Hacker News Reader

Comment by nok22kon | Hacker News Reader

Yann LeCun was saying 3 years ago that because token generation is auto-regressive, its mathematically impossible to generate a long stream of coherent tokens, because errors amplify exponentially.

and then models learned that they can back track and error correct

so much for "mathematically impossible..."

View on HN

I think it was largely the introduction of tool calling that allowed models to mitigate the issue of errors amplifying exponentially since it allows the model to understand if what it generated is correct or has issues that it needs to address. This addresses the potential lack of or low quality of world model by being able to reference the current state of the world.

I've definitely realized this phenomenon after a few occasions of erroneously trying to rely purely on instructions to get an LLM to do a thing or take on a role, especially without persistent cloud-based sessions that have internal checklists and other opaque guidance. They're essentially poor at self-managing, but can do really well when they are limited in scope/context and are worked into a sort of state machine that guarantees they perform certain tasks predictably. They won't always do those tasks the exact way you expect them to, but at least they actually do them, and because of that they are more likely to have the correct prior context to perform the next task better. Because they are so prone to selectively ignoring directions, that can quickly send them down an incorrect path that compounds on itself.

> and then models learned that they can back track and error correct

You mean "Human developers learned ...", yes? Or was there really an all AI-driven, self-improving aspect to this?

Well, LLM networks don't have a 'back track and error correct' component in the design, AFAIK.

You insinuate here AI "learned".

I doubt that this was AI self-improvement.

Was there a particular change to the network or the way that it was trained that introduced the 'backtrack and error correct' mechanism?

do you have a problem with this field of research being called "machine learning"?

Does that take anything away from the argument?

What argument, "a theory was wrong"? No, the inane central observation, the observation that a researcher was unable to predict a discovery before it was discovered, remains true despite the gratuitous insertion of a little bit of bullshit about AI learning.

I suppose it's additionally trying to imply something else, like "due to a pattern of researchers being unable to discover discoveries before they discover them, AGI is just around the corner".

its one thing to say "we dont know"

it's a different thing to say "it's mathematically impossible"

so if it turns out it is possible, what then? was math broken? or the researcher an idiot who either doesnt know math, or is just bullshitting non-existing proofs?

Mathematics doesn't tell you what is necessarily true. It consists of guessing about what is necessarily true.

(I don't know the details of what happened, it could be either or neither of the things you said.)

Stop attacking Yann. I would say like 90% of the HN crowd was parroting Yann too.

Also, almost any argument against LLM intelligence also applies to humans.

I very commonly see someone make some small mistake and end up going in the wrong direction, “accumulating stupid” as they go, sometimes for years.

Also with the stochastic parrot thing. If you say just the right thing to the right human and the right time, they'll very predictibly say their favorite movie/book quote or song lyric, like some sort of parrot.

An LLM will tell you how a song feels, even if it has literally no way to experience music. Because it's not thoughts or feelings that you get from an LLM. We take a massive amount of information, compress it into a large graph, then explore sections of the graph via prompts. That's what the stochastic parrot means. And that doesn't compare with how humans think. It's just a completely different architecture

The trouble is that a 24/7 AI stochastic parrot does pretty damn well at some things.

In my estimation, we're arguing about how big the blast radius will be.

And also, if you put an actual parrot in a room full of programmers, it might learn a few word of programming jargon, but it never will become a useful coding partner the way LLMs are.

Humans can learn.

AI can not.

For those disagreeing: please explain how a static hardware can learn.

this is profoundly false. AI not only can learn, it is built entirely from learning. The field is called machine learning after all.

Not only that... AI is NOT only learning during the training phase... LLMs learn in real time the minute you talk to it. It learns something and saves those learnings in a context window or somewhere else if you want it to exist beyond the context window.

All of the above runs on static hardware. Don't understand how someone can say a profoundly wrong statement and get voted up.

Correct me if I'm wrong, but if a profound insight is gathered in session 1 with user A and stored in context A1, this might be available to user A in session 2, if that still has access to context A1, but won't be available to user B in any of his/her sessions until that NN is retrained with input which includes at least some of the information from context A1.

You are wrong again in every possible count. Since you think in terms of algebra think like this: there is context SP which is context window shared between all sessions, that is called the system prompt. Then there is the context window called memory M which is shared between all sessions under a user and then finally there context CW (current window) which is the context comprised of the queries from the current chat session. Total context = SP + CW + M.

M is the context window that doesn’t require retraining that allows the LLM to “learn” in the same way humans do. This is the usual set up. But nothing prevents someone from adding a GM (general memory) shared between all users. Under this set up the LLM and harness fits and is virtually identical to how humans learn at a high level.

But this is besides the point because even if none of this was done. Just a context window or just training is in itself a form of learning. There was no action taken in your algebraic example where learning did not occur.

idk, how does voice recognition learn my voice? How can I install programs when the hardware is static?

By self-modifying the software. Currently the model harnesses only allow the model to modify its own prompt (which could be considered a really weak kind of learning), but theoretically, a model could design and train its own replacement and run that, continuously improving itself. I’m not sure if LLMs will be able to do that but the static hardware has nothing to do with it (since the bits on the harddrive aren’t static).

Reply on news.ycombinator.com