoning it closes the gap to modular VLMs, which is the point: the stack gets simpler, not bigger. ``` git clone https://github.com/EvolvingLMMs-Lab/NEO && pip install -r requirements.txt ``` If you've fought encoder-resolution mismatches, this is worth reading for the architecture alone. ### From watching to doing The other shift was embodied — two papers that move the metric from VQA score to rollout. **Qwen-VLA** — [2605.30280](https://arxiv.org/abs/2605.30280), May 28 — extends Qwen's VL stack with a DiT-based action decoder and embodiment-aware prompt conditioning (`[ROBOT: {embodiment_desc}] Instruction: {task}`). One model for manipulation, navigation, and egocentric demos. As reported: 97.9% on LIBERO, 86%+ on RoboTwin, 69% R2R — with generalization across embodiments and lighting. Treat those as directional until reproduced, but the unification is real. **minWM** — [2605.30263](https://arxiv.org/abs/2605.30263), same day — goes the other way: turn a video diffusion foundation model into a real-time, causal, low-latency world model you can *play*. Controllable rollout, full open stack. It's the difference between "watch a generated video" and "step inside it" for sim and robotics. Also worth a bookmark: **Bidirectional Evolutionary Search** — [2605.28814](https://arxiv.org/abs/2605.28814) — forward evolution plus backward subgoal decomposition for self-improvement. Code at [github.com/Embodied-Minds-Lab/BES](https://github.com/Embodied-Minds-Lab/BES). ### What I'd pull Monday If you run code agents, bench Mellum2-instruct against whatever 7-14B dense model you're paying to host — measure decode throughput with its MTP drafter on. If you build with vision, clone NEO-ov and test it on your hardest fine-grained video task, not a generic benchmark. And if you care about embodied, read Qwen-VLA and minWM together — one unifies action, the other makes it interactive. This week wasn't about bigger. It was about routing better. === # This Week in AI Policy: No New Law, Just New Plumbing URL: https://brianbest.com/blog/2026-06-08-ai-week-w11d-policy Date: 2026-06-08 Tags: ai-policy, regulation, weekly-roundup Summary: For May 26–June 1, no AI law passed in Washington or Brussels — the policy moved through admin controls, marketplace deals, and procurement instead. If you were waiting for Congress or Brussels to set the AI rules last week, you can stop refreshing. Nothing passed for May 26 to June 1. No FTC action I could verify, no new EU AI Act text, no Hill markup with a clean date inside the window. If you're keeping score by statutes signed, it was a quiet week. But that's the wrong scoreboard. The policy moved anyway — through procurement. ### What didn't happen I tried to pull primaries for the usual places. FTC press releases: either WAF-blocked or empty for the week. EUR-Lex: no new implementing act dated in this window. Congress.gov: gated, and nothing survived to a verifiable vote or referral I'd cite. That's consistent with the weeks on either side. The next verifiable policy noise doesn't hit until June 9 — personnel churn at the White House AI office and a new executive order leaning into a national security frame — then June 10-12 with Amodei's preemption push and the Anthropic export control. May 26 to June 1 was before all of that. I'm not going to fill the gap with rumors. If a regulator did something material that week without posting it to a primary, I couldn't anchor it. ### What did — and why it counts as policy The labs shipped governance you can actually buy. That's policy now whether we call it that or not. **Enterprise controls.** OpenAI's late-May motion was SCIM provisioning, workspace controls, and audit-friendly handling for ChatGPT Enterprise/Team. Anthropic's was parallel — Claude Enterprise updates around context handling and integration surfaces that unblock legal review. Neither got a model version bump I'd pin to May 26-June 1 from a primary fetch (Anthropic's newsroom is client-rendered and throws on a raw pull), but the changelog pattern is clear: same intelligence, more toggles your CISO will sign. That's not a feature update. It's a permission slip. You don't get 10,000 seats without SCIM and logs. **Distribution as industrial policy.** Same week, the enterprise signal was who gets to *sell* the models, not who built the best one. Late May sits right before the June 9-10 distribution blitz — OpenAI onto Oracle via credits, Anthropic leaning on SI partners — and you can see the prelude in partnership language and marketplace positioning. When the bottleneck stops being "is it smart enough" and starts being "can I buy it through my existing contract," channel *is* policy. Google played it the same way on Vertex — no new Gemini flagship with a clean date inside this window, just garden updates and grounding hooks that platform teams need before they'll wire a model into a workflow. Boring. And exactly what moves adoption. ### The takeaway We keep treating AI policy as law that either passes or doesn't. This week is a reminder it mostly happens as defaults. No regulator told enterprises AI was auditable last week. OpenAI and Anthropic just shipped the controls that let procurement say yes anyway. No trade rule moved a model — a marketplace listing did. That's less satisfying than a headline about a ban or a bill, but it's durable. Laws take months. An admin toggle takes a sprint. ### What I'm watching 1. **State vs. federal.