“They don’t care that the output is mediocre, that the product is unprofitable, that there is a quickly-approaching wall that generative AI can’t leap over as it runs out of training data, or that the transformer-based architecture of Large Language Models has hard limitations that are impossible to overcome.”
Our reading:
Generative AI faces a quickly-approaching wall it cannot get over as it runs out of training data.
Assessment: Unresolved
What happened
Did the forecast come true?
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
Epoch AI's June 2024 study projects full use of public human-generated text between 2026 and 2032 if trends continue; it does not establish exhaustion. [e1] The DeepSeek-R1 paper's January 2026 revision reports reasoning gains without human-labeled reasoning trajectories, but does not demonstrate gains after usable training data runs out. [e2] Neither item establishes the exhaustion condition or a resulting capability ceiling, so the forecast remains unresolved. [e1, e2]
Caveats
- Public human-generated text is only one possible source of training data.
- The packet establishes neither a lasting ceiling caused by data exhaustion nor an escape from that ceiling.
- Progress before verified exhaustion cannot settle this conditional forecast.
- The claim supplies no definite deadline or capability measure.
Evidence (2 items)
Context only: Epoch AI,
between 2026 and 2032
Why it matters: This projects exhaustion of public human text; it does not report that exhaustion occurred. The discussion explicitly allows continued progress through other data, improved data efficiency and different training choices.
- Publisher
- Epoch AI
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: DeepSeek-AI via arXiv,
advanced reasoning
Why it matters: The authors report reasoning improvements without human-labeled reasoning trajectories. This is relevant counterevidence to assuming that more human annotations are always necessary, but it does not establish that the forecast's exhaustion condition occurred.
- Where
- Abstract, version 2
- Publisher
- DeepSeek-AI via arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
How we judged it
- Kind of claim
- Forecast
- Deadline
- None given. We don’t add one.
Conditions
- as it runs out of training data (their words)
Where we had to interpret
- The claim predicts a lasting capability ceiling caused by exhaustion of usable training data.
- Exhaustion of public human-written text alone does not establish exhaustion of every usable source of training data.
- Quickly-approaching supplies no definite deadline. I do not substitute another researcher's projected date.
- The data-exhaustion condition must be established before scoring the predicted inability to overcome it.
Right if: Verified exhaustion of usable training data, followed by a demonstrated capability ceiling caused by that exhaustion, with evidence that alternative data sources or training methods cannot overcome it.
Wrong if: Verified exhaustion of the relevant training data followed by reproducible capability gains that overcome the same asserted constraint.
We wrote this test after the outcome was known.
What would settle it
Verified exhaustion of the relevant usable training data, followed by either a demonstrated capability ceiling that alternative data and training methods cannot overcome, or reproducible gains overcoming that same constraint.
We found 28 claims in this post. See them all.
Full record
Source and dates
Where's Your Ed At, The Shareholder Supremacy, paragraph 113.
- 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 2024-07-01.
Where it was said (1)
Said once in the posts we read.
: Where's Your Ed At, The Shareholder Supremacy
They don’t care that the output is mediocre, that the product is unprofitable, that there is a quickly-approaching wall that generative AI can’t leap over as it runs out of training data, or that the transformer-based architecture of Large Language Models has hard limitations that are impossible to overcome.
- Where
- paragraph 113
- Published
- 2024-07-01
- 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
Both reviewers disagreed with the draft’s verdict.