[ainch]

> It would be foolish to deny the effectiveness of image recognizers based on generalized adversarial networks (GANs), the key neural network technology underlying LLMs

I could be misreading this, but I hope the author doesn't think GANs are used in LLMs. They are cool, though.

[andy99]

It’s unfortunate - he’s an author, it would have been fine to stick with an authors perspective that LLMs can’t write (which is true), as well as the copyright stuff (which I don’t agree with but he certainly has standing to give an opinion on).

But he’s made the error of trying to come at it from a technical perspective, when he clearly knows nothing about that side of things, which discredits the rest.

[magicalist]

> But he’s made the error of trying to come at it from a technical perspective

Huh? That quote appears to be the entirety of the "technical perspective" of the post and is an aside from his larger points that you have blessed as "fine". Literally nothing in the rest of the post relies on that incorrect statement.

Let me quibble with what is discredited here, given the entirety of your point is built upon an error.

[red75prime]

Here's another technical point:

> They're word-association mechanisms with no embodiment and no way to associate the text vectors they manipulate with real-world phenomena.

This is wrong too. RLVR grounds foundational models in reality.

[Alpha3031]

Aren't RLVR signals typically based off formal systems not natural phenomena? The formal sciences are certainly useful for producing tools used in natural science, but it's not entirely clear they alone are sufficient to associate text to natural (real-world) phenomena. Honestly, the pretraining and RLHF are probably more tied to the world than RLVR, for all that RLVR might be useful (maybe even more useful) for making the model perform better in certain tasks (such as working with formally defined systems).

[red75prime]

Yeah, I should have said RL, not RLVR. The point is RL interacts with external world, while autoregressive pretraining is limited to the passive ingestion, and RLHF relies on a model of human preferences that has no access to truth sources besides the limited training data it was built upon.

[azakai]

If you want a more concrete example, then LLMs are also trained on visual data these days, which means they do have access to the world in an important way. This directly contradicts the blogpost's claim that LLMs have

> no way to associate the text vectors they manipulate with real-world phenomena.

Historically, that LLMs were text-only used to be a major argument for why they "lack access to meaning", see the Stochastic Parrot paper and the Octopus paper that it references. But even the authors of those papers have (grudgingly) conceded that the argument no longer holds due to multimodality.

[miltonlost]

How does a tiny error, in an aside and unrelated to the greater point, made at near the end discredit the rest of it?

[andy99]

A lot of people will stop reading or become completely distracted when encountering something so blatantly wrong, and it also indicates the author is willing to say such things. I see some people downplaying it, that’s fine, it’s clearly not helping the author make a case. Even im this discussion, the current top comment and a long thread are solely about that error (granted that’s not too unusual here)

[scarmig]

Things that are vaguely GAN-shaped might play some role in post training. Though no one would call them GANs or identify them as the key underlying technology.

[bonoboTP]

There is RLHF with a model that's trained to emulate a human evaluator, but the evaluator is not really trained jointly with the main model to adapt to its distribution and tell it from real text. Though I'm sure there are some niche cases when this is done. But definitely not a prominent thing.

[threethirtytwo]

He hallucinated. This is how you verify non AI nowadays.

AI is so good that if you see an hallucination as obviously wrong like this one it’s a sign it’s written by a human.

[sxp]

Yeah. Ironically, if he used an LLM to proofread his work, it would have told him that GAN's are generative not "generalized". That LLMs are primarily built on Transformers rather than GANs. And that they're more famous for image generation rather than image recognition.

[bonoboTP]

Transformers are an architecture and GANs are a training method for the architecture. There are GAN Transformers.

[andy99]

Do you believe it was this edge case he was referring to in making his point?

[bonoboTP]

No, I simply think the person is not familiar with these concepts, which is not a crime, it's okay to be wrong, I didn't want to banish that person from here, just corrected the statement. But might have been also a brain fart. Happens. Just wanted to correct it in the interest of beginner reading this.

Once Musk said he believes in Transformers instead of diffusion (or the other way around). When actually diffusion transformers are very popular and mainstream. What he meant was autoregressive inference vs diffusion. People like to use buzzwords while not knowing them. Old issue, it was the same decades ago.

To make it even more confusing, there are also diffusion LLMs, which typically but not necessary, use transformers also.

And independently of diffusion and transformer and llm or image generator, you can optionally put a GAN discriminator adversarial loss on any of them.

[nullstyle]

Examples for us less educated folk?

[Alpha3031]

Jiang et al. (2021) TransGAN: https://dl.acm.org/doi/10.5555/3540261.3541391

In general probably not much of a stretch to get rid of convolutions or recurrence by replacing with attention and see if it works hence the title of the original transformers paper.

[bonoboTP]

Analogy "it's not GAN, it's transformer!" - - "it's not a recursive implementation, it's object oriented!" if you are more familiar with CS.

Or "it's not 4-wheel-drive, it's diesel!", if familiar with cars.

[FeepingCreature]

And that GANs haven't been relevant in image generation since 2021.

[bryanrasmussen]

sorry but I see incorrect AI responses almost every day in things I query on Google, generally because I am often querying on something I am an expert in and I just want a linking source. There will be a good link in the result, generally, but the AI almost always messes up some elemental fact, sometimes it is because the AI has "misunderstood" what I asked which is reasonable, but most of the time the result shows that it has "understood" what I said but there will be a statement of fact that is just wrong.

[red75prime]

The AI you are talking about is a low-inference-cost model optimized for massive throughput.

It's not like every model has the same amount of those shortcomings.

[threethirtytwo]

Of course but for obvious mistakes like this one it’s more likely a human hallucinated than an LLM

[chpatrick]

Yeah, it's hard to take the rest seriously.

[pasquinelli]

maybe you're looking for an excuse to ignore what's being said.