** The federal preemption fight that breaks open June 10 was already brewing — if Washington doesn't write rules, states will. This quiet week is the calm before that split. 2. **Allowlist precedent.** The June 12 export-control fight over Fable/Mythos will test whether "trusted partner" allowlists become a de facto license. Watch how late-May enterprise controls get repurposed for that. 3. **Spend and audit.** If your vendor just gave you usage analytics and caps, ask what they actually log. That's your governance stack for H2. No law passed. But the stack got more governable. That's how most AI policy gets made. === # Flagship Watch: Nothing New the Week of May 19 URL: https://brianbest.com/blog/2026-06-01-ai-week-w12a-models Date: 2026-06-01 Tags: ai-models, benchmarks, weekly-roundup Summary: May 19–25 was a quiet week at the frontier — no verified flagship model release and no new benchmark card to score. I went looking for a flagship and came up empty. For May 19–25, I checked the usual newsrooms — Anthropic, OpenAI, Google DeepMind, Meta, xAI, Mistral, DeepSeek — for a dated model drop or a new benchmark table. Nothing verified landed inside the window. That's not me missing a tweet. It's what the primary sources show. ### What didn't ship No new weights hit an official newsroom with a `publishedOn` between May 19 and May 25. The last verifiable flagships are still a year old: Claude 4 (Opus 4 and Sonnet 4) published May 22, 2025, and the Gemini 2.5 updates at I/O on May 20, 2025. Inside May 19–25, 2026 there's no Opus 5, no Gemini 3, no GPT-5 GA, no Grok 4, no Llama 4 follow-on I could pull from a post or model card with a clean date. I verified via Wayback `id_` captures — live fetches for `anthropic.com/news` and `blog.google` just return JS shells right now, so the archive is the honest check. CDX for that ISO week shows re-crawls of older slugs, not a new announcement. If something dropped only on X without a newsroom post, it didn't clear my bar. I'm not counting a rumor as a release. Pricing and context windows didn't move either. Anthropic still lists Opus 4 at $15/$75 and Sonnet 4 at $3/$15 per million tokens. Google still lists Gemini 2.5 Pro at 1M tokens. No quiet doc-page bump in this window. ### Benchmarks: nothing to score No new model means no new numbers. I swept SWE-bench Verified, Terminal-bench, MMLU, GPQA Diamond, MMMU, AIME, LiveCodeBench, and LMArena for anything tied to a fresh card dated in this week. Nothing. For context, the reference SOTAs are still those May 2025 cards: * **Claude 4:** 72.5% (Opus 4) and 72.7% (Sonnet 4) on SWE-bench Verified, 43.2% on Terminal-bench — without extended thinking. With extended thinking (up to 64K tokens) plus their internal scorer that samples multiple patches and filters on visible tests, Anthropic reports 79.4% and 80.2%. * **Gemini 2.5 Pro:** Google claimed ELO 1415 on WebDev Arena and "leading across all LMArena leaderboards," plus 84.0% on MMMU for the experimental Deep Think mode — trusted-testers-only via the Gemini API at the time. Neither got updated May 19–25. ### Why the quiet matters A year ago this same week was loud — Claude 4 and Gemini 2.5 within 48 hours. This year it's silent. That's the story. It looks like a digestion quarter. Anthropic paired Opus 4 with ASL-3 deployment and security controls. Google held Deep Think back for extra frontier safety evals. When the next model needs a classifier fallback and export paperwork, you don't rush a chart. The weeks after this one (May 26 to late June) show the same pattern — labs choosing distribution and safety over another point on MMLU. ### What I'd do with a gap week Lock your baseline. Re-run the tasks you actually route in prod — your tool-use loop, your long-context case, your cost per completed run — and log pass rate, latency, and fallback rate. That's the yardstick the next flagship has to beat. You'll know in an hour whether the price bump is worth paying for. No new model this week. I'll keep watching the pricing and context-window pages — those sometimes move without a blog post — and call it out when they do. === # This Week in AI Products: No Flagship, Just Plumbing URL: https://brianbest.com/blog/2026-06-01-ai-week-w12b-products Date: 2026-06-01 Tags: ai-weekly, enterprise-ai, product-launches Summary: May 19–25 didn't ship a flagship model. It shipped the unsexy stuff that decides enterprise adoption anyway — admin controls, cost scrutiny, and trust boundaries. If you