“Large language models are here to stay, their recipe well understood, and not hard for people skilled in the art to replicate; large pretrained models are out their for the taking on the dark web.”
Our reading, the sentence that was checked:
LLMs are here to stay: their recipe is well understood, skilled practitioners can replicate them without much difficulty, and large pretrained models are freely available on the dark web.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Should we worry more about short-term AI risk or long-term AI risk?, paragraph 24.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-should-we-worry-more-about-short-139304-a1
We recorded 11 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2023-03-14. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- 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
- Large language models are here to stay
- their recipe well understood
- not hard for people skilled in the art to replicate
- large pretrained models
- for the taking on the dark web
Our interpretation
- In this observation, here to stay summarizes the difficulty of reversing existing diffusion of models and technical knowledge; it is not assigned an invented forecast deadline.
- Well understood refers to practical construction recipes, not complete scientific understanding of model internals.
- Replication could mean training a comparable base model, adapting existing weights, or running a downloaded model. These are assessed separately.
- Dark web ordinarily means hidden services requiring specialized access. Public torrents and ordinary websites establish unofficial access but do not alone establish that specific location.
- The ongoing accessibility claim is assessed during the publication month through March 14, 2023.
Counts as supported if: Contemporaneous records establish practical recipes, replication without substantial difficulty by skilled practitioners, and freely obtainable large pretrained weights on the dark web.
Counts as contradicted if: Direct evidence establishes that a material asserted component was false in the target period. Public-web availability alone does not disprove parallel dark-web availability.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
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
The evidence supports important parts of the statement: large-model weights had leaked through a verified public torrent, and Stanford demonstrated adaptation using a published recipe. It does not establish the specific dark-web location or easy replication from scratch. Calling the entire claim supported would require treating dark web as any unofficial distribution and replication as reuse of existing weights. Those readings were not specified. Public-torrent evidence also does not prove that dark-web copies were absent, so contradiction is unwarranted.
Limitations
- The leak is established through reputable reporting rather than a directly inspected primary distribution archive.
- Fine-tuning and running pretrained weights require different resources from pretraining.
- The permanence phrase cannot independently be settled as an indefinite forecast.
Evidence (3 items)
Supports the claim: Stanford Center for Research on Foundation Models,
a model fine-tuned from the LLaMA 7B model
Why it matters here: The team demonstrated practical adaptation and published its recipe and training code. It relied on an existing pretrained model, so it does not establish that training a comparable base model from scratch was easy.
- Where
- Opening summary
- Publisher
- Stanford Center for Research on Foundation Models
- Kind of source
- Demonstration
- Describes events of
- Public from
Supports the claim: The Verge,
leaked online.
Why it matters here: The report dates a public torrent to March 3 and describes researchers verifying that the leaked weights matched the official model. This supports unofficial availability, but does not establish distribution through dark-web hidden services.
- Publisher
- The Verge
- Kind of source
- News report
- Describes events of
- Public from
Context only: arXiv; GPT-NeoX-20B authors,
training and evaluation code
Why it matters here: The earlier primary record documents independent construction and release of a large model and its code. It provides background for diffusion of the recipe, but cannot by itself establish March 2023 accessibility or low replication difficulty.
- Publisher
- arXiv; GPT-NeoX-20B authors
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Authenticated March 2023 hidden-service distribution records, together with a defined replication task and contemporaneous evidence of the effort and resources it required. Clarifying that dark web meant unofficial public distribution would materially change the assessment.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Should we worry more about short-term AI risk or long-term AI risk?
Large language models are here to stay, their recipe well understood, and not hard for people skilled in the art to replicate; large pretrained models are out their for the taking on the dark web.
- Where
- paragraph 24
- Published
- 2023-03-14
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-should-we-worry-more-about-short-139304-a1: Current. Outcome: Unresolved. 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.
About assessment marcus-should-we-worry-more-about-short-139304-a1.