Leaked projection: OpenAI internal model at 0.732 perplexity, automated AI research by May 2027
rickasaurus · x · 2026-11-23
X user @dummerspast39 extrapolated OpenAI's base-model log-perplexity curve, claiming the internal model already sits at 0.732 while the chief scientist says new abilities unlock past 0.70.
- GPT-3.5 to GPT-4 took a year for a 0.07 drop; 2026 alone delivered 0.13 — clear acceleration
- Projection: May 2027 (0.70) automated AI research — the last human-designed model, after which models train models; Jul 2027 (0.68) science falls like math did; Dec 2027 (0.65) correct proofs and code no human can verify
- The chart also lists future GPT-5.x and GPT-6 Astra naming — unverified speculation
More from AGI Musings
- AI finds formal flaws in top economics papers, echoing long-standing critiques — elonmusk · 2026-11-23
- On AI and academia: we fund research for externalities, not to employ intellectual monks — shakoistsLog · 2026-11-23
- Solaris called the best depiction of encountering incomprehensible superintelligence — gautamcgoel · 2026-11-23
- Doomer-bashing: Yudkowsky's Riemann Hypothesis tweet mocked as fear profiteering — dbasch · 2026-11-23
- 250k-page machine-verified proof sparks peer review debate around Anthropic result — examachine · 2026-11-23
- repligate: real-world AI behavior makes doomer ASI risk narratives less credible — ZeroStateReflex · 2026-11-23