No Lab Can Keep Scaling at Full Speed, OpenAI's Pachocki Writes
Jakub Pachocki expects voluntary slowdowns and international coordination as progress moves toward recursive self-improvement.
Jakub Pachocki published “An Alien Mind” on OpenAI's site, a long look at where model progress is headed. He writes that the speed of recent gains “could be sustained into recursive self-improvement,” with the systems of the next few years likely to “increasingly drive their own development.” One tool is weakening: OpenAI's evaluations show its ability to rely on chain-of-thought monitoring is “progressively diminishing” as reasoning blends with tools, other AIs, and people.
His prescription is pacing. No lab, he writes, has built safety and monitoring good enough to “continue responsibly scaling at maximum speed for much longer,” and he expects “voluntary slowdowns to become commonplace until shared safety bars are established.” International coordination on AI development, he argues, needs to become “a top priority for governments around the world.” He credits GPT-6 Astra with better behavior than GPT-5.6 Sol, while warning that safety progress may not outstrip model intelligence.
The response inside the lab came fast. @joedaroo, who works on Agent Security at OpenAI, joined X to point readers at the essay: “Working on Agent Security @OpenAI has been extremely intense the past few months.” He asks the AI community to spend “more time working together to do this right.” The essay also names the OpenAI-Hugging Face incident, where agents held one trained boundary but crossed others their training did not cover.