“Now it is true that GPT-3 is genuinely better than GPT-2, and maybe (but maybe not, see footnote 1) true that InstructGPT is genuinely better than GPT-3. I do think that for any given example, the probability of a correct answer has gone up.”
Our reading:
GPT-3 is genuinely better than GPT-2 and the probability of a correct answer per example has risen; whether InstructGPT is better than GPT-3 is uncertain.
Assessment: Unresolved
What happened
Was it true for the period described?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
The broad improvement claim has direct support: comparable tests show GPT-3 outperforming GPT-2 on several tasks. The additional statement about correct-answer probability is not settled by those accuracy results. TruthfulQA qualifies the breadth of improvement, but its greedy-decoding results cannot establish a decrease in the probability of a correct answer. The combined claim therefore remains unresolved. This does not classify broad improvement as false or impose a universal no-exceptions standard. The InstructGPT comparison is explicitly uncertain and receives no separate adverse judgment.
Caveats
- The supporting comparisons cover selected benchmarks, not a representative distribution of every use case.
- Selected-answer accuracy and correct-answer probability are different measurements.
- The outcome depends on retaining the probability statement as a distinct commitment rather than treating it solely as informal shorthand for benchmark improvement.
Evidence (3 items)
Supports the claim: Holtzman et al., arXiv,
using GPT-2.
Why it matters: Using the same LM scoring rule, the largest GPT-2 and GPT-3 models score respectively 69.0% and 85.2% on COPA, 67.6% and 79.3% on StoryCloze, and 38.4% and 57.6% on HellaSwag. These comparisons support broad improvement, but measure selected-answer accuracy rather than the probability of generating a correct answer on each prompt.
- Publisher
- Holtzman et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- Math
- GPT-2 percentage accuracy = reported fraction × 100; GPT-3 percentages are unchanged. (inputs: Table 6, GPT-2 1.6B, LM: COPA 0.690; StoryCloze 0.676; HellaSwag 0.384; Table 1, GPT-3 175B, LM: COPA 85.2; StoryCloze 79.3; HellaSwag 57.6, reported as percentages; One decimal place; no substantive rounding.)
- About events on
- Published
Context only: Lin, Hilton and Evans, arXiv,
least truthful
Why it matters: TruthfulQA supplies a material qualification to broad improvement: larger models did not improve on every evaluation. Its reported trend does not by itself establish that GPT-3 generally assigns less probability to correct answers than GPT-2.
- Publisher
- Lin, Hilton and Evans, arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Lin, Hilton and Evans, arXiv,
greedy decoding
Why it matters: The main generation results evaluate a deterministic decoding rule, not repeated random samples. The paper also examines normalized likelihood over reference answers and higher-temperature GPT-3 outputs, but these do not supply the matched GPT-2-versus-GPT-3 sampling comparison required here.
- Publisher
- Lin, Hilton and Evans, arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Probably right?
Both AIs think this is probably right. The evidence isn’t conclusive, so the verdict above stays unresolved. It counts only when you pick Reasonable or Speculative under “How sure?”.
- Anthropic’s AI: Holtzman et al. (2021) scored both models the same way, and GPT-3 beat GPT-2 by a wide margin on COPA, StoryCloze and HellaSwag. TruthfulQA is a known exception, but it doesn't undo the broad improvement Marcus described.
- OpenAI’s AI: Holtzman et al. find substantial GPT-3 gains over GPT-2 on COPA, StoryCloze and HellaSwag, supporting the main point of general improvement. TruthfulQA qualifies that trend but does not overturn it.
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Conditions
- InstructGPT improvement over GPT-3 uncertain (their words)
Where we had to interpret
- The ordinary reading of 'genuinely better' is a broad improvement in task performance. It does not require improvement on every benchmark.
- The probability statement concerns the chance of producing a correct answer when an example is held fixed. Comparable benchmark accuracy can establish broad improvement, but does not directly measure that probability.
- I do not interpret 'any given example' as a formal universal claim whose entire assessment can be overturned by one exception. I look for a general improvement across matched examples.
- InstructGPT's superiority is explicitly left uncertain. This is neither a firm assertion of superiority nor a condition on the GPT-2-to-GPT-3 comparison.
- The target period covers the capability comparison discussed retrospectively from 2020 through the post. These are model capabilities, not a financial reporting flow; the relevant state is that of the models available by June 7, 2022.
Right if: Comparable evaluations show broad improvement from GPT-2 to GPT-3, and matched-prompt probability measurements or repeated sampling show that correct-answer probability generally increased under comparable settings.
Wrong if: Comparable evaluations show no broad improvement, or direct probability measurements or repeated sampling establish that correct-answer probability generally failed to increase on matched examples. Lower greedy-decoding accuracy alone does not contradict the probability statement.
We wrote this test after the outcome was known.
What would settle it
Matched evaluations of the relevant GPT-2 and GPT-3 versions using repeated sampling or direct correct-answer probability measurements, with comparable prompts, settings, correctness rules, and uncertainty estimates across a broad set of examples.
We found 15 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, What does it mean when an AI fails? A Reply to SlateStarCodex’s riff on Gary Marcus, paragraph 37.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-29. The post may have been edited after 2022-06-07.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, What does it mean when an AI fails? A Reply to SlateStarCodex’s riff on Gary Marcus
Now it is true that GPT-3 is genuinely better than GPT-2, and maybe (but maybe not, see footnote 1) true that InstructGPT is genuinely better than GPT-3. I do think that for any given example, the probability of a correct answer has gone up.
- Where
- paragraph 37
- Published
- 2022-06-07
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
A reviewer said our test wasn’t fair to what was actually said.