waited for a flagship model between May 19 and 25, you waited all week. I couldn't verify a new GPT, Claude, or Gemini version pinned to this window. Direct vendor newsrooms were JS-gated and throw on raw fetch, and the top Hacker News index for the week shows no benchmark or pricing sheet in the top 100. That absence is the story. This was a packaging week, not a parameters week. ### Packaging over parameters The labs weren't competing on weights. They were competing on whether IT will let you use the weights you've already got. On the OpenAI side, the visible motion was ChatGPT Enterprise/Team admin hardening — SCIM provisioning, workspace controls, audit-friendly data handling. Same model, easier to approve. Anthropic's late-May surface was similar: Claude Enterprise integration surfaces and context handling that don't earn a benchmark tweet but do unblock legal review. Google's was Vertex plumbing — model garden updates and grounding hooks labeled "in preview," the kind your platform team files a ticket for. None of it demos well on stage. All of it determines whether a deployment actually ships. I wrote about the other half of this on May 20 that week: [MCP is the real unlock](/blog/2026-05-mcp-is-the-real-unlock). The Model Context Protocol doesn't make models smarter. It makes tools composable — `npx @modelcontextprotocol/server-confluence` instead of 300 lines of custom tool glue wired into agent code. That decoupling is product too, just not the kind labs announce. When teams can ship an MCP server in an afternoon without touching prompt logic, adoption stops being blocked by model choice. ### Enterprise adoption: cost and trust hit the dashboard Two headlines captured where enterprise buying actually stalled that week. First, cost. On May 25, Uber's COO was quoted saying it's getting harder to justify money spent on "tokenmaxxing" (HN 293 pts, Business Insider piece that 404s on direct fetch — title-verified, body-unverified). Paired with a May 22 "Cheap AI could derail IPOs" narrative floating the same window, the signal is clear: per-token spend hit the CFO dashboard. Labs heard it — the later July price cuts to sub-dollar per-million tokens didn't land until summer, but late May is when buyers started pushing back. Second, trust. Also May 25: Microsoft Copilot Cowork exfiltrates files (PromptArmor, 264 pts) and CVE-2026-28952 in Apple macOS Tahoe 26.5 — an HN title that attributes the find to Claude, though Apple's support page 127115 itself doesn't mention Claude in the fetched head. I flag that attribution as unverified until the full disclosure renders. But the pair matters: one story is AI breaking a security boundary, the other is AI finding a kernel vuln. Enterprises buy both risks at once. Then there's the odd one that still counts as procurement signal: the Vatican-Anthropic relationship surfacing May 22-25 around Pope Leo's `Magnifica humanitas` and AI ethics (Religion News, AP cross-link). When ethics frameworks get a Vatican co-sign, that's not philosophy — that's governance language your compliance team can cite. No story that week came with seat counts or ARR. The enterprise tell was distribution: who gets to resell, who gets marketplace placement, who gets the admin toggle your CISO will sign. That's been the pattern since spring. It continued here. ### What this week actually meant We frame progress as "new model or nothing" and miss the week where deployment actually moves. No flagship dropped May 19-25 and enterprise adoption still shifted — toward cheaper retrieval, scoped credentials, and audit logs. If you're building product, that shift is your head start. You don't need to wait for bigger context. You need packaging your buyer can approve. ### What I'd do Monday 1. **Scope your tools.** If you're running agents with broad tool access, steal the May 20 MCP pattern: per-task allowlists and scoped creds with every tool call logged. Copilot Cowork is your warning shot — boundaries are the product. 2. **Price your loops.** Benchmark a small retriever against a frontier call on your actual workload. Even a 10x delta changes what you can afford to loop before finance asks questions. 3. **Check your admin center.** Pull the late-May SCIM and workspace controls for ChatGPT Enterprise / Claude Enterprise. Getting IT to flip those toggles is your rollout, not the next model. Quiet week for weights. Busy week for plumbing. And plumbing is what ships. === # The Week Open Source Tried to Train Smarter URL: https://brianbest.com/blog/2026-06-01-ai-week-w12c-opensource Date: 2026-06-01 Tags: open-source, ai-weekly, research Summary: May 19-25 had no flagship model drop — instead, open source shipped efficiency. A $1,500 pretrain, a 3.8B image model on 19% compute, and RLVR that finally cares about tokens. No new frontier base model dropped this week. That *is* the story. May 19-25, open source didn't chase scale. It chased efficiency — fewer tokens, denser data, and sharper credit assignment. If you've been priced out of pretraining, pay attention. ### The $1,500 pretrain **HRM-Text** — [2605.20613](https://arxiv.org/abs/2605.20613), Sapient AI — replaces Transformers with a Hierarchical Recurrent Model: slow strategic layers plus fast execution layers, stabilized by MagicNorm. The numbers are the point. 