Sora Can’t Handle the Truth
We recorded 11 claims from this post. 8 were checked against evidence. 3 can’t be checked. 5 more are held. Read the post at its source.
Claims from this post (11)
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.
Assessment: SupportedSaid
“Turns out Sora makes stuff up, too,”
What we tested: OpenAI's Sora, like LLMs, makes things up.
What happened: OpenAI identifies an unintended physical error in a launch demonstration. This supports the existence of realistic-looking errors without depending on Marcus's chessboard analysis. (OpenAI, )
Assessment: SupportedSaid
“Aside from showing that (not surprisingly) Sora is pretty clueless about chess, it’s revealing about how the mysterious Sora even works.”
What we tested: Sora is pretty clueless about chess.
What happened: The report attributes the same elementary depiction errors to Rahul Agarwalla, rather than quoting Marcus. The event date refers to the original monkey demonstration, whose timestamp was checked on Altman's post; March 6 is the reporting date. (The Times of India, )
Assessment: SupportedSaid
“Nowadays, we have “multimodal” systems, trained on both text and images, and the aforementioned “hallucinations” (perhaps better called failed approximations, as Gerben Wierda has pointed out) and other errors persist and even crop up in new ways.”
What we tested: Multimodal systems trained on text and images still hallucinate and make other errors, sometimes in new ways.
What happened: The launch material documents invented entities in generated video. This supports the ordinary reading that multimodal generation retains fabrication errors and exhibits visual consistency failures. (OpenAI, )
Assessment: Can’t be checkedSaid
“It’s probably not fair to blame the weird board on a lack of data.”
What we tested: Sora's erroneous chessboard is probably not due to a lack of training data.
Assessment: UnresolvedSaid
“Rather, Sora is failing to apprehend a cultural regularity of the world, despite ample evidence.”
What we tested: In the chessboard example, Sora is failing to apprehend a cultural regularity of the world (the rules of chess) despite ample evidence.
Assessment: UnresolvedSaid
“But it’s drawing the wrong generalization.”
What we tested: Sora arguably shows some generalization, but in the monkey-chess case it is drawing the wrong generalization.
Assessment: UnresolvedSaid
“What’s missing is any global comprehension by the system.”
What we tested: In most or all of the Sora errors seen so far, virtually every bit of the images is locally correct, and what's missing is any global comprehension by the system.
Assessment: UnresolvedSaid
“Sora fundamentally revolves around trying to take pixels and reconstruct what pixels might be nearby in space and time. The forest is never perceived apart from the trees, and the chessboard is never perceived apart from its constituent squares.”
What we tested: Sora fundamentally works by reconstructing nearby pixels in space and time rather than perceiving whole objects.
Assessment: Can’t be checkedSaid
“From AGI perspective, however, this is an immense warning sign.”
What we tested: From an AGI perspective, Sora's fixation on the local at the expense of a deep representation of the global is an immense warning sign.
Assessment: Can’t be checkedSaid
“The sooner the AI industry -- and the world at large — recognizes this, the better.”
What we tested: The AI industry and the world should recognize soon that generative AI shares a persistent factuality and comprehension problem.
Assessment: SupportedSaid
“Gary Marcus is astonished that some people still haven’t recognized that generative AI has struggled mightily when it comes to world models, high level reasoning and factuality, even as evidence for that conclusion continues to pour in, even with data of ever-greater magnitude and sophistication..”
What we tested: Generative AI has struggled mightily with world models, high-level reasoning and factuality, even with ever more data.
What happened: The evaluation reports substantial failures to generate requested object relationships despite strong visual quality. This supports difficulties with spatial world consistency, without proving the absence of internal world models. (arXiv; Gokhale and colleagues, )
Held from this post (5)
These records did not pass every check, so nothing about them is published. Held is not a finding.