Modeling AI takeoff with Fable 5.2: lab metrics run 4.7 years ahead of real value
ronbodkin · x · 2026-11-18
Ron Bodkin builds a quantitative three-stage model of AI research progress — built with Claude Fable 5.2 — to examine rapid-takeoff arguments sparked by METR's report on AI acceleration at Anthropic.
Core model: progress flows through small-scale experiments (gated by researcher-taste judgment) → best results run at scale (governed by a transfer rate) → real value (governed by metric validity). Hard-to-supervise tasks like research taste lag behind easily supervised ones like performance optimization.
Key findings (primary scenario: AI ideas/transfer judgment +3 SD/year, metric judgment +0.5 SD):
- With no pacing, progress pace hits 15x/year by Feb 2029;
- Progress on the lab's own metrics (11.4 years) outpaces real value (6.7 years) by 4.7 years, since metric validity lags;
- Comparing four pacing policies: comprehension audits delay the takeoff moment to Mar 2030, a 9-month internal-use lag to Jun 2030, capping compute at 9.6% is weaker (Dec 2029);
- A lagged lab working with older AI can actually move faster.
The piece supports a staged model of research taste: lags in hard-to-supervise steps naturally slow takeoff, offering a testable framework for interpreting METR's findings.
More from AGI Musings
- A Godelian question: do dumb rocks defending against ASI need a proof of security? — jessi_cata · 2026-11-18
- Can superintelligence get past a wall of dumb rocks? An AGI speculation thread on Turing completeness and security proofs — jessi_cata · 2026-11-18
- 'Permanently unemployable' rhetoric sparks pushback: we still have agency over AI's future — kylekabasares · 2026-11-18
- Does intelligence level n build a 'super lance' that 10n can't beat? — jessi_cata · 2026-11-18
- Judea Pearl: econometricians don't even understand potential outcomes themselves — yudapearl · 2026-11-18
- A designer pitches a mini-series on what the "End of Software" future actually looks like — vinvan · 2026-11-18