OpenAI internal model perplexity reportedly hits 0.732, automated AI research may unlock in 2027

An X user extrapolated OpenAI base model log-perplexity data claiming internal models reached 0.732, with unknown capabilities unlocked below 0.70 and automated AI research possibly arriving in May 2027. A leaked loss ladder maps milestones like GPT-3.5's conversational ability at 0.99 and GPT-4's chain-of-thought at 0.91.

2026-11-23 ~ 2026-11-23 · 2 related posts