Leaked OpenAI base-model perplexity chart shows roadmap up to GPT-6.2 Astra and an internal model cracking a Millennium Prize problem
iruletheworldmo · x · 2026-11-23
KOL iruletheworldmo shared an OpenAI-style "Base model log-perplexity" chart and claims Google is training a much smaller model as capable as fable, hinting at a major breakthrough.
The chart maps releases from GPT-3.5 through GPT-6.2 Astra with capability milestones: fluent conversation (GPT-3.5), thinking before answering pays off (GPT-4 era), useful coder (GPT-4.5), cybersecurity expert (GPT-5.6 Sol), and a solved Millennium Prize math problem (GPT-6.2 Astra and an unnamed "Internal model"), plus an unlabeled "?" zone implying further gains. Unverified leak, but effectively a roadmap disclosure if genuine.
More from Companies & People
- CMU tops first AI-Ready Campus Index; Prof. Conitzer quips it's like Phelps rated best for tsunamis — conitzer · 2026-11-23
- Tesla engineers praise Musk's rare hands-on technical depth as CEO — mitchdeg · 2026-11-23
- vboykis reverses course: 10x engineers exist, but their skill is relative to team and environment — vboykis · 2026-11-23
- Corey Quinn Mocks Docker for Spinning Off "the Part of the Logo That Runs Containers" — kingslyj · 2026-11-23
- Pre-LLM era's hardest ML problem was convincing colleagues to let you try it — neal_lathia · 2026-11-23
- Who Owns an AI-Generated Proof Nobody Understands? The Authorship Dilemma — gerardsans · 2026-11-23