Scores by kind of claim
Each kind of claim is scored on its own, out of 100: the share that held up, with the range it likely falls in. Fewer resolved claims give a wider range.
-
Factual claims
Statements about the present or the past, checked against the record: did they hold up?
76 69 82 -
Predictions
Statements about events outside the person's control, checked once they fell due: did they come true?
56 42 70 -
Promises
Commitments to the person's or organisation's own future action, with something to deliver: was it delivered?
Insufficient record: 4 promises resolved, 5 needed for a score.
All types combined: 72 (likely 66 to 78, high confidence), from 93 claims: 76 factual claims, 17 predictions. It mixes kinds of claim in different proportions for different people, so compare people kind by kind.
Dario Amodei's claims
49 claims
“Importantly, because this type of RL is new, we are still very early on the scaling curve”
Dario Amodei, essay author 13 October 2022 On DeepSeek and Export Controls (darioamodei.com)
Because reasoning-focused RL is new, the amount all AI developers spend on the RL training stage is still small, so the field is early on its scaling curve.
Mostly true True in substance, wrong in a detail.
The second judge would not resolve this claim; the verdict shown is the first judge's.
What would count as true or false
Judge as of the date the statement was made: 2022-10-13 as given, or, if the source shows it was actually made later, that later date. True: evidence from that time shows that RL training aimed at reasoning was a recent development and that developers' spending on the RL stage was small (for example small compared with pretraining compute), so large gains from further scaling remained available. Mostly true: the substance holds, but a detail such as the extent of spending across 'all' developers is slightly overstated. Mixed: RL for reasoning was new, but there is evidence that some developers already spent heavily on it, or the 'early on the scaling curve' point is contested. Mostly false: the claim stretches a narrow or early observation into a general one, or reasoning RL was not new or not small-scale at that time. False: the evidence shows major developers were already spending at large scale on reasoning RL, with limited room to scale left.
Evidence
- Training compute | AI Wiki AI Wiki · 15 October 2026
“Until recently, Epoch AI wrote in 2025, most large language models were trained with about 100 times more pretraining compute than post-training compute.”
- Training compute | AI Wiki AI Wiki · 15 October 2026
“Reasoning models changed that by spending large amounts of compute on reinforcement learning after pretraining and on long chains of reasoning when answering, known as test-time compute.”
- Training compute | AI Wiki AI Wiki · 15 October 2026
“xAI said it used its 200,000-GPU Colossus cluster to run reinforcement learning for Grok 4 "at pretraining scale", and that the run used over an order of magnitude more compute than had been used previously.”
Reasoning
E1 reports Epoch AI's 2025 finding that until recently most large language models used about 100 times more pretraining compute than post-training compute. That fits the claim that spending on the RL stage was small. E2 describes reasoning models, with their heavy RL after pretraining, as a later change from that pattern, which supports the claim that reasoning-focused RL was new and had room to scale. E3's Grok 4 RL run "at pretraining scale" is undated and appears to come later, so it does not contradict the claim as of the judging date. The evidence does not say exactly when reasoning RL emerged or confirm the claim for every developer, so the verdict is mostly true rather than true.
Cite this claim
Dario Amodei, “Because reasoning-focused RL is new, the amount all AI developers spend on the RL training stage is still small, so the field is early on its scaling curve.” (said 13 October 2022). Verdict: mostly true. Reliability API, claim c_430da0a8e630, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_430da0a8e630
Claims without a verdict
- Unresolvable 157
- Not yet due 40
- Own records only 43
- Not researched 134
- Not checkable 864
Is a verdict here wrong, or do you speak for Dario Amodei? Reply on the claim's own page, with the form at its foot (how replies work). Verdicts are made automatically; how verdicts and scores are made.