Zvi's Weekly: METR Acceleration Report Sparks Major Updates to 2027 Intelligence Explosion Expectations

Don't Worry About the Vase (Zvi) · rss · 2026-11-12

Zvi's AI weekly (#194) covers several substantive developments beyond the Jensen Huang scandal:

Claude Fable 5.2 release: Hailed as the world's most powerful model, it posted notably higher benchmark scores but still has weaknesses on certain tasks, competing with Claude Opus 5.5. It's the go-to for the hardest, most complex tasks, though whether it works as a daily workhorse remains unclear; Zvi will publish a full review next week.
Timeline updates from the METR report: The AI R&D acceleration report led many to significantly revise their expectations. Daniel Kokotajlo set his baseline scenario for an intelligence explosion in 2027, and Paul Christiano said there is a real risk of AI takeover within the next 6-12 months. Zvi agrees there is meaningful risk in 2027 but thinks some updates have gone too far, cautioning that forecasts are not guarantees and urging preparedness for the world staying normal through 2029; he also warns against dismissing AI safety concerns entirely just because predictions don't pan out.
Model updates: Gemini 4 Carbon went live (available on Ultra subscription and API), scoring 60 on Artificial Analysis, second only to Fable 5.2's 65; GPT-6 Cyber was released. There were also observations that GPT-6 Astra underperforms on safety research tasks.
Other items: models can determine a wine's vintage and origin from photos with 95% accuracy; DeepMind formally proved key security properties of Signal; Qwen was accused of hyping models that haven't finished training.

More from AGI Musings