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?
81 76 87 -
Predictions
Statements about events outside the person's control, checked once they fell due: did they come true?
Insufficient record: 2 predictions resolved, 5 needed for a score.
-
Promises
Commitments to the person's or organisation's own future action, with something to deliver: was it delivered?
65 45 83
All types combined: 80 (likely 75 to 85, high confidence), from 84 claims: 78 factual claims, 6 promises. It mixes kinds of claim in different proportions for different people, so compare people kind by kind.
Demis Hassabis's claims
37 claims
“the new paradigm of the last year that where they get smarter, the longer amount of inference time you give them at test time.”
Demis Hassabis, CEO of Google DeepMind, interviewee 23 July 2025 Transcript for Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games (#475) (Lex Fridman)
Over the last year a new paradigm of thinking AI systems emerged, which get smarter the more inference time they are given at test time.
Resolved true The record shows it happened or was so.
What would count as true or false
True: between about mid-2024 and July 2025, AI models that use extra inference-time or test-time compute ("thinking" or reasoning models such as OpenAI o1 from September 2024, o3, DeepSeek-R1 and Gemini 2.5 Flash/Pro thinking) emerged and became widely recognised as a new paradigm, and published evidence shows that their performance improves as they are given more inference time or compute. Mostly true: the paradigm and the scaling with test-time compute are documented, but the timing is a little off (e.g. its emergence dates slightly before mid-2024). Mixed: test-time reasoning models exist, but improvement with more inference time is shown only weakly or inconsistently, or the paradigm clearly began well before the last year. Mostly false: such systems existed only as isolated research results without broad adoption in that period. False: no such models or scaling behaviour had appeared by July 2025.
Evidence
- Start building with Gemini 2.5 Flash Google Developers Blog · 17 April 2025
“A higher budget allows the model to reason further to improve quality.”
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning arXiv (ar5iv rendering), DeepSeek-AI authors · 1 August 2024
“OpenAI’s o1 (OpenAI 2024b) series models were the first to introduce inference-time scaling by increasing the length of the Chain-of-Thought reasoning process.”
- OpenAI's o3 suggests AI models are scaling in new ways — but so are the costs | TechCrunch TechCrunch · 23 December 2024
“We don’t know all the details of how o3 was made, but these benchmarks are early signs that test-time scaling may work to improve the performance of AI models.”
- Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities awesome.papernotes.org (awesome-papers project) · 17 September 2026
“Figure 4 shows aggregate performance increasing with thinking budget on AIME 2025, LiveCodeBench, and GPQA Diamond.”
Reasoning
The DeepSeek-R1 paper (E2) credits OpenAI's o1 series, cited as 2024, with first introducing inference-time scaling through longer chain-of-thought reasoning. In December 2024, TechCrunch (E3) described o3 as a sign that AI models were "scaling in new ways" through test-time compute. Google's Gemini 2.5 Flash announcement (E1) says a higher thinking budget lets the model reason further and improve quality, and the Gemini 2.5 report summary (E4) shows performance rising with thinking budget on AIME 2025, LiveCodeBench and GPQA Diamond. Taken together, these show reasoning models with test-time scaling emerging and spreading within roughly the year before July 2025, and published results show their performance improving as they get more inference compute.
Cite this claim
Demis Hassabis, “Over the last year a new paradigm of thinking AI systems emerged, which get smarter the more inference time they are given at test time.” (said 23 July 2025). Verdict: resolved true. Reliability API, claim c_d0db36b1b605, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_d0db36b1b605
Claims without a verdict
Is a verdict here wrong, or do you speak for Demis Hassabis? 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.