Michael Darius Eastwood · Independent AI alignment researcher
Published
Recursion in this context means letting a model work over its own outputs. Parallel recursion runs several branches at once and combines them at the end; sequential recursion runs a chain, with each step depending on the last. The claim about the parallel case is deflationary. When you fit the exponent alpha-parallel from the data (log capability against log rounds), it sits close to zero. Adding more parallel branches at fixed compute does not buy meaningful scaling. The finding is the programme's strongest replicated quantitative result. It is not the same claim as "parallel is useless" (parallel has uses that the exponent does not capture), and it is not the same claim as "sequential is better" (that is Claim 10, and its priority is credited to Sharma & Chopra).
The measurement is described in Papers I, II and IX. The version-three harness runs a set of medium-tier arithmetic evaluators across five or more models from three families, comparing eight sequential rounds against four parallel branches of two rounds followed by a merge and two further rounds. Compute is matched so that the difference is the recursion form and not the budget. The programme's own runs find alpha-parallel close to zero across every model tested. Brown et al. (2024) reported the diminishing-returns direction under the "Large Language Monkeys" framing on their own harness, which is why the programme cites their paper as prior art on the direction. Sharma & Chopra later quantified the sequential-versus-parallel gap independently and are cited on the sequential half of the picture.
The claim's status is Needs replication, not Confirmed. The programme's harness has reproduced alpha-parallel approximately zero in the scaling-challenge example and in every model configuration tested internally, and the finding is consistent with Brown et al., but a full six-model rerun by an independent lab is pending. The claim survives with different task domains and different merge strategies; the exact number for alpha-parallel is less important than the sign, which is roughly zero rather than positive. If parallel recursion turns out to scale under a differently constructed harness, the claim is refuted; that possibility is the reason for the replication contract.
The falsification contract asks a second lab to build the serial-versus-parallel comparison at matched compute, run it across at least five models from three families, include at least two task domains and two merge strategies, and compute the paired difference alpha-serial minus alpha-parallel with bootstrap confidence intervals. Confirmation requires the difference to be greater than zero in at least four out of five models at p<0.05. If the difference is approximately zero, the claim is refuted. If the difference reverses (parallel actually higher than serial), the claim is stronger than refuted and the dashboard will state the direction reversed. The number that decides is the sign of alpha-parallel, not any comparison against a headline benchmark.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.