Hacker News Reader

Hacker News

Comment by lagt_t | Hacker News Reader

Parent Full thread

lagt_t

Everytime they say LLMs are the path to AGI, I cringe a little.

View on HN

1. AGI needs an interface to be useful.

2. Natural language is both a good and expected interface to AGI.

3. LLMs do a really good job at interfacing with natural language.

Which one(s) do you disagree with?

I think he disagrees with 4:

4. Language prediction training will not get stuck in a local optimum.

Most previous things we train on could have been better served if the model developed AGI, but they didn't. There is no reason to expect LLMs to not get stuck in a local optimum as well, and I have seen no good argument as to why they wouldn't get stuck like everything else we tried.

There is very little in terms of rigorous mathematics on the theoretical side of this. All we have are empirics, but everything we have seen so far points to the fact that more compute equals more capabilities. That's what they are referring to in the blog post. This is particularly true for the current generation of models, but if you look at the whole history of modern computing, the law roughly holds up over the last century. Following this trend, we can extrapolate that we will reach computers with raw compute power similar to the human brain for under $1000 within the next two decades.

More compute also requires more data - scaling equally with model size, according to the Chinchilla paper.

How much more data is available that hasn't already been swept up by AI companies?

And will that data continue to be available as laws change to protect copyright holders from AI companies?

It's not just the volume of original data that matters here. From empirics we know performance scales roughly like (model parameters)*(training data)*(epochs). If you increase any one of those, you can be certain to improve your model. In the short term, training data volume and quality has given a lot of improvements (especially recently), but in the long run it was always model size and total time spent training that saw improvements. In other words: It doesn't matter how you allocate your extra compute budget as long as you spend it.

In smaller models, not having enough training data for the model size leads to overfitting. The model predicts the training data better than ever, but generalizes poorly and performs worse on new inputs.

Is there any reason to think the same thing wouldn't happen in billion parameter LLMs?

This happens in smaller models because you reach parameter saturation very quickly. In modern LLMs and with current datasets, it is very hard to even reach this point, because the total compute time boils down to just a handful of epochs (sometimes even less than one). It would take tremendous resources and time to overtrain GPT4 in the same way you would overtrain convnets from the last decade.

True but also from general theory you should expect any function approximator to exhibit intelligence when exposed to enough data points from humans, the only question is the speed of convergence. In that sense we do have a guarantee that it will reach human ability

It's a bit more complicated than that. Your argument is essentially the universal approximation theorem applied to perceptrons with one hidden layer. Yes, such a model can approximate any algorithm to arbitrary precision (which by extension includes the human mind), but it is not computationally efficient. That's why people came up with things like convolution or the transformer. For these architectures it is much harder to say where the limits are, because the mathematical analysis of their basic properties is infinitely more complex.

LLMs aren't improving at things they're unable to do at all. An example being reasoning.

LLMs can reason. You can verify this empirically by asking questions which require reasoning to e.g. GPT-4.

False. https://arxiv.org/abs/2308.03762

This is not peer reviewed research and has some serious issues: https://news.ycombinator.com/item?id=37051450

There is evidence that they actually can't reason:

https://arxiv.org/abs/2311.00871

https://arxiv.org/abs/2309.13638

https://arxiv.org/abs/2311.09247

https://arxiv.org/abs/2305.18654

https://arxiv.org/abs/2309.01809

It sounds like you're arguing against LLMs as AGI, which we're on the same page about.

The underlying premise that llms are capable of fully generalizing to a human level across most domains, i assume?

Where did you get that from? It seems pretty clear to me that language models are intended to be a component in a larger suite of software, composed to create AGI. See: DALL-E and Whisper for existing software that it composes with.

The comment said that LLMs are the path to AGI, which implies at least that they’re a huge part of the AGI soup you’re talking about. I could maybe see agi emerging from lots of llms and other tools in a huge network, but probably not from an llm with calculators hooked up to it.

You're arguing that LLMs would be a good user interface for AGI...

Whether that's true or not, I don't think that's what the previous post was referring to. The question is, if you start with today's LLMs and progressively improve them, do you arrive at AGI?

(I think it's pretty obvious the answer is no -- LLMs don't even have an intelligence part to improve on. A hypothetical AGI might somehow use an LLM as part of a language interface subsystem, but the general intelligence would be outside the LLM. An AGI might also use speakers and mics but those don't give us a path to AGI either.)

The comment I was replying to was referencing OpenAIs use of the phrase "the path to AGI". Natural language is an essential interface to AGI, and OpenAI recognizes that. LLMs are a great way to interface with natural language, and OpenAI recognizes that.

While it's kind of nuts how far OpenAI pushed language models, even as an outside observer it's obvious that OpenAI is not banking on LLMs achieving AGI, contrary to what the person I was replying to said. Lots of effort is being put into integrating with outside sources of knowledge (RAG), outside sources for reason / calculation, etc. That's not LLMs as AGI, but it is LLMs as a step on the path to AGI.

I don’t know if they are or not, but I’m not sure how anyone could be so certain that they’re not that they find the mere idea cringeworthy. Unless you feel you have some specific perspective on it that’s escaped their army of researchers?

Because AI researchers have been on the path to AGI several times before until the hype died down and the limitations became apparent. And because nobody knows what it would take to create AGI. But to put a little more behind that, evolution didn't start with language models. It evolved everything else until humans had the ability to invent language. Current AI is going about it completely backwards from how biology did it. Now maybe robotics is doing a little better on that front.

I mean, if you're using LLM as a stand-in for multi-modal models, and you're not disallowing things like a self-referential processing loop, a memory extraction process, etc, it's not so far fetched. There might be multiple databases and a score of worker processes running in the background, but the core will come from a sequence model being run in a loop.

how come?

I just snicker.

Yea the idea that the computers can truly think by mimicking our language really well doesn't make sense.

But the algorithms are black box to me, so maybe there is some kind of launch pad to AGI within it

Reply on news.ycombinator.com