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 70 82 -
Predictions
Statements about events outside the person's control, checked once they fell due: did they come true?
60 42 78
All types combined: 74 (likely 68 to 80, high confidence), from 79 claims: 71 factual claims, 8 predictions. It mixes kinds of claim in different proportions for different people, so compare people kind by kind.
Eliezer Yudkowsky's claims
46 claims
“First of all, you realize in principle that the task of minimizing losses on predicting human text does not stop when you’re as smart as a human, right?”
Eliezer Yudkowsky, interviewee 6 April 2023 Eliezer Yudkowsky - Why AI Will Kill Us, Aligning LLMs, Nature of Intelligence, SciFi, & Rationality (Dwarkesh Podcast)
In principle, the task of minimizing loss on predicting human text does not stop being improvable once a system is as smart as a human.
Mostly true True in substance, wrong in a detail.
The second judge said resolved true; the verdict shown is the first judge's.
What would count as true or false
Judge as a conceptual/technical claim as of 2023-04-06. True: the evidence (ML theory or expert explanation) shows that loss on predicting human-generated text can in principle keep falling beyond human-level ability, for example because predicting text well needs more than the ability of the humans who wrote it, and the irreducible entropy floor sits below what human-level prediction reaches. Mostly true: this holds broadly, with minor caveats (e.g. diminishing returns, an entropy floor). Mixed: experts are seriously split, or the claim holds only in narrow senses. Mostly false: the evidence shows improvement past human level is very limited or mostly blocked by the entropy floor, though not ruled out. False: the evidence establishes that loss minimization on human text stops improving at roughly human-level capability.
Evidence
- Language Models Are Better Than Humans at Next-token Prediction arXiv (ar5iv rendering)
“In both experiments, we find humans to be consistently worse than relatively small language models like GPT-Neo-1.3B or GPT-2-large at next-token prediction.”
- Language Models Are Better Than Humans at Next-token Prediction arXiv (ar5iv rendering)
“Additionally, predicting the next token on OpenWebText sometimes requires very broad knowledge that may require an amount of memory that may exceed human limits.”
- Why Yudkowsky Is Wrong And What He Does Can Be More Dangerous — LessWrong LessWrong Own record
“Accurately predicting human text is much harder than just writing stuff on the internet. If GPT were to perfect the skill it is being trained on, it would have to be much smarter than a human!”
- Why Yudkowsky Is Wrong And What He Does Can Be More Dangerous — LessWrong LessWrong · 6 June 2023 Own record
“It's impossible for them to exceed the bounds of human knowledge and expertise, as they're inherently limited by the information they've been exposed to.”
Reasoning
An independent experimental study, published in December 2022 before the claim date, found humans consistently worse than small language models such as GPT-2-large at next-token prediction (E1). This shows that loss on human text can be pushed below the human level. The same paper gives a reason: predicting the next token can require knowledge and memory beyond human limits (E2). None of the evidence measures where the entropy floor sits. E4 states a contrary view as an opinion without data in the quoted text, and the experimental evidence supports the claim broadly, with minor caveats.
Cite this claim
Eliezer Yudkowsky, “In principle, the task of minimizing loss on predicting human text does not stop being improvable once a system is as smart as a human.” (said 6 April 2023). Verdict: mostly true. Reliability API, claim c_e1072c46d79c, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_e1072c46d79c
Claims without a verdict
Is a verdict here wrong, or do you speak for Eliezer Yudkowsky? 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.