Karpathy: you know RL is working when models stop thinking in English
joeddav · x · 2026-10-21
Karpathy quipped that you can tell RL training is done properly 'when the models cease to speak English in their chain of thought', and followed up confirming 'the RL is working now all right.' The remark highlights how post-RL models increasingly reason in mixed or non-English chains of thought — a humorous but telling signal of effective RL training.
More from Models
- Don't mention recent AI math results to Opus 5.5 if you want it to focus — repligate · 2026-10-21
- User finds codex now beats ChatGPT for everything but quick phone searches — verrsane · 2026-10-21
- Nace.AI Launches Drex Max 2.0, a 27B Local Decision Model for Computer Use — DotaMate · 2026-10-21
- Speculation: Codex/Work also cap at 12 hours — OpenAI quietly eats cache-miss costs — dotey · 2026-10-21
- Using GPT-6 Pro for strategic business discovery: a 51-minute multi-subagent run and reusable prompt — cedric_chee · 2026-10-21
- Gemini 4 Argon processed ~25T tokens in first 72 hours, 3x Gemini 3 — clmt · 2026-10-21