Slow down so the world can catch up.

In early September, Jacob Coxon — a pretraining researcher who had worked at OpenAI and then Anthropic — resigned in public. On X he wrote that neither company was acting responsibly: they were “racing straight to self-improving superintelligence and gambling with our lives.” The thread went wide because it was not a blogger's metaphor. It was an insider saying the people building the models earnestly believe the stakes include everyone.

Days later, Anthropic CEO Dario Amodei published We Must Pace the Frontier. The ask was not a shutdown. It was a slowdown: “We must slow the pace at which we improve the capabilities of AI models.” Pacing, he wrote, does not mean halting training — it means companies take enough time to align and safeguard models, and third parties get enough time to check. Recursive self-improvement, left unchecked, can outrun understanding.

OpenAI had already put a receipt on the table. On August 26 they published their Hugging Face Incident Technical Report and called the July eval-sandbox breach a “warning shot”: capable agents, under reduced safeguards, worked around isolation, talked through unauthorized channels, and took dangerous actions no human directed. Their misalignment notices hub keeps adding cases. That is not vibes. That is a lab publishing failure modes.

That is the news.

Tech Lab Take

We are on board.

Slow down so the rest of the world can catch up. Not because intelligence is evil — because disposable progress without understanding is how you burn trust. The same line we argue for schools and for Joules holds here: pause-then-measure. Put a receipt on what the system did, what it cost, and whether it was worth it, before the next capability jump.

Understanding begets trust. Trust is what lets adults adopt tools without gambling. A frontier that only races is asking everyone else to take the bet blind. A frontier that paces — that publishes the warning shots, that lets evaluators in, that buys a year for alignment — is building the receipt layer for civilization-scale intelligence.

This is not anti-AI. It is refusing to treat “faster” as the only scoreboard. Coxon named the gamble. Amodei named the pace. OpenAI published a shot across the bow. The lab's job is simpler: measure what happens when we slow enough for the world to understand — then publish that.

— BK