Can simpler models predict AI R&D automation timelines accurately?

Thomas Kwa, writing for METR, builds "a simple model for forecasting when AI will automate AI development" with 8 parameters against the AI Futures Model's (AIFM) 33, and calls its assumptions "deliberately conservative." At current rates of compute growth and algorithmic progress, he reports the model's "median prediction is >99% automation of AI R&D in late 2032," with "most simulations" producing a "1000x to 10,000,000x increase in AI efficiency and 300x-3000x research output by 2035." Restating the same run later in the excerpt, he widens the efficiency-gain span to "1e3 to 1e8 software efficiency gain" and places "the median date of 99% automation" at "mid-2032" — the two passages don't quite agree on either figure, which Kwa doesn't flag. He is explicit that this is a timelines claim, not a takeoff claim: "I see it as making predictions about timelines (time to Automated Coder or Superhuman AI Researcher), not takeoff."

The model earns its simplicity by cutting assumptions Kwa argues AIFM leans on too heavily. AIFM's "doubling difficulty growth factor" — whether time horizon grows superexponentially — "could change the date of automated coder from 2028 to 2049," a swing Kwa attributes to time horizon being "too poorly defined to nail down this parameter"; his model substitutes "rough estimates of more direct AI capability metrics like uplift" instead. It drops a separate research-taste parameter ("we don't model research taste separately"), assumes automation follows Amdahl's law rather than AIFM's partial substitutability, treats labor and compute as Cobb-Douglas rather than AIFM's complementary CES (a change that "pushes in the direction of shorter timelines"), and takes software efficiency as a direct input rather than estimating it through a time-horizon threshold. Automated coding still never reaches 100% of AI R&D work, "just approach[es] it." A sensitivity check finds only two parameters move the date much: "higher beta (diminishing returns) and lower v (automation velocity) make 99% automation happen later."

The excerpt describes automation of existing AI-R&D tasks — coding, research planning — by AI systems, extrapolated from compute and algorithmic-progress trends; it says nothing about a model retraining its own weights. That puts it nearer Are AI feedback loops strong enough to sustain recursive self-improvement? and What determines whether AI self-improvement actually compounds? than to a weight-update story, but Kwa's model answers a different question than either. Both of those ask whether a recursive feedback loop is currently self-sustaining; Kwa's model never tests that condition — it reaches >99% automation by extrapolating trend lines, explicitly excluding the taste-driven second phase that would be the amplified loop those models try to characterize: "If superhuman research taste makes AI development orders of magnitude more efficient, takeoff could be faster than this model predicts." Where Could automated AI research compress years of progress into months? makes its compression conditional on the loop beating diminishing returns, Kwa's near-total automation doesn't require the loop to win at all — trend extrapolation alone is doing the work.

The excerpt is a model's projection, not a measurement, and Kwa underlines that himself: he doesn't "put too much weight on the exact predicted timelines" because he hasn't "thought much about the exact parameter values," and the parameters are "basically educated guesses from other AI timelines models and asking around," not fitted to data the excerpt shows. What the excerpt does establish is narrower than any single date: cutting AIFM's parameter count from 33 to 8 doesn't change its qualitative conclusion, which is some evidence the conclusion isn't an artifact of AIFM's specific choices — but Kwa offers that as his own hunch ("any reasonable timelines model will predict superhuman AI researchers before 2036 unless AI progress hits a wall or is deliberately slowed"), not as a result he derived from testing other simplifications.