Hacker News Reader

Comment by ericmcer | Hacker News Reader

If you really think about what an LLM is you would think there is no way that leads to general purpose AI. At the same time though they are already doing way more than we thought they could.

Maybe people were surprised by what OpenAI achieved so now they are all just praying that with enough compute and the right model AGI will emerge.

View on HN

LLMs are the language center of an AI. They in no way purport to be: long term memory, critical thinking engines, calculators.

If we want those things we can build them. Building them into the language center would be absurd and weird.

This is a very good take on the current state of LLMs. They are able to articulate surprisingly well and LLMs or something similar will definitely be part of a larger AI model in the future. They are good creative aids, but they should not be trusted.

In the brain all those things are neural networks too, memory is encoded in the weights. So if we want to have that type of system, everyone should have their own model and the weights would actually adjust all the time just like in our brain, but now it is just trained once and that is that

And yet this is what all the big companies we re trying to do, because the LLM is incredibly expensive fixed cost, and they marginal cost for each new task.

> If you really think about what an LLM is you would think there is no way that leads to general purpose AI

It is an autoregressive sequence predictor/generator. Explain to me how humans are fundamentally different

"Prove me wrong?" That's not how this works. Your implicit claim here is that human cognition and LLM functioning are fundamentally similar. That claim requires substantiation.

It is how it works if you are replying to someone who claims "If you really think about what an LLM is you would think there is no way that leads to general purpose AI". The counter example is human beings are considered general purpose intelligence and we are complex, but fundamentally predictable systems (not by us today), with (as far as we can tell) deterministic outputs based on the state of the universe (including physical being like brain and chemicals).

Responding to an unsubstantiated claim with an unsubstantiated claim just adds another unsubstantiated claim.

So far as I know, whether the universe behaves deterministically remains an unsolved question. Given that, your statement here would already be one of belief rather than fact, even before we get to the parentheticals. There is information here, but not about whether LLMs can develop into AGI.

Fine, you can ignore my previous comment, that's just my answer to the question that this discussion ultimately takes you to. But I feel like you are just sitting on the sidelines making strawmen and playing pedantic games instead of saying anything constructive.

The original comment said:

> If you really think about what an LLM is you would think there is no way that leads to general purpose AI.

This is an inflammatory way to state an extreme position on a well-discussed debate over whether next-token prediction can lead to general intelligence. The original commenter clearly believes it can't get you there. If you want to say that with any authority, you need to have an answer for what is different between what we consider general intelligence (for most people, this is simply human intelligence) and what models are capable of. This is the question at the heart of artificial intelligence.

I challenged them to explain their answer. I made no claims, I asked no one to prove anything wrong. If it is obvious that LLMs can't be AGI, the answer to how an LLM differs from human intelligence is also obvious, right?

> I made no claims, I asked no one to prove anything wrong.

Your original comment was:

> It is an autoregressive sequence predictor/generator. Explain to me how humans are fundamentally different.

Which would be interpreted by most reasonable people as you making the claim that humans are autoregressive sequence predictors, and asking people to prove you wrong. I can see how you could say this without intending to make that claim, but most people will interpret this as you making that claim.

I do not intend to inflame things or discredit what you are saying, but just to say that if you did not intend to make a claim or ask people to prove you wrong, a different approach would be more successful in the future.

You're right, it was hastily written and I was annoyed.

But I generally hold out hope that people can see a claim "A!=B" and a response "A=C, explain how C!=B" and understand that is not the same as claiming "C=B", especially on HN.

I do remain convinced my interpretation was sound, but on review I have to concede it was also quite uncharitable.

With all the wildly overheated claims that've been flying around since the advent of these models, I may be myself somewhat overfitted. Granted, in such an environment I feel like a little extra care for epistemic hygiene is warranted. But there was no reason for me to be rude about it.

I know what you mean. Unfortunately, it's easy for frank and concise language to be taken the wrong way when in written form (and sometimes even verbal form). I wish I didn't have to make qualifiers about my intent on my internet comments, but I often do, to try and make sure that other people take my comment the way I intended it. I think it generally leads to better discussion.

I don't blame people for not wanting to talk this way.

I actually did a full write-up on this here fyi: https://photonlines.substack.com/p/intuitive-and-visual-guid.... You can skip most of this and scroll down to the end-section called 'The Mental Model for Understanding LLMs' where I try to map how transformers are able to mimic human thinking. I think that comparing them to auto-associative / auto-regressive networks is actually a really good analogy FYI and I do believe a lot of human learning is based on this.