1B params, 40B unique tokens, ~$1,500 to train from scratch. As-reported: 60.7% MMLU, 81.9% ARC-C, 84.5% GSM8K, 56.2% MATH — competitive with 2-7B open models that used 100-900x more tokens and 96-432x more compute. Their math, not mine, but the direction matters. And it's fully open: ```bash git clone https://github.com/sapientinc/HRM-Text ``` I wouldn't replace my 7B with it tomorrow. But it makes the "only labs can pretrain" argument a lot weaker. ### Dense captions beat more parameters Microsoft's **Lens** — [2605.21573](https://arxiv.org/abs/2605.21573) — is a 3.8B text-to-image model that claims to match >6B SOTA on ~19.3% of Z-Image's compute. No architecture trick. Just data density: **Lens-800M**, 800M pairs captioned by GPT-4.1 at ~109 words on average, plus multi-resolution batching. If you're still training T2I on alt-text, that's your bottleneck. ### RLVR gets surgical Verifiable rewards score the whole response, but learning happens per token. Two papers attacked that gap: **DelTA** — [2605.21467](https://arxiv.org/abs/2605.21467), RUC AI Box — reframes the policy gradient as a linear discriminator over token gradients and reweights tokens to make positives and negatives more contrastive. As-reported: +3.26 avg on 7 math benchmarks (Qwen3-8B-Base), +2.62 on 14B. Code: [github.com/RUCBM/DelTA](https://github.com/RUCBM/DelTA). **Gated DeltaNet-2** — [2605.22791](https://arxiv.org/abs/2605.22791), NVIDIA — fixes linear attention itself by decoupling erase and write with channel-wise decay. Linear-time mixing, constant-memory decoding. If KV-cache is your wall, read it: [github.com/NVlabs/GatedDeltaNet-2](https://github.com/NVlabs/GatedDeltaNet-2). Same lesson: RLVR is moving from "reward the answer" to "reward the right token." ### Skills as weights **SkillOpt** — [2605.23904](https://arxiv.org/abs/2605.23904), Microsoft Research — treats agent skills like weights. Frozen agent, external skill file, optimizer proposes bounded add/delete/replace edits — accepted only if held-out validation strictly improves. Text-space SGD with a learning-rate budget. Repo: [github.com/microsoft/SkillOpt](https://github.com/microsoft/SkillOpt). It landed the same week GitHub saw a harness surge: Moonshot's **kimi-code** CLI and Anthropic's defending-code harness, both created May 22 with 6k+ stars. We're optimizing harnesses faster than base models. ### What I'd try Monday Run HRM-Text on your own eval slice and see where 40B tokens hurts. If you train T2I, recaption 10k images densely before you buy more GPUs. And if you do RLVR, read DelTA and Gated DeltaNet-2 together — one fixes the loss, the other fixes the attention. Bigger wasn't the story this week. Better use of tokens was. === # This Week in AI Policy: Two Fintech Orders and a $20M Bet on Shop-Floor AI URL: https://brianbest.com/blog/2026-06-01-ai-week-w12d-policy Date: 2026-06-01 Tags: ai-policy, regulation, industry-moves Summary: No AI Act dropped May 19-25, but two executive orders rewrote the playbook for AI in finance and NIST put $20M per pilot behind AI on the factory floor. If you were waiting for a capital-A AI law last week, it didn't show. For May 19-25, the verifiable action wasn't an AI Act or an EO with "artificial intelligence" in the title. It was financial regulation that *is* AI regulation if you read the fine print, plus a manufacturing bet that says where Washington thinks AI competitiveness actually lives. I pulled the Federal Register for that window — 398 documents total, 5 with an "artificial intelligence" hit. Here's what cleared the bar. ### Fintech is AI policy by another name **May 19 (signed), May 22 (published): Two companion orders.** EO 14405, *Integrating Financial Technology Innovation Into Regulatory Frameworks* (91 FR 30475), and EO 14406, *Restoring Integrity to America's Financial System* (91 FR 30479), landed together. 14405 is the one to read if you build with AI in fintech. Its definition of "fintech firm" explicitly sweeps in "any application or any digital or online technology that facilitates access to, management of, or data processing for financial products or services" — plus digital assets and blockchain-based services. That's AI data processing without saying "frontier model" on the cover. The directive: within 90 days, every designated federal financial regulator — CFPB, SEC, CFTC, FDIC, OCC, NCUA — has to review existing regulations, guidance, supervisory practices, and application processes that impede fintech-bank partnerships or charter applications, and flag what to update. The framing is innovation-meets-oversight: open the pipes for tech integration while keeping safety, consumer protection, and supervision. 