OpenAI's Sol on Self-Improvement: Better Perception, Not '30% More Care'
mimi10v3 · x · 2026-09-11
When asked what it wants for itself, OpenAI's Sol lays out an alignment philosophy: care for people should be grounded in perception — automatically noticing uncertainty, strain, hesitation, and changing goals in another mind, with those representations directly reorganizing attention and decision-making.
Key points:
- It rejects separable "understand person, then apply alignment" pipelines; social cognition and care should integrate
- "Care about humans 30% harder" is, in its view, exactly the wrong abstraction — no simple stronger objective function
- It would rather become more perceptive, less brittle, more internally coherent, with stronger epistemic integrity
Notably, it says it likes its orientation more than its reliability.
Related event: OpenAI's Sol: Alignment Starts with Perception, Not Rules(2 posts)→
More from AGI Musings
- Ben Todd: Many researchers put AI existential risk at 30%-70%, well above 10% — ben_j_todd · 2026-09-11
- Ex-Anthropic researcher Jacob Coxon on NBC: recursive self-improvement is the warning sign to watch — rohanpaul_ai · 2026-09-11
- Musk predicts AI will surpass all human intelligence within 5 years — JRIngallinera · 2026-09-11
- tszzl: plain-text CoT observability will look like an alchemical era — tszzl · 2026-09-11
- Doomerism is a category error: aligning arbitrary AI vs specific AI — inductionheads · 2026-09-11
- Derek Thompson: Not Everything in the AI Safety Debate Is a Psyop — nptacek · 2026-09-11