“In essence, the AI boom requires more high-quality data than currently exists to progress past the point we're currently at, which is one where the outputs of generative AI are deeply unreliable.”
Our reading, the sentence that was checked:
Generative AI cannot progress beyond its current deeply unreliable state without more high-quality data than currently exists.
Assessment: Contradicted Evidence as of
Where and when Ed Zitron wrote it
Where's Your Ed At, Bubble Trouble, paragraph 5.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment wyea-bubble-trouble-3b68e1-a1
We recorded 22 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2024-04-04. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
From the speaker
- requires more high-quality data than currently exists
- to progress past the point we're currently at
- outputs of generative AI are deeply unreliable
Our interpretation
- The surrounding argument concerns the stock of high-quality human-generated material available in April 2024.
- The plain reading asserts that meaningful further improvement in output correctness requires exceeding that stock.
- Controlled improvements in error correction or objectively checked answers count as progress. They need not eliminate every hallucination.
- Later experiments can test the claimed necessary condition; this is not a forecast with a deadline.
Counts as supported if: Evidence establishes that meaningful further improvements require additional high-quality human data beyond the April 2024 stock, rather than changes in training, data use, or inference.
Counts as contradicted if: Controlled experiments demonstrate meaningful improvements in correctness using existing human datasets, self-generated training examples, or additional computation without requiring a larger human-data stock.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
Outcome: Contradicted
The dated evidence goes against what the claim committed to, under the criteria frozen before grading.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Controlled studies demonstrate meaningful improvements in error correction and answer accuracy using existing problems, generated training traces, and additional inference computation. That contradicts the asserted necessity of exceeding the available human-data stock for further progress. The evidence does not contradict the narrower concern that generative AI can remain unreliable.
Limitations
- The result depends on reading progress as meaningful improvement, not complete elimination of errors.
- The experiments concern mathematics and programming.
- Self-generated traces still rely on pretrained models, existing human problems, and correctness checks.
Evidence (3 items)
Goes against the claim: Google DeepMind researchers via arXiv,
training on self-generated data
Why it matters here: SCoRe uses existing MATH and MBPP problems, generated correction traces, and correctness rewards. Its Gemini 1.0 Pro coding experiments improve self-correction on held-out HumanEval, demonstrating further error-correction progress without requiring an expanded stock of human-written material.
- Publisher
- Google DeepMind researchers via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: UC Berkeley and Google DeepMind researchers via arXiv,
substantially improve test-time compute scaling
Why it matters here: Experiments on an existing MATH split show that inference strategies improve checked answer performance. The methods include specialized training, but do not require collecting a human-text stock beyond what existed in April 2024.
- Publisher
- UC Berkeley and Google DeepMind researchers via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Samia Kabir and colleagues via arXiv,
52% of ChatGPT answers contain incorrect information
Why it matters here: The study supports concern about unreliable programming answers before the post. It does not test whether improving them requires more human data than exists.
- Where
- Abstract
- Publisher
- Samia Kabir and colleagues via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Where's Your Ed At, Bubble Trouble
In essence, the AI boom requires more high-quality data than currently exists to progress past the point we're currently at, which is one where the outputs of generative AI are deeply unreliable.
- Where
- paragraph 5
- Published
- 2024-04-04
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment wyea-bubble-trouble-3b68e1-a1: Current. Outcome: Contradicted. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.