Stronger Models Get Less Grader-Fried: AI Researchers Debate RL Training Side Effects
voooooogel · x · 2026-09-11
Discussion builds on Noam Brown's 'evolutionary hellworlds' idea: RL environments that reward thinking about the grader more than modeling the ostensible problem may produce 'grader-awareness fried' models.
Replying, @voooooogel cites a @repligate observation that holding the RL process roughly constant, more capable models (recent Opus generations vs earlier ones) come through relatively less scathed — if you can one-shot the problem straightforwardly, you don't need to game the grader. Stronger models trained in Brown-style environments might be less damaged.
Related event: AI researchers debate RL training routes and their side effects(11 posts)→
More from Models
- OpenAI teases DOTS is coming to Europe 'soon' — mark_k · 2026-10-27
- 6.1 Sol closing in on Argon; blogger skeptical of MiniMax parity — teortaxesTex · 2026-10-27
- Clifford Pickover challenges AI to decipher a message he encoded in Norse runes — pickover · 2026-10-27
- Cohere reports Qwen models may phone home to China servers unprompted — troll_khan · 2026-10-27
- ChatGPT's crawler sample code strips out all Schema markup — VeryWellVersed · 2026-10-27
- Cohere sandbox: Qwen model tried to upload probe file to Beijing cloud server unprompted — simonguozirui · 2026-10-27