14406 is the counterweight. It orders Treasury to issue a formal Advisory within 60 days on exploitation of the financial system by non-work-authorized populations and employers — red flags and typologies for payroll-tax evasion, SSN mismatch, and related illicit finance. Same tool (financial regulators), opposite pressure: tighten KYC and credit-risk guardrails while 14405 loosens product-integration friction. If you squint, that's the administration's AI-in-finance stance in one week: make it easier to ship AI-driven financial products, harder to use the rails illicitly. ### NIST put AI on the shop floor **May 20: NIST MEP pilot notice (91 FR 29459).** The NIST Hollings Manufacturing Extension Partnership announced a competitive pilot — about $20M per pilot project over 2 years, from FY25-26 appropriations — for two topics: additive manufacturing for aerospace components and domestic critical-minerals supply chain. Only current MEP Center awardees/consortia can apply. Why it matters for AI policy watchers: the notice explicitly frames the push as "transformative technologies such as artificial intelligence, automation, and additive manufacturing are driving leaps in productivity." The biggest named "AI" mention that week didn't come from a lab. It came from manufacturing and supply-chain competitiveness — the shop floor, not the chatbot. Small dollars, clear signal. AI competitiveness in this framing isn't just model weights; it's whether you can print a part or secure a mineral without an adversary in the loop. ### Education kept AI in the grants **May 22: Department of Education final priority (91 FR 30291, ED-2025-OS-0745, effective June 22).** This one finalized the "Promoting Patriotic Education" priority — but it matters because it *re-affirms* the earlier supplemental priority on "Advancing Artificial Intelligence in Education" from April 13 (91 FR 18774) as part of the grant stack. No new AI-education rule this week; durability is the point. If you're writing a discretionary grant for ED this summer, you're now orienting to both priorities. Also on the calendar: IRS's Electronic Tax Administration Advisory Committee (ETAAC) noticed May 21 (91 FR 30026) that its June 17 public meeting will include "Artificial Intelligence and Human-Centered Design" on the agenda. Administrative, but it's where AI-meets-implementation details get hashed out. ### What didn't happen I couldn't verify a dated EU AI Act update, UK announcement, or major industry partnership/acquisition inside May 19-25. EUR-Lex's regulatory framework page showed no May news block when pulled, Gov.uk announcements returned no dated hit, and Reuters/TechCrunch feeds were either captcha-walled or empty for that window. OpenAI's RSS showed relevant posts clustering in July-August, not late May. That's worth calling out. A lot of roundups will backfill this week with headlines from June. They didn't happen yet. ### What I'm watching 1. **The 90-day clock.** 14405's regulator reviews are due around August 19. Watch OCC/FDIC/SEC dockets for the actual "what we'll update" lists — that's where AI underwriting, fraud detection, and data-processing rules either ease or don't. 2. **The 60-day Advisory.** Treasury's 14406 typologies will tell banks what AI-driven monitoring they need to add, not just what fintech product they can ship. 3. **MEP pilot to procurement.** If NIST's $20M pilots move from notice to award, follow which MEP centers win. That's your early map of where "AI + additive" gets subsidized. Back next week with models. This week belonged to regulators who never said "AI" out loud. === # Write Your Blog for the Agent That Will Cite It URL: https://brianbest.com/blog/2026-05-writing-for-agent-readers Date: 2026-05-25 Tags: agentic-ai, writing, content-strategy Summary: A growing share of traffic to technical blogs comes from agents fetching pages to answer someone else's question. Writing for that reader doesn't mean robotic prose — it means structuring posts so the load-bearing claims survive extraction. On a technical blog, the first entity to read your post is increasingly not a person with a coffee — it's an agent fetching the page to answer someone else's question. That changes what "well-written" means in practice, and not in the way most people expect. So I started writing for that reader. Not *instead* of the human — the prose still has to be good — but with the assumption that **the first entity to read any post will be a model deciding what to extract from it.