It would help if you stated what your own (clearly incorrect) beliefs are in this area, so we can help correct them.

The point is that working with natural language tokens is very different than tokens that represent an image.

A simple relevant example is that if you ask an LLM to write a psychological thriller about a poor former student who commits murder and deals with intense moral guilt, in classic Golden Age Russian literature style, it is unlikely to sign it with "Fyodor Dostoyevsky."

It does that when generating images because, at a high level, image generation doesn't benefit from the kind of reasoning that language generation is able to.

▲

johnnyanmac 3 hours ago | parent [-]

>It would help if you stated what your own (clearly incorrect) beliefs are in this area, so we can help correct them.

Sure, let's re-examine what this chain is doing

1. "This is a pretty solid argument against people who argue that LLMs are more than just (very massive) next token predictors."

2. (you) "This issue has nothing to do with LLMs."

3. (me) "yes, it does"

4. (you) "an LLM does not generate images"

5. (me) "this is an LLM generating images"

6. (you) "an LLM does not understand what a 'signature' is"

So we are getting lost in minutae to asset that..."LLMs aren't much more than just (very massive) next token predictors.", agreeing with what the original comment is claiming.

There's a bit of meta-commentary seemingly missing from your context here. so I'll mention it. Some people are trying to claim that LLM's are "reasoning" with data, and that the way they "learn" isn't actually too different from human learning. Aspects of an LLM like this, being unable to reason about with the image it generated, are disproving such notions as of 2026. That is all the top comment in this chain is saying.

I hope that helps.

	▲	duskwuff 2 hours ago \| parent [-]
		Re. 5: LLMs do not generate images. LLMs prompt distinct, separately trained image models to generate images. The LLM has no ability to introspect the image model and cannot provide feedback during the image generation process. If the image model misinterprets the LLM's prompt (which can happen!) or inserts unexpected content, the LLM may become "aware" of that during subsequent chat steps as it ingests the generated image, but it cannot provide detailed control over their generation.