OpenAI's chief scientist wants a training speed limit before AI accelerates itself in 2027

Jakub Pachocki asked rival labs to agree on perplexity thresholds they will not cross until model monitoring improves, warning that AI-driven breakthroughs in training itself could arrive as early as 2027.

By · Published

Several people in white lab coats watch glowing particles move through clear curved tubes beside a transparent panel.

Jakub Pachocki asked rival labs to agree on perplexity thresholds they will not cross until model monitoring improves, warning that AI-driven breakthroughs in training itself could arrive as early as 2027.