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
12 claims
“we were measuring these trends back when models cost $1,000 or $10,000 to train”
Dario Amodei, Anthropic CEO 10 March 2025 CEO Speaker Series with Dario Amodei of Anthropic (Council on Foreign Relations)
Amodei's group measured AI scaling trends when models cost $1,000 or $10,000 to train.
Mixed Partly true and partly false in roughly equal measure.
The second judge would not resolve this claim; the verdict shown is the first judge's.
What would count as true or false
The test is whether Amodei and his collaborators (at Baidu, OpenAI or Anthropic) measured AI scaling trends in work where individual models cost roughly $1,000 to $10,000 to train. Treat the figures as order-of-magnitude. TRUE: dated publications or contemporaneous records show his group studied scaling trends with models whose training cost was in the thousands to low tens of thousands of dollars. Examples are early scaling work such as Baidu's 2017 'Deep Learning Scaling is Predictable, Empirically' or early OpenAI scaling studies, with compute or cost estimates in that range. MOSTLY TRUE: the group did do early scaling measurements, but the models cost somewhat more or less than stated (for example around $100 or about $50,000), or the timing detail is slightly off. MIXED: scaling work in that era is documented, but whether Amodei's own group did it, or the cost scale, is only partly supported. MOSTLY FALSE: Amodei was involved in scaling work, but the earliest measurements involved models far more expensive (hundreds of thousands of dollars or more), so the claim overstates how early or cheap the work was. FALSE: the evidence shows his group did no such scaling measurements at that cost level.
Evidence
- The Making Of Dario Amodei Big Technology · 29 July 2025
“In their experiments, Amodei and his colleagues found that AI’s performance improved meaningfully as they added more of these ingredients. The team released a paper on speech recognition that showed that model size directly correlated with performance.”
- The Making Of Dario Amodei Big Technology · 29 July 2025
“But the jump from GPT-2 to GPT-3 would be massive, a 100x project costing tens of millions of dollars.”
- When Will We Spend Enough to Train Transformative AI EA Forum · 28 March 2023
“estimates put the cost of training GPT-2 in 2019 around $20k-$100k. Training GPT-3 in 2020 cost $2m-$10m.”
- The Making Of Dario Amodei Big Technology · 29 July 2025
“We’ve done scaling laws. We could see the path to the models getting good.”
- The Making Of Dario Amodei Big Technology · 29 July 2025
“In their experiments, Amodei and his colleagues found that AI’s performance improved meaningfully as they added more of these ingredients.”
- The Making Of Dario Amodei Big Technology · 29 July 2025
“The team released a paper on speech recognition that showed that model size directly correlated with performance.”
- The Making Of Dario Amodei Big Technology · 29 July 2025
“Amodei, who’d become a research director at OpenAI, got directly involved in the next iteration, GPT-2, which was effectively the same model as GPT, just bigger.”
- The Making Of Dario Amodei Big Technology · 29 July 2025
“But the jump from GPT-2 to GPT-3 would be massive, a 100x project costing tens of millions of dollars.”
- AI and compute OpenAI · 16 May 2018 Own record
“Dario Amodei, Danny Hernandez”
- AI and compute OpenAI · 16 May 2018 Own record
“2012 to 2014: Infrastructure to train on many GPUs was uncommon, so most results used 1-8 GPUs rated at 1-2 TFLOPS for a total of 0.001-0.1 pfs-days.”
- Scaling laws for neural language models - OpenAI OpenAI · 23 January 2020 Own record
“Jared Kaplan, Sam McCandlish, Tom Henighan, Tom Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, Dario Amodei”
- Deep Learning Scaling is Predictable, Empirically arXiv
“In speech recognition, Amodei et al. (2016) show word error rate improvement for a Deep Speech 2 model on varying sizes of training data.”
Reasoning
Independent sources show that Amodei and colleagues ran early scaling experiments. E1, E5 and E6 describe work by Amodei and colleagues, including a speech-recognition paper linking model size to performance, and E12 cites Amodei et al. (2016) on how Deep Speech 2 error rates changed with the amount of training data. None of the evidence gives a dollar training cost for that early speech-recognition scaling work. E10 gives compute only in petaflop/s-days, and the paper credits in E11 carry no cost figure. The nearest cost figure is a third-party estimate of $20k–$100k for GPT-2 (E3), a model Amodei worked on (E7). That figure is near the claimed range but above it, so the cost part of the claim is only partly supported.
Cite this claim
Dario Amodei, “Amodei's group measured AI scaling trends when models cost $1,000 or $10,000 to train.” (said 10 March 2025). Verdict: mixed. Reliability API, claim c_96cb67eda7c1, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_96cb67eda7c1
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.