Monday, September 14, 2026

TL;DR

Anthropic's pacing plan turned into concrete commitments this weekend: permanent evaluator access, a Microsoft governance draft, a second lab losing a researcher to METR — while China, Trump and the market all pushed back, with AI stocks selling off Monday despite no lab cutting spend. Separately, practitioners split over whether a reported research speedup counts as real recursive self-improvement.

Act on this

  • Training a general-purpose model above 10^25 FLOPs for the EU market? File your first systemic-risk evaluation with the AI Office by September 15. EU AI Act tracker #

Signals

Zvi Mowshowitz says OpenAI and Anthropic have both entered a recursive self-improvement era — but researchers closest to the work aren't convinced #

Zvi's September 13 post flagged an unreleased OpenAI model that grew 'more capable than Astra across the board' within two weeks of training, calling that pace more alarming than the Navier-Stokes proof it produced. Two days earlier, John Schulman, Beren Millidge and Charlie O'Neill told Dwarkesh Patel they see diminishing returns and persistent bottlenecks in AI defining its own objectives, not a takeoff. Zvi · Dwarkesh Patel

Coding agents are starting to manage their own testing and coordination, not just write code  #

Thorsten Ball's September 12 newsletter describes routinely spawning isolated agent 'orbs' with Fable 5.1 and GPT-6 Astra that test features independently, generate video proof, and report back — with the lead agent adjusting its own prompts based on what they find, unsupervised. He calls it working at 'a higher altitude' than manual review allows. Ball

Agent transparency is becoming the harder problem than agent capability  #

Simon Willison's September 12 test had GPT-6 Astra chain OpenStreetMap tools into a working running route in 27 minutes — but the code behind it vanished once the conversation compacted, leaving no way to inspect how it was built. His point: agent frameworks need to preserve pre-compaction reasoning, not just final outputs. Willison

News

WATCHAnthropic to give outside evaluators permanent access; OpenAI commits to match  #

Anthropic will give third-party evaluators like METR employee-level, standing access to verify safety practices and training runs. Altman called it 'a great idea' and pledged OpenAI will do the same. Amodei · CNBC

WATCHMicrosoft drafts its own AI 'Code of Conduct' for public comment  #

Nadella said Microsoft will publish rules governing its first-party MAI models for public consultation, citing 'deliberate pacing' and floating its own embedded evaluators. Unite.AI

WATCHChina and Trump both reject the pacing call  #

China's commerce ministry called Amodei's essay proof of 'US technological hegemony'; state media called it 'hostile.' Trump told reporters 'whoever wins AI wins,' dismissing the warnings as exaggerated. via ANI · NPR

WATCHAI stocks sell off as pacing talk lands, spending doesn't  #

Nvidia fell 2.1% and AMD 4.8% premarket Monday; SoftBank dropped over 10% in Tokyo. Microsoft still guides toward $175 billion in fiscal 2027 capex, unchanged since Friday. Yahoo Finance

WATCHA second lab loses a safety researcher to METR  #

Google DeepMind's Josh Engels joined METR alongside Anthropic's Joe Benton, whose departure was already known; both told NBC that labs' incident transparency 'is entirely voluntary.' NBC News

ACTEU AI Act systemic-risk filings are due tomorrow #

Providers of general-purpose models trained above 10^25 FLOPs must submit systemic-risk evaluations, red-teaming methodology and energy-use disclosures to the EU AI Office by September 15. EU AI Act tracker

The long view

A year ago, outside evaluators like METR reviewed frontier models under access terms labs set and could revoke — time-limited testing agreements, not standing access. This week Anthropic committed to giving METR permanent, employee-like access, badges and all, as the first concrete step in Dario Amodei's pacing plan, and a second lab lost a safety researcher to METR itself over incident transparency that otherwise remains purely voluntary. If this holds, the test becomes whether what evaluators find ever gets published against a lab's own interest.

Also noted

  • Jesse Vincent's episodic-memory v1.6.0 added a fifth supported harness and AWS Bedrock summarization, September 8. obra/episodic-memory  #
  • The blocking CUDA crash in GLM-5.3-Flash's llama.cpp port was fixed September 11; the merge still awaits maintainer review. llama.cpp PR #27754  #
  • Claude Code v2.1.267 fixed a path-traversal bypass in marketplace plugin installs, September 9. Claude Code changelog  #
  • vLLM 0.29.0 shipped September 9, defaulting to Model Runner V2 and dropping ten deprecated model architectures. vLLM releases #