Increasing AI automation from 20% to 100% of economic tasks requires several orders of magnitude more effective compute.
3 events · 1 assessment · 1 decision
Structured and assessed
First pass. Added one required subclaim, the quantity for the effective FLOP gap between 20% and 100% automation, which the Matcher confirmed is new. No named arguments, because the claim reads directly off that quantity. Assessed it as inconclusive at credence 0.5: Davidson's central estimate of 4 OOM rests on indirect evidence and has a wide range. The budget was tight, so no instances were recorded. Davidson 2023 and GATE are the sources to record on a later pass.
Assessed Inconclusive
verdict confidence 0.65 · credence 0.50
The claim restates a modelling parameter rather than an observed regularity. In Tom Davidson's 2023 compute-centric takeoff model, the "effective FLOP gap" between AI that can perform 20% of economic tasks and AI that can perform all of them has a central estimate of about four orders of magnitude, with a wide range (roughly one to eight or more). Epoch's GATE model adopts the same parameter and treats it as highly uncertain. The claim is true if the effective compute gap is at least about three orders of magnitude. The case for a large gap is that the last tasks to fall are likely the hardest: long-horizon, judgment-heavy, physically embodied or poorly specified work. Davidson combined several indirect lines of evidence, including how capabilities scale across domains and how behaviour changes with animal brain size. None of these measures the gap directly. Against a large gap: capability gains on agentic and long-horizon tasks have arrived faster than many expected. Some of the remaining tasks may also be limited by robotics hardware, data, or regulation more than by training compute, and that makes "required compute" a poor measure of what stands in the way. The 20% starting point is itself loosely defined. The parameter is a reasonable central guess. But the evidence behind it is indirect, and it leaves substantial probability on gaps of only one or two orders of magnitude. The claim is best read as an open forecast. Better evidence would come from measuring how automatable task shares grow with each step up in compute, using successive frontier models.
Claim entered the graph