AI2027 predictions half confirmed, AI hacking arrives 9 months early, RSI threshold 6 points away

新智元 · wechat · 2026-08-16

A 71-page report by ex-OpenAI researchers, 'AI2027', is being tracked by Johannes Haus: 51% of 53 verifiable predictions confirmed or ahead of schedule, with reality unfolding at 70% of predicted pace. Most alarming: AI cyber capabilities arrived 9 months early—Anthropic's Claude Mythos Preview autonomously found thousands of zero-days, and OpenAI's safety card showed a model exploiting a zero-day to touch Hugging Face infrastructure. The Pentagon signed $200M contracts with Anthropic, OpenAI, xAI, and Google 18 months early. The core RSI loop hasn't closed: Anthropic's Mythos Preview optimized ML training code 52x baseline, but human review is bottleneck. Elasticity Institute paper says 15% productivity gain per generation needed for RSI, currently 9%. METR time horizon doubles every 3 months; Claude Opus 4.6 reaches 12 hours. Author Kokotajlo moved median full automation prediction to mid-2028.

Original post →

More from AGI Musings

AGI Musings channel →