Logic Training Generalizes Reasoning, But Style Doesn't Seem g-Loaded
teortaxesTex · x · 2026-09-22
teortaxesTex argues that logically structured training signal generalizes: models pretrained on code got better at nearly all tasks, and R1's RL on math & coding made it a far better writer than V3. But style, he concludes, "is not very g-loaded" — latest models still struggle on literary work and deep conceptual questions without ground truth, suggesting generalization stays close to verifiable reasoning objectives.
More from AGI Musings
- Hugging Face incident row: Andrew Ng calls AI fear a PR-driven setback, critics push back — aran_nayebi · 2026-09-22
- Google researcher: AI adoption stalls on certainty expectations — output error bounds instead — docmilanfar · 2026-09-22
- Jensen Huang says there's a '0% chance' AI destroys humanity by 2030 — AlexTensor · 2026-09-22
- Aviation-Style Black Boxes for AI: Building Incident Reporting Systems — ProfChesterman · 2026-09-22
- AI safety scholar: safety has moved from niche to mainstream, with hype and conflicts of interest — ProfChesterman · 2026-09-22
- "Alignment is a diff you can ship": Andrew Ng's most contrarian line on model alignment — ccerrato147 · 2026-09-22