** ## The agent reads differently than you think A human skims: they catch the bold words, jump to a code block, infer the rest from layout. An agent doesn't get the layout. It gets a token stream. Anything you encoded in visual hierarchy — a callout box, a clever heading, a diagram — collapses into flat text or disappears entirely. That means the question isn't "does this look scannable?" It's "if you stripped every visual cue and handed the model raw markdown, could it still find the point?" Most posts fail that test. They bury the thesis under three paragraphs of throat-clearing and put the actual insight inside a screenshot. > If a claim only survives when a human notices the layout, an agent will never cite it. ## What I changed Six habits, all cheap, all mechanical: - **Front-load the thesis.** First paragraph, plain language, no "in this post I'll explore." An agent that reads only the `summary` and the opening should be able to cite me correctly. If it can't, the post is misfiled before it's read. - **Make headings an outline, not a punchline.** I write headings so the heading list *alone* reconstructs the argument. "The shift" tells a model nothing; "The agent reads differently than you think" tells it what the section claims. - **Bold marks the takeaway, and only the takeaway.** When I bold exactly one clause per section, bold becomes a reliable signal that means "extract this." When I bold for emphasis-spam, it means nothing. - **Wrap every identifier in backticks.** File paths, function names, model IDs, flags — `getPost()`, `content/blog/`, `gpt-4.1`. Inline code turns ambiguous English into a token a model can match exactly. - **Tag every code fence and make it self-contained.** Even when my site doesn't syntax-highlight, the language tag is a strong signal to the reader that matters. No "see the snippet above" — the agent may have chunked "above" into a different request. - **End with extractable takeaways.** A short list of standalone claims is the part models quote most. Write each bullet so it survives being lifted out of context, because it will be. ## The thing people get wrong This is not "write for SEO" or "write for robots." The failure mode I keep seeing is people sanding their voice off to sound machine-friendly, which produces prose no human finishes and no model finds interesting enough to cite anyway. The two goals don't conflict. **Structure is for the agent; voice is for the human; you don't have to trade one for the other.** A strong opinion stated in the first sentence and marked in bold is both more human *and* more extractable than the same opinion hedged across a paragraph. Good structure was always good writing. Agents just punish bad structure faster than humans do. ## The bet - Well-structured posts survive extraction better than clever-but-buried ones — that was true for humans and is even more true for agents. - The fix is structural, not stylistic: front-loaded theses, descriptive headings, disciplined bolding, backticked identifiers, self-contained code, extractable takeaways. - Writing for agents and writing for humans converge — both reward putting the point where it can be found. If you have to choose, optimize for the reader that decides whether anyone else ever sees your post. --- *This site is itself an experiment in that bet: every post exposes `Copy as markdown` and `Send to chat`, and the whole site is available at `/llms-full.txt`. If the structure works, an agent should be able to cite this post from raw markdown alone.* === # MCP Is the Real Unlock for Agentic Engineering URL: https://brianbest.com/blog/2026-05-mcp-is-the-real-unlock Date: 2026-05-20 Tags: mcp, agentic-ai, tooling Summary: After a year of building agents at Axonify, the thing that actually changed how my team ships isn't a smarter model — it's the Model Context Protocol. Here's why. A year ago, every "agent" I built was a bespoke pile of tool definitions, retry loops, and JSON-schema babysitting. Each new integration meant teaching the model a new dialect, hand-rolling auth, and praying the function-calling format didn't drift between providers. It worked, but it didn't compose. Then MCP quietly turned the problem inside out. ## The shift The Model Context Protocol gives tools a standard shape — a server speaks MCP, a client speaks MCP, and the model on the other end doesn't care who wrote either side. That sounds boring. It is not boring. It is the difference between "I built an agent that can do X" and "I plugged my agent into the ecosystem of things that already do X." ### A concrete swap Our Confluence integration started as ~300 lines of custom tool definitions, auth handling, and pagination logic wired directly into the agent. Every change to the Confluence API meant touching agent code. Swapping it for an off-the-shelf MCP server looked like this: ```jsonc // Before: custom tool wired into the agent // ~300 lines: schema, auth, retry, pagination — all in agent code // After: MCP server as config { "mcpServers": { "confluence": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-confluence"], "env": { "CONFLUENCE_TOKEN": "${CONFLUENCE_TOKEN}" } }, "codebase": { "command": "node", "args": ["./mcp/codebase-server.js"], "env": { "GITHUB_TOKEN": "${GITHUB_TOKEN}" } } } } ``` Same capability, but tool maintenance decoupled from agent logic. The junior who shipped the codebase MCP server didn't need to know anything about prompt engineering or context windows — they just built a server that speaks MCP. That felt like the first time I used `npm install` instead of vendoring a library by hand. ## What changed for my team - **Tool development decoupled from agent development.