新智元 (微信公众号)T2·72 pts
- Altman separates two layers: AI already speeds up OpenAI's R&D, but unsupervised autonomous iteration remains far off
- OpenAI brings in METR and other external evaluators on-site, conceding internal self-review is insufficient for safety claims
- Recursive self-improvement is framed as gradual rather than a sudden threshold, mirroring AGI narratives