METR presents a deliberately stripped-down forecasting model for AI R&D automation, reducing the parameter count from their earlier AI Futures Model’s 33 down to just 8. The model combines a Cobb-Douglas production function for research output, a Jones-style law for software efficiency growth, and a sigmoidal automation curve over log-scale effective compute. Fed with current estimates — algorithmic efficiency growing at roughly 5x per year, 25-50% of AI R&D tasks already automated as of January 2026, and compute scaling at 2.6x annually through 2029 before tapering off — the model’s median trajectory lands on 99% automation of AI R&D by mid-2032, with research output scaling 300x to 3,000x by 2035 and software efficiency gains spanning three to eight orders of magnitude.

Key Insights

The most valuable contribution here is not the headline date but the transparency about what drives it. The sensitivity analysis reveals that two parameters dominate uncertainty: the automation velocity (how quickly the fraction of automatable tasks grows with each increment of effective compute) and the diminishing-returns exponent on software improvement. Pessimistic draws on both can push 99% automation past 2036, while optimistic draws pull it closer to 2030. The authors are also careful to note hard structural limitations: the model covers the path to automated coding but explicitly does not model what happens after — the “takeoff” dynamics where superhuman research taste could compound progress further.

Our Take

The candor about the model’s construction is refreshing — the authors note it was built in roughly 15 hours of work, and that pre-2026 backtesting is essentially infeasible because meaningful AI R&D automation barely existed before 2025. This honesty cuts both ways: it makes the work easy to trust as a good-faith estimate and easy to question as a provisional sketch. The choice to assume zero task substitution (strict Amdahl’s Law) is conservative, since in practice partially-automated workflows often restructure tasks rather than simply speeding up the automatable fraction. Still, as a legible, reproducible baseline that anyone can stress-test by adjusting eight numbers, this model fills a genuine gap. The bottom line — that nearly any reasonable parameterization yields superhuman AI researchers before 2036 absent a hard technological wall or deliberate slowdown — is a claim worth taking seriously, even if the precise year remains noisy.