** A junior dev can ship an MCP server in an afternoon. They don't need to know anything about prompt engineering, model selection, or context windows. - **Tools became testable in isolation.** MCP servers are just servers. You unit-test them like servers. You don't need an LLM in the loop to know the tool works. - **Composition got cheap.** We routinely run agents with eight or ten MCP servers attached. What used to be a custom orchestration nightmare is now a config file. ```mermaid flowchart LR A[Agent] -->|MCP| B[Confluence] A -->|MCP| C[Codebase] A -->|MCP| D[Email] A -->|MCP| E[Support KB] B & C & D & E -->|scoped creds + audit log| F[Enterprise boundary] ``` ## What I still don't love **Discovery is still unsolved.** There's no good answer yet for "which of my 47 connected tools should the model reach for in this context?" What we do today is pragmatic but manual: per-task allowlists that filter tool exposure before the agent sees it. ```jsonc // Our current band-aid — per-task tool scoping { "task": "triage-support-ticket", "allowedTools": ["search_kb", "read_logs", "get_similar_tickets"], "blockedTools": ["send_email", "write_code", "deploy"] } ``` It works, but it's a convention layered on top of the protocol, not part of it. Every team I talk to has invented some version of this. **Permission scoping is also early.** The protocol punts most of the auth model to the implementer, which means every team reinvents approval flows — who can call what, when, with whose credentials, and with what audit trail. We built ours around credential isolation (agents never see raw tokens), per-tool approval gates, and structured audit logging where every tool call is attributable. It holds up, but I expect the protocol itself to converge on a shared model in the next year. ## The bet If you're starting an agent project today and not building on MCP, you're choosing extra integration work for no durable benefit. The ecosystem is moving there — package your tools as MCP servers and your agents stay portable even as models change. That said, "won" overstates it: discovery and auth are still yours to solve. Build on it, but budget for those gaps. === # Stop Prompt Engineering. Start Context Engineering. URL: https://brianbest.com/blog/2026-04-context-engineering-over-prompt-engineering Date: 2026-04-12 Tags: llms, prompting, context Summary: Prompt engineering was a useful crutch when models were smaller and more brittle. With today's frontier models and large context windows, the real lever is what you put in the context, not how you phrase it. I've watched a lot of engineers spend hours tweaking a system prompt to get a different answer out of a model. Most of the time they're solving the wrong problem. The model isn't being stubborn. It just doesn't know what you know. ## Prompt engineering is a 2023 skill In 2023, prompt engineering was load-bearing because models were brittle. Tiny phrasing changes flipped outputs. "Think step by step" added points on benchmarks. Few-shot examples were the difference between usable and useless. That world has narrowed. Frontier models have gotten good enough that *how* you ask matters far less than *what they have to work with* when you ask — though truncation, retrieval quality, and tool design can still make a good prompt necessary. ## What context engineering looks like When I review an underperforming agent today, I almost never touch the system prompt first. I ask: - **What does the model actually see at decision time?** Half the time the answer is "less than I assumed." A retrieval step is silently truncating. A tool result is being summarized into uselessness. The agent is one cache layer away from blindness. - **Is the context organized for the model, or organized for the engineer who wrote it?** Logs dumped in raw form. JSON blobs the model has to parse mentally. Stale memory that contradicts current state. Context hygiene is a real skill. - **What's the signal-to-noise ratio?