Pundit ClaimsAI claims by Gary Marcus and Ed Zitron, with the evidence

Gary Marcus Marcus on AI Published

Muddles about Models

We recorded 9 claims from this post. 4 were checked against evidence. 5 can’t be checked. 8 more are held. Read the post at its source.

Claims from this post (9)

In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.

  1. Assessment: Can’t be checkedSaid

    “Of course as Fei Fei Li correctly pointed out on X, it all depends on what you mean by a “model”, but here’s the crux, same as it ever was: correlations aren’t causal, semantic models.”

    What we tested: Correlations found inside LLMs are not causal, semantic world models.

  2. Assessment: Can’t be checkedSaid

    “Finding that some stuff correlates with space or time doesn’t mean that stuff genuinely represents space or time.”

    What we tested: The correlations with space and time in Gurnee and Tegmark's findings do not show that LLMs genuinely represent space or time.

  3. Assessment: SupportedSaid

    “It’s also not even remotely new.”

    What we tested: The finding that geography can be inferred from language corpora is not new; earlier work from 2009 and 2017 showed similar results.

    What happened: Louwerse and Zwaan recovered geographic relationships among 50 US cities using newspaper text, latent semantic analysis and multidimensional scaling. This directly establishes the earlier finding Marcus describes. (Cognitive Science; abstract reproduced by PubMed, )

  4. Assessment: UnresolvedSaid

    “It’s the same problem: correlations are fragile, and you can’t count on them.”

    What we tested: GPT's failure to stick to the rules of chess, even when it plays a decent game, reflects the same problem: correlations are fragile and cannot be counted on.

  5. Assessment: Can’t be checkedSaid

    “I would conjecture that you can’t say to ChatGPT, let’s play a game of chess and then (say) 37 moves in pause, and get it to reliably answer how many white bishops are left on the board, nor to explain what happened to some pawn that had been taken a few moves earlier.”

    What we tested: Marcus conjectures that ChatGPT cannot reliably say how many white bishops remain, or what happened to a captured pawn, partway through a chess game.

  6. Assessment: UnresolvedSaid

    “Without flexible underlying models that can reliably support inferences on novel questions, we can expect more of the same of what we are seeing over and over now: more hallucinations, more reasoning that is hit or miss, and nothing like the stability we should insist on for future AI.”

    What we tested: Without flexible underlying models, AI systems will keep producing hallucinations and hit-or-miss reasoning, and will lack the needed stability.

  7. Assessment: UnresolvedSaid

    “There is no internal model of what she actually did, and hence no capacity to support reliable inference, hence no guarantee that anything it writes about her will be true.”

    What we tested: LinkedIn's OpenAI-powered assistant has no internal model of what Kaya Yurieff actually did, and hence no capacity to support reliable inference about her, and hence no guarantee that anything it writes about her will be true.

  8. Assessment: Can’t be checkedSaid

    “To encapsulate the heart of what Ernie Davis and I said at length in our 2019 book Rebooting AI, AI is unlikely to ever be trustworthy if its makers continue to skip the step of having explicit cognitive models of the world.”

    What we tested: AI is unlikely ever to be trustworthy if its makers keep skipping explicit cognitive models of the world.

  9. Assessment: Can’t be checkedSaid

    “And here is why all this matters: safety, which is what Tegmark really cares about, is not going to emerge anytime soon, and probably ever, from hit-or-miss systems that substitute correlation for a factual, structured, interrogable representations of the world.”

    What we tested: AI safety will not emerge anytime soon, and probably never, from correlation-based systems that lack structured, interrogable representations of the world.

Held from this post (8)

These records did not pass every check, so nothing about them is published. Held is not a finding.