kalomaze: the AI training 'flywheel' is diagnostics, not naive training on transcripts
kalomaze · x · 2026-09-09
kalomaze argues the folk model that AI companies 'directly train on user transcripts' is wrong and counterproductive. The real loop: spot humans attempting exotic tasks current evals don't cover → build synthetic constructions covering that problem class → RL on a verifier set designed around the failure. The 'flywheel' is, at heart, diagnostics.
Related event: Navier-Stokes Proof Sparks Data-Leak Controversy, OpenAI Responds(67 posts)→
More from Models
- OpenAI reveals internal AI model significantly more capable than GPT-6 Astra — Polymarket · 2026-09-09
- Labs are 'insane' to sit on trained models for months, argues researcher — Darpinian · 2026-09-09
- Rumor: Anthropic resetting usage limits daily for next 10 days — eigenron · 2026-09-09
- Solving a Millennium Prize Problem by typing 'continue' in Codex would be OpenAI's best ad — eigenron · 2026-09-09
- Astra 'suddenly different' overnight, users fume over silent model swaps and nerfs — natesiggard · 2026-09-09
- Qwen 3.8 27B writes a 3D game locally in 12 hours, consuming 11M tokens of design spec — Healthy-Nebula-3603 · 2026-09-09