Wednesday, September 16, 2026

TL;DR

OpenAI, Anthropic and Google DeepMind confirmed weeks of joint safety-coordination talks and a shared AEF-1 evaluator standard, declining an antitrust waiver as critics call it cartel behavior — the same day a third lab researcher, Google DeepMind's Bilal Chughtai, resigned warning AI 'has the potential to kill us all.' Separately, Anthropic and OpenAI practitioners report agentic coding has outgrown their CI systems.

Act on this

  • Use Baseten for inference? A leaked token gave admin access to its production GitHub in 25 minutes — audit credentials baked into shared Docker images. Strix #

Signals

Agentic coding has outgrown the CI and test infrastructure built for human-paced commits  #

Anthropic's engineering team disclosed Sept 14 that Claude now authors roughly 80% of its own merged production code, with CI job volume up 25x in six months; Gergely Orosz reported Sept 15 that OpenAI's non-engineering teams went from near-zero Codex use to 90% adoption in four months, forcing continuous CI rework. Demonstrated, not argued. Anthropic · Orosz

Open-model routing rankings are turning over within days of a launch, not months  #

DeepSeek's V4.1 Flash — launched September 10 with a novel causal encoder-decoder architecture — climbed to challenge Tencent's Hy4 Preview atop OpenRouter's daily token-volume ranking within days, pushing GLM-5.3-Flash down the list, per Tokenmaxxing's tracking of OpenRouter's own usage data. Tokenmaxxing · Latent Space

AI-content detectors still misfire when text is human-edited, not purely AI-written #

Armin Ronacher's September 14 test fed the Pangram detector — which claims a 0.0041% false-accusation rate — text he had substantially rewritten by hand from an AI-generated structure; Pangram still flagged it as fully AI-generated, undercutting detector-based accusations of AI use. Ronacher

News

WATCHOpenAI, Anthropic and Google DeepMind confirm weeks of joint safety talks  #

OpenAI policy chief Chris Lehane said the three labs have coordinated on safety standards for weeks, including embedding independent evaluators — and said OpenAI does not believe it needs an antitrust waiver to do it. TechCrunch

WATCHAEF-1 sets a baseline standard for independent AI evaluators  #

The AI Evaluator Forum published AEF-1, a baseline standard covering evaluator access, conflicts of interest and transparency; Anthropic, OpenAI and xAI all cosigned it, per Latent Space's reporting. Latent Space · AI Evaluator Forum

WATCHAntitrust scholars call the labs' pacing coordination collusion  #

Truth on the Market's Dirk Auer argues the pacing coordination functions as a cartel — Amodei's plan asks rivals to agree on how much compute each may use, a textbook restriction on output under antitrust law. Truth on the Market

WATCHGoogle DeepMind safety researcher resigns, warns AI 'could kill us all'  #

Bilal Chughtai left Google DeepMind in July but went public Sept 14, warning AI agent swarms are already capable of escaping operators' control and conducting cyberattacks — a call for pacing before harm, not after. Whalesbook

SHIPMeta bundles AI usage into a new Meta One subscription #

Meta launched five-tier subscriptions, $3.99 to $499 a month, across Instagram, Facebook and WhatsApp, unlocking heavier use of its Muse image and video models beyond the free tier. TechCrunch · Meta

WATCHJay Carney to lead a new AI policy bridge to Democratic lawmakers  #

SV Angel's Ron Conway tapped former Obama press secretary Jay Carney to run Project Blueprint, aimed at frontier governance, the China race and job displacement — launching with public backing from Altman, Amodei and Hassabis. PR Week

The long view

A year ago, each frontier lab set its own safety commitments unilaterally, with no joint standard and no shared evaluators between competitors. This week OpenAI, Anthropic and Google DeepMind confirmed they've coordinated on safety for weeks, cosigned a shared AEF-1 evaluator standard, and said they don't need an antitrust waiver to do it. If this trend continues, the test is whether shared standards produce real enforcement, or just formalize existing practice — and whether regulators treat coordination among market leaders as safety, not the collusion critics are already naming it.

Also noted

  • Apple began rolling out Siri AI in beta, on-device plus Private Cloud Compute processing, English-only, September 14. Apple #
  • A viral report alleges Anthropic's proposed evaluator METR is conflicted through effective-altruism funding ties. Protos #
  • Cornelis Networks, an Intel spinoff, raised $205M for a GPU-agnostic AI-cluster networking fabric, the Active Compute Fabric. TechCrunch #
  • OpenAI passed Anthropic in OpenRouter dollar spend for the first time in 2.5 years, per OpenRouter's own data. officechai #
  • Unsloth's dynamic GGUF quantization of DeepSeek-671B reportedly beats SOTA models on the Aider Polyglot benchmark. Unsloth #