AI News Briefing — Altman and Musk endorse Amodei's pacing plan

Model releases

• [2026-09-14] Three rivals signed on to Dario Amodei’s pacing plan inside a day. Sam Altman called independent evaluators with employee-like access “a great idea, and we will do the same”; Elon Musk posted “Dario is right”; Satya Nadella endorsed the framework over the weekend. Altman separately ruled out an OpenAI listing this year. The Information reports Anthropic, OpenAI and Google were already in quiet talks about a shared safety standards body — step two of the plan, minus the announcement. (source, source, source)

  Three endorsements inside a day reads as momentum. The Information’s reporting puts the talks before the speech, which makes this an announcement of something already moving rather than a proposal anyone had to weigh.

• [2026-09-13] Washington did not move with them. Trump put the warnings down to “very negative forces” and said “whoever wins AI, wins”. Speaker Mike Johnson ruled out an emergency session, arguing that regulating now hands the race to China, while allowing that some guardrails are needed. The Register reads the three steps as a list of asks — chip controls on China, and a standards body the incumbents would help staff. (source, source, source)

  No emergency session means no deadline, so every commitment in the story above is one the labs set and keep themselves.

• [2026-09-14] DeepSeek called off the cutover it had scheduled for today. V4 Pro requests were due to start landing on V4.1-Flash at Flash billing from 04:00 UTC; the changelog now says the API keeps serving V4 Pro past September 14 with billing unchanged, after customers asked for it. The older deepseek-v4-flash and deepseek-v4-flash-vision-exp names stay retired and still route to V4.1 Flash. (official)

  Deferred, not cancelled. The changelog names no new date, so the migration is still owed and now has no clock on it.

  For ML / Data Engineers: While you are in the config, grep for deepseek-v4-flash and deepseek-v4-flash-vision-exp as well — those two stayed retired, so any harness still pinned to them has been scoring V4.1 Flash rather than the model its config names.

• [2026-09-12] Microsoft added Grok to Copilot inside Word, Excel and PowerPoint through its Frontier Program. An admin has to enable it per tenant, and it is not offered to Frontier customers in the EU, EFTA or the UK during the preview. (official)

  Per-tenant enablement makes this an admin decision, not a user one. The EU, EFTA and UK carve-out means an org spanning those regions cannot offer the same Copilot everywhere while the preview runs.

  For Solution Architects: Enabling the Frontier Program is a tenant-wide default, so decide it as one — which model answers inside Word, Excel and PowerPoint, and whether your regional footprint permits a single answer at all right now.

MCP

• [2026-09-14] MCP has an exam now. The Agentic AI Foundation and Linux Foundation Education opened enrollment for the Model Context Protocol Associate, 120 minutes, online and proctored, pinned to the 2026-07-28 spec. The domain weights say where the protocol’s hard parts sit: interactions and execution 26%, security and governance 24%, fundamentals 16%. (official, details)

  Pinning the exam to the 2026-07-28 spec means a pass certifies that version of the protocol rather than the protocol in general — worth knowing before treating the credential as a durable one.

Practice & craft

• [2026-09-14] Simon Willison published commit-rewriter 0.1 after Datasette’s security-release commits turned out to carry coding-agent cruft and issue IDs from a private repository. uvx commit-rewriter path/to/repo opens a local web UI; it cuts a timestamped branch before rewriting anything from the first edited message forward. If an agent writes your commit messages and your history is public, that is the failure this addresses. (source)

  Worth noting where the cruft surfaced: Datasette’s security-release commits, which are the ones outsiders read most closely.

  For Software Developers: Point it at a repo where an agent has been writing your commit messages and read what is actually in there before deciding whether to rewrite anything. A timestamped branch is cut before the first edit, so looking costs one branch you can delete.

Watch list

• OpenAI’s evaluator access, quote to mechanism. Altman said OpenAI will match Anthropic’s employee-like access for outside evaluators. A named evaluator and published terms settle it; a post does not.

  Neither lab has named who such an evaluator would be. Access on employee terms is a contract before it is a policy, so a name is what arrives first or not at all.

• The safety standards body. The Information puts Anthropic, OpenAI and Google in talks about one. A charter or a founding roster is the artifact — three labs talking has been reported before and produced the 2023 Frontier Model Forum.

  What to read in a charter, if one appears: whether it names thresholds or only topics. That is what would separate this from the body it would otherwise resemble.

• OpenAI’s misalignment disclosure framework. Seventeen days to Senator Josh Hawley’s October 1 document deadline, still nothing published.

  With a deadline somebody else set and nothing out yet, the first public version of this may turn out to be a document produced for a subcommittee rather than a framework posted for everyone.

• AWS’s bedrock-agentcore namespace. Off September 17; three days for the grep.