“There is no way to create enough data to train these models, and little that we've seen so far suggests that generative AI will make anybody but Nvidia money.”
Our reading:
There is no way to create enough data to train these models, and little suggests generative AI will make money for anyone but Nvidia.
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 evidence available by April 4, 2024 does not settle either clause [e1–e4]. May 2023 experiments found diminishing returns from repeated training data, while June 2023 research demonstrated useful synthetic data for a 1.3-billion-parameter coding model [e2, e1]. An April 1, 2024 paper reported that accumulating real and synthetic data prevented model collapse in its tested settings [e3]. These results establish neither sufficient generated data for continued frontier-model improvement nor its impossibility [e1–e3]. On February 27, 2024, Klarna projected $40 million in profit improvement from its AI assistant for calendar 2024 [e4]. This supplies an indication of prospective profit outside Nvidia, but one company projection does not settle whether such indications were scant or substantial by publication [e4].
Caveats
- The research does not establish whether generated data could meet the frontier training requirements contemplated in the claim [e1–e3].
- Klarna's estimate is a company projection, not verified realized profit or a comprehensive account of commercial prospects [e4].
- The claim specifies no frontier data requirement or threshold for 'little' evidence of prospective profit.
Evidence (4 items)
Context only: Gunasekar et al., arXiv,
Despite this small scale
Why it matters: The abstract describes a 1.3-billion-parameter coding model trained with selected web data and GPT-3.5-generated textbooks and exercises. This is a concrete earlier success for synthetic training data, but its scale and task do not establish sufficient data for frontier successors.
- Publisher
- Gunasekar et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Muennighoff et al., arXiv,
the value of adding compute
Why it matters: The experiments identify diminishing returns from repeatedly using limited data, while also finding benefits from limited repetition. They substantiate a data constraint, but do not establish that every method of creating usable data must fail.
- Publisher
- Muennighoff et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Gerstgrasser et al., arXiv,
accumulating data prevents model collapse.
Why it matters: The paper reports theoretical and experimental mitigation of collapse by retaining accumulated real and synthetic data. This challenges an inevitable-collapse argument, but does not establish the quantity or quality of data needed for the frontier training contemplated by the post.
- Publisher
- Gerstgrasser et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Klarna Bank AB, distributed by PR Newswire,
$40 million USD in profit improvement
Why it matters: The event assessed is the public announcement of a prospective profit benefit from Klarna's OpenAI-powered assistant. It is relevant evidence suggesting future profit outside Nvidia, but remains a company projection and cannot by itself establish how much evidence exceeds 'little'.
- Publisher
- Klarna Bank AB, distributed by PR Newswire
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Measures
- Klarna, Projected profit improvement from its AI assistant, USD millions, company reported, 2024-01-01 to 2024-12-31
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Where we had to interpret
- Assess the evidence available by publication on 2024-04-04. 'Seen so far' includes earlier demonstrations without an arbitrary starting date.
- In context, 'enough data' means usable training data sufficient to continue improving frontier language models. Training a smaller model successfully is relevant but does not alone establish that frontier data requirements can be met.
- The ordinary reading of the commercial clause concerns evidence suggesting future profit outside Nvidia. Credible projections, demonstrated savings and commercial adoption can bear on it; realized profits are not required.
- 'Anybody' includes companies using generative AI as well as companies supplying it.
- The financial scope records prospective profit, but the speaker supplies no accounting basis, amount or profit period. No accounting convention is imposed.
- The target period has no lower bound because the statement explicitly concerns everything seen so far, rather than a particular reporting quarter.
Right if: Contemporary evidence would need to establish that available data-generation methods could not supply usable data for the frontier training discussed, and that credible indications of future profit outside Nvidia were scant. Failures of particular methods or a lack of realized profits would not alone establish both assertions.
Wrong if: Contemporary evidence would need to demonstrate an adequate data-generation method at the relevant frontier scale, or provide substantial, credible indications of prospective profit outside Nvidia. A small-model demonstration or an isolated promotional forecast would be relevant counterevidence but would not automatically settle the entire statement.
We wrote this test after the outcome was known.
What would settle it
Research available by April 4, 2024 demonstrating adequate generated data for the relevant frontier training, or establishing that available methods could not supply it, would settle the data component. A broader commercial record available by that date, including credible profit projections, attributable costs and savings, and adoption evidence, could establish whether indications of future profit outside Nvidia were scant or substantial; realized profit is not required.
We found 24 claims in this post. See them all.
Full record
Source and dates
Where's Your Ed At, Bubble Trouble, paragraph 49.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2024-04-04.
Where it was said (1)
Said once in the posts we read.
: Where's Your Ed At, Bubble Trouble
There is no way to create enough data to train these models, and little that we've seen so far suggests that generative AI will make anybody but Nvidia money.
- Where
- paragraph 49
- Published
- 2024-04-04
- 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.