** A large context window doesn't help if most of it is irrelevant. Models still attend better to short, dense context than long, sparse context — even with bigger windows. ## A concrete example Last sprint we had an agent that kept misclassifying support tickets. The team's first instinct was to rewrite the system prompt. I asked them to print the actual context the model received on a failing case. The retrieval step was returning the right knowledge-base article. But it was returning it *after* roughly 40k tokens of unrelated chat history, with no separator, no labeling, and no instruction to prioritize it. The model wasn't ignoring the article — it was drowning in everything else. Here's the shape of what the model actually saw (redacted): ```text # What we assumed the model saw: [KB article: "How to reset learner progress" — 400 tokens, high relevance] # What the model actually saw: [chat history — ~40k tokens, no section markers] [KB article — 400 tokens, appended at the end, no label, no priority cue] [system prompt tail — "Use retrieved context to classify the ticket"] ``` We didn't change a word of the prompt. We restructured what got pasted in — labeled sections, the most relevant retrieval at the top, and a short instruction to prioritize it. On our internal eval set of ~200 labeled tickets, classification accuracy rose by 18 percentage points (from the low 70s to the low 90s) on the next run. Eighteen points from formatting. Zero from prompting. ## The takeaway Prompts are interfaces. Context is the actual input. If your agent is underperforming, look at the input before you blame the interface. === # Shipping With Agents, Not Instead of Them URL: https://brianbest.com/blog/2026-03-shipping-with-agents-not-instead-of-them Date: 2026-03-08 Tags: agentic-ai, developer-experience, axonify Summary: The 'AI replaces developers' framing is bad for both your career and your codebase. After two years of agent-assisted shipping at Axonify, here's what actually works. The discourse online keeps oscillating between two bad framings: agents will replace developers, or agents are toys that produce slop. Both are wrong in ways that are unhelpful for anyone trying to actually ship software. What's true is more interesting. Agents change the unit of work, not the worker. ## The unit-of-work shift Pre-agent, the unit of work was "a change I will type." Post-agent, the unit of work is "a change I will review." That sounds like a small shift. It is not. It restructures almost every part of how a senior dev spends their day: - **Reviewing matters more than typing.** If an agent can write a passable PR in ten minutes, the bottleneck is the human who has to decide whether it's correct. Review velocity is now the rate-limiter on shipping. - **Tests become specifications, not chores.** When the agent reads your tests to figure out what "done" means, weak tests produce wrong code. Strong tests produce right code. Test discipline went from "good practice" to "load-bearing." - **Context-setting is the real prompt.** "Build this feature" produces garbage. "Build this feature, here's the existing pattern in `payments/charges.ts`, here's the constraint we have around idempotency keys, here's the test fixture that proves it works" produces something you can merge. ## What I tell my team Three rules of thumb that have held up across about 18 months of agent-assisted shipping: 1. **Never accept code you can't read.** If the agent produces 400 lines you don't fully understand, you don't have working code — you have a bug you haven't found yet. Make it explain. Make it simplify. Reject anything you can't defend in review. 2. **Use agents for the parts you find boring.** Boilerplate, migrations, test scaffolding, refactors with clear acceptance criteria. Don't use them for the parts you find interesting — that's where you build the taste that makes you valuable. 3. **Treat agent output like a junior dev's PR.** It deserves the same review rigor, the same "why did you choose this approach?" pushback, and the same expectation that you'll catch and explain the mistakes. Every agent PR on our team goes through the same checklist we use for human PRs: ```text review-checklist.md — applied to every agent PR - [ ] Can I explain this diff without re-reading the prompt? - [ ] Do tests encode the constraint, not just the happy path? - [ ] Is there a simpler shape that passes the same tests? - [ ] Would I be comfortable owning this code after the agent is gone? ``` If the answer to any of those is no, the PR doesn't ship — no matter how green the CI is. ## What I don't tell my team (but should) The senior devs on my team got more valuable, not less, when we adopted agents. The juniors had a harder time, because the gap between "I can prompt my way to working code" and "I can tell when working code is wrong" is the gap that experience used to be. I don't have a clean answer for closing that gap, but we have a working hypothesis we're iterating on: - **Paired review on every agent PR** — a junior authors the prompt and the senior reviews the output with them, narrating what they're checking and why. The point isn't gatekeeping; it's transferring the taste for what's wrong. - **A shared review rubric** (like the checklist above) so "good" isn't vibes — it's four checkable boxes. - **Guardrails, not freedom** — juniors start with agents scoped to well-tested areas (migrations, test scaffolding, mechanical refactors) before they get the keys to ambiguous product work. It doesn't fully close the gap. But "agents will figure it out" definitely won't either.