1,324 frontier-lab employees asked Washington to pace their own field

Two open letters in nine days, and the signature lists carry the argument better than the text does. Simon Willison pulled both together this morning. “Open Weights and American AI Leadership,” shepherded by Microsoft and dated July 24, has 235 AI-adjacent companies on it: NVIDIA, Amazon, Y Combinator, the Linux Foundation, and OpenAI as a later addition. “Pacing the Frontier,” published July 28, carries 1,324 employees of frontier AI companies, including Jakub Pachocki, Ilya Sutskever, Dario Amodei, and Jack Clark, asking the U.S. government to support an international effort to deliberately pace automated AI development. Anthropic is absent from the first list and present on the second, and put out its own position on open-weights models three days after the Microsoft letter went out.

The open-weights letter makes a security argument rather than an ideological one, and that is the part worth borrowing: closed models “can be breached, misused, or fail in ways that outsiders cannot detect,” and consolidating capability behind a handful of providers creates single points of failure nobody outside those providers can audit. It also defends distillation by name as legitimate model development, which sits awkwardly next to Dario Amodei’s call to crack down on industrial-scale distillation operations. The pacing letter is the more interesting document for anyone shipping agents, because its stated worry is not hypothetical. Anthropic says it produces 80% of its code with Claude Code, OpenAI credited Sol with a 20% cut to end-to-end serving costs, and Kimi K3 designed a chip to serve a nano model built on its own architecture. When 1,324 people who work at the labs write down that the loop is closing, the open question is what a pacing mechanism could look like when the compounding happens inside their own repos.

A rebuttal to yesterday’s math news went up on PhilArchive overnight and reached the Hacker News front page: OpenAI’s claimed disproof of Connes’ Rigidity Conjecture is invalid. OpenAI’s “Ten advances in mathematics and theoretical computer science” post ran Saturday, claiming results across geometry, cryptography, and complexity, with the Connes item among the headline claims. The rebuttal is early and thinly discussed, 23 points and four comments as it crossed the front page, so it reads as a live dispute rather than a settled one. The pattern underneath it is the thing to track: machine-generated proofs arrive faster than the review capacity to check them, and the checking is where the epistemics actually live. Verification lag is the recurring cost of this whole category, and it does not shrink when the generator gets better.

The sharpest practitioner writing of the day came from a Reddit thread. A developer sorted all 79 rules in their global CLAUDE.md against Boris Cherny’s advice at YC Startup School on Monday to delete your CLAUDE.md, your skills, and your hooks every six months and see what the model does. The audit produced a taxonomy worth stealing. Model patches, rules compensating for behavior the model should have gotten right, came to 17 of 79. Environment facts the model cannot observe came to 37: a zsh shell where ${PIPESTATUS[0]} reads empty so a build that exited 0 gets reported as failed, a repo under iCloud sync turning a one-second cold test run into four minutes. Preferences and decisions with no derivable right answer came to 25. Only the first bucket, 22% of the file, is what the delete-it advice targets, and Claude Code shipping Opus 5 with 80% of its system prompt cut is explicable precisely because a system prompt is almost entirely that one category. The best point in the post is about the experiment itself: “delete it and see what happens” is a test that mostly cannot fail, because a rule firing only on a deploy or a migration returns “seems fine” for every week you don’t deploy. Write down what each rule was for before deleting it, and “seems fine” becomes checkable.

That context-budget problem got a runtime answer from Google in the same window. Genkit Go added Agent Skills built on progressive disclosure: instructions, scripts, and references package into SKILL.md bundles where only the frontmatter metadata sits in the system prompt, and middleware loads the full body when a task matches the description. It is the same discipline the CLAUDE.md audit arrives at by hand, moved into the framework where it does not depend on anyone remembering to prune.

Governance moved on two fronts. GitHub restricted npm granular access tokens configured to bypass 2FA from performing sensitive account, org, and package management actions, which now require an interactive 2FA challenge, closing one of the larger credential-based attack paths in the registry and following the supply-chain disruption work GitHub described last week. GitLab published a piece on governing agentic AI, MCPs, and code assistants that names the structural version of the problem: code completion kept a human in the review loop by construction, since a person accepted every suggestion before it existed, and an agent that opens a merge request, calls a tool, and modifies CI configuration removes that property without supplying a replacement.

Two infrastructure notes worth a bookmark. A wafer.ai writeup on running Kimi K3 on MI355X, claiming better performance per dollar than B300, took 110 points on Hacker News, and inference economics on AMD silicon is a story that matters more for where serving capacity goes next year than any single benchmark table. Cloudflare introduced Meerkat, an internal globally consistent control-plane service built on the QuePaxa consensus algorithm instead of Raft, which permits leaderless writes while holding strong consistency.

What to watch: whether the Connes rebuttal survives contact with reviewers who work in the field, and whether anyone who signed the pacing letter proposes a mechanism concrete enough to argue about.