Base log perplexity is an extremely strong predictor of Epoch and OpenAI ECI, researchers note
stochasticchasm · x · 2026-11-23
A tweet thread on whether we're past the era where base-model perplexity determines everything: one poster quipped it feels nostalgically like 2023 again, since we're supposedly beyond that. @eliebakouch countered with a chart showing base log perplexity remains an extremely strong predictor of both Epoch and OpenAI ECI benchmarks. The original poster conceded the correlation looks pretty strong.
More from Models
- Gemini 4 xhigh costs ~2.8x more but passes the same 12 checks as high in coding test — mhdfaran · 2026-11-23
- Dylan Patel: efficient pretraining makes post-training misalignment exponentially easier — dylan522p · 2026-11-23
- Leaked OpenAI loss curve suggests GPT-6.2 Astra solved a Millennium Prize problem, internal model already lower — kimmonismus · 2026-11-23
- Dev baffled by Google's new Neon, Argon, Carbon, Hydrogen model codenames — eigenhector · 2026-11-23
- Gemini Pro usage screenshot shows Claude/GPT weekly quota fully exhausted at 0% — MediocreInside8628 · 2026-11-23
- AI agents play a 3D murder mystery: GPT-6.2 solves it in 3m14s — venturetwins · 2026-11-23