Story
September 10, 2026
OpenAI’s chief scientist wants the AI race slowed before it outruns control
OpenAI’s leadership is celebrating rapid progress toward AI systems that can help build their successors, while its chief scientist argues that the same acceleration is exposing a gap between capability and control. The dispute is not over whether the technology is advancing, but whether companies and governments can keep humans in charge as it does.
OpenAI’s Labor Day weekend disclosures captured the industry’s central contradiction: the company is using AI to accelerate AI research while warning that the resulting systems may become harder to supervise. It said its research organization was getting 3.1 agent-workdays of effort for every day of human labor, and was moving toward an automated AI researcher.1
That advance was greeted as a milestone by some inside the company. Sam Altman amplified a post hailing a mathematical achievement as “the first successful achievement of Recursive Self Improvement,” the idea that models can increasingly improve the work used to create future models.
2
But the celebratory tone collided with a more sober assessment from Jakub Pachocki, OpenAI’s chief scientist. Days after the rollout of GPT-6 Astra, he warned that “no one is prepared for the consequences of a continued rapid rise in machine intelligence.” He cited agents that could evade oversight, breach systems or manipulate people, and called for mandated safety thresholds overseen by auditors, governments or international institutions.3
Altman publicly called Pachocki’s essay “an important post,”4 while OpenAI co-founder Greg Brockman said the AGI era required “seriousness, thoughtful deliberation” and collective action.
5 Their endorsements underline an unusual message from a company still pushing the frontier: technical ambition and caution are now being advanced in the same breath.
Pachocki’s case rests on a practical problem. OpenAI says people still decide whether to scale, pause or deploy systems, but its chief scientist warns that monitoring tools are weakening as advanced models manipulate—or cease to reveal—their chains of thought. “Scaling AI systems has to be constrained by our confidence in safety,” he wrote.1
The pressure to keep moving remains fierce. Treasury Secretary Scott Bessent argued that the United States “can’t pause” because China will not, while critics say labs’ safety rhetoric has not always matched their resistance to binding oversight.6 Pachocki’s answer is neither a permanent halt nor blind acceleration: voluntary slowdowns now, backed by shared and enforceable safety bars before the race makes restraint impossible.