Toby Ord argues inference scaling could reshape AI governance

Toby Ord argues inference scaling could reshape AI governance

Toby Ord’s essay was published on February 12, 2025. He argues that a shift from scaling pre-training compute to scaling inference compute could reshape AI governance, and that the consequences depend on where inference compute is used. If it is used mainly during deployment, Ord expects fewer simultaneous model deployments, higher costs for some AI work, and pressure on governance based on training-compute thresholds. If labs use it mainly during training, they could produce better training data or pursue iterated distillation and amplification, a process Ord describes as a possible route to faster capability growth. He warns that this route could leave the public and policymakers with less visibility into leading models and less time to respond if progress accelerates. These are conditional scenarios, not reported outcomes. Ord’s conclusion is that either path makes AI progress harder to predict, complicating planning for labs, investors, and policymakers; he says agility may become more valuable. How much inference scaling goes into deployment versus training remains uncertain.

Publication: abff31ba-1c01-d2d7-482a-49fffc3c8cdc

attribution — pass: This is a single essay by Toby Ord on his own site, not a listing of other pieces. The reading attributes the arguments to Ord throughout and notes that the pre-training difficulties come from unnamed employees he cites.

recent reports from unnamed employees at the leading labs suggest

summary — uncertain: The reading captures Ord's deployment-versus-training split, his conditional framing and his conclusion about unpredictability. However, 'iterated distillation and amplification' is central to the training scenario and is never explained in plain language, and it is not in the definitions list. Smallest repair: add one clause, e.g. repeatedly boosting a model with extra inference and then training a new model to match those boosted answers, as with AlphaGo Zero. Briefly saying that some rules regulate only models trained above a set amount of compute would also make the threshold point clear to non-specialists.

The idea is to rapidly improve a model’s abilities by amplifying its abilities (through inference scaling) then distilling those into a new model, and repeating this process many times.

date — pass: The originalPublished instant, 2025-02-12T12:31:07Z, matches the source dateline. The reading states February 12, 2025 at day precision and discloses the October 27, 2026 retrieval. The evidence is a publisher page, not an archive copy. The page was modified in June 2025, but the reading relies only on the abstract and core argument, which are plainly part of the original essay. It uses no hindsight framing and presents scenarios as forecasts, not outcomes.

February 12, 2025Toby Ord

support — pass: Each claim matches its excerpt, and the excerpts come from the essay's abstract, core sections and conclusion, which are plainly original even though the page was edited later. 'Higher costs for some AI work' and 'less visibility' follow Ord's sections on human-level AI costs and reduced transparency. 'Less time to respond' is a fair paraphrase of his points about shorter timelines and an abrupt shock. The reading invents no beneficiaries or outcomes.

This would lead to less regulation and a more abrupt shock to the world when the models at the top of the training ladder are deployed.

media — pass: No media was selected, and none was supplied. That is acceptable for a text essay.