OpenAI pushes perplexity-based safety bars, reveals model capability chart and internal 178-BECI model
StartupYou · x · 2026-11-23
OpenAI researcher merettm published "Perplexity and Intelligence", arguing for shared safety bars based on perplexity and other measurements, with progress beyond them gated on understanding AI behavior — a new level of pretraining transparency.
An accompanying chart maps BECI capability against base-model log-perplexity from GPT-3.5 to GPT-6.2 Astra: +27 BECI from March 2023 to December 2025 on only 0.025 loss reduction; 2026 projects +20 BECI at 0.127 loss drop (1.6 BECI per 0.01 of loss); an internal model at 0.732 log-perplexity hits 178 BECI. Reasoning adds +9 BECI for both o1 and o3 over their base models.
More from Models
- OpenAI opens up pretraining progress, but critics say key compute data is missing — SonglinYang4 · 2026-11-23
- 5 free uncensored GGUF models on Hugging Face: MiniMax, Qwen, Kimi, Nemotron — Promptmethus · 2026-11-23
- OpenAI admits pretraining barely improved since 2022 — but flat curves aren't flat capabilities — petrusenko_max · 2026-11-23
- Pretraining sets the floor, RL raises the ceiling: on how the two divide the work — JacquesThibs · 2026-11-23
- Google AI Studio's periodic-table model selector reveals Gemini 4 Neon: 1M context, $3/$15 pricing — patloeber · 2026-11-23
- Domingos: Log-Perplexity Fell to 0.75 Bits While Cost Rose Orders of Magnitude — pmddomingos · 2026-11-23