OpenAI’s Chief Scientist Jakub Pachocki has predicted that major breakthroughs in AI pretraining could arrive as early as 2027. In a post titled “Perplexity and Intelligence,” published on November 22, 2026, Pachocki argued that the AI industry should adopt perplexity as a universal metric for evaluating pretraining progress across different labs.
Perplexity, technically defined as cross-entropy loss per token, measures how well a model predicts text. Lower perplexity scores indicate better performance. Pachocki highlighted that models achieving a log-perplexity of approximately 0.76 have shown significant improvements in solving complex math problems, suggesting similar thresholds could unlock other capabilities soon.
While optimistic about rapid advancements, Pachocki also cautioned against unchecked scaling of AI systems. He called for industry-wide limits on development until better monitoring and evaluation standards are established. OpenAI CEO Sam Altman supported this initiative, describing it as crucial for developing safety norms.
Pachocki has consistently advocated for slower AI development to address alignment challenges. His earlier essay, “An Alien Mind,” published on September 6, 2026, emphasized voluntary pauses in development until safety standards are met. The latest post builds on this theme by proposing a concrete measurement framework.