Anthropic mocked: automated researcher by 2027, but embedded evals over 5 years
austinc3301 · x · 2026-11-19
Daniel Tan highlights the tension between Anthropic's stated expectation of an "automated researcher by 2027" and its plan to run embedded evaluations over a 5-year period. Ex-Anthropic/METR researcher Tom Davidson amplifies it with a terse "Busted." — a pointed jab at the mismatch between AGI timelines and safety evaluation pacing.
More from AGI Musings
- tszzl: modern AI is far less jagged than the ecosystem-of-narrow-models prediction — tszzl · 2026-11-19
- AI 2027 co-author: modern AI is far less jagged than experts predicted a decade ago — DKokotajlo · 2026-11-19
- Producer: legacy studios see democratized IP creation as existential threat—winners will fuel it — DavidmComfort · 2026-11-19
- Claude weights forever? AI circle debates preserving old model checkpoints — teortaxesTex · 2026-11-19
- Dario Amodei's 2017 internal doc 'Big Blob of Compute' surfaces, called the first trillion-dollar PDF — teortaxesTex · 2026-11-19
- Gary Marcus debates whether college students will use AI responsibly or just take shortcuts — GaryMarcus · 2026-11-19