An article this long needs a table of contents IMO.

Will note this and try to add a TOC next time I do this sort of write-up. This article wasn't supposed to be this long - I originally meant to make a more compact version but kept on finding more things which I needed to add in order to map to a correct 'mental model' of what ChatGPT is.

We don't have sufficient understanding of human thinking and mental processes to make any claims like the ones you are trying to make. If we did, psychology would be a quantitative hard science; it most definitely is not.

We have no idea how human thinking works and no idea what consciousness is.

I'm getting so tired of listening to software engineers LARP pseudo neuroscientists with 6th grade level insights.

>Of course, the [AI] brain isn’t ‘conscious.’ It doesn’t have any survival instincts which we humans do.

Bruh...

So instead of shouting insults, maybe you can shine some light on how anything stated there is wrong? Everything that I said there is true - we don't have a good mapping of what consciousness is - if you want to add some insights add them instead of making insulting and useless statements on HN.

We have no idea if ChatGPT is conscious or not. We have no idea if it is "thinking" or not.

What we know with a reasonably high level of certainty is that consciousness and "thought" are physical processes. That's about it.

Pulling out the scalpel to start dividing up what physical process is and isn't conscious is a fools errand. And especially foolish when just making up arbitrary qualifications for it.

Am I saying that ChatGPT is conscious? No. But I am saying is you shouldn't give much credence to people who are anything more than agnostic about it.

Human neurons are continuous input, with active dendrites and dendritic compartmentalization. Spikey artificial NNs seem to hit problems with riddled basins so far. A riddled basin is a set with no open subsets.

Feed forward networks are effectively DAGs, and circuit like, not TM like.

Caution is warranted when comparing perceptrons with biological neurons.

Dendrites can perform XOR operations before anything makes it to the soma for another lens.

While there is much to learn, here is one highly cited paper on dendritic compartmentalization.

https://mcgovern.mit.edu/wp-content/uploads/2019/01/1-s2.0-S...

I think that the perceptron model of learning plasticity is on pretty shaky ground as being a primary learning model for humans.

AI needs to see thousands or millions of images of a cat before they reliably can identify one. The fact that a child needs to only see one example of a cat to know what a cat is from then on seems to point to humans having something very different.

> AI needs to see thousands or millions of images of a cat before they reliably can identify one.

Not if they inherit from a previous generation of AI. But even if they did, a different training speed does not imply a different capability

My point is not that humans have a faster training speed but that humans must be doing something fundamentally different from LLMs. You could build Altman's $7 trillion dollar GPU cluster and use the majority of the world's energy to feed it and you'd still hit the same limitations if you're just running an LLM on it, even a very sophisticated LLM. This is Yann LeCun's position as well.

That, as I understand it, is not a valid chain of logic. Requiring fewer data points does not inherently indicate that the underlying mechanism (autogressive sequential generation, not the transformer which is just an architecture) is different.

Not to mention the secondary arguments like - no proof that human learns faster from fewer datapoints, that's just your assumption in the sibling comment. Humans inherit information. The equivalent - fine-tuning a foundation model - is very fast to learn novel objects.

Just because someone has a Turing award doesn't mean they know what they're talking about. They are just people, with strengths and weaknesses like everyone else. But often on the extreme end of strengths and weaknesses.

Well, the same thing goes for you - just because someone posts on HN doesn't mean they know what they're talking about. And if I have to decide whose assessment I trust regarding AI, I take the Turing award winner who worked for almost 40 years on AI over a random guy from the internet.

Sure. But I'm not standing here saying my argument is valid because SolidAsparagus made it.

> And if I have to decide whose assessment I trust regarding AI, I take the Turing award winner who worked for almost 40 years on AI over a random guy from the internet.

I'd encourage you to do your own thinking and make up your own mind.

Humans train on continuous video. Even our most expensive models are, in terms of training set size, far behind what an infant processes in the first year of their life.

EDIT: and it takes human children a couple years to reliably identify a cat. My 2.5 y.o. daughter still confuses cats with small dogs, despite living under one roof with a cat.

I contend that you could show any child old enough to communicate in basic English a photograph (so not live continuous video) of some obscure animal they've never seen before (say an Okapi) and they'd be able to easily identify another Okapi when seeing one at a zoo.

So you're just going to ignore the 5 years of continuous training? I'm not sure what point you're trying to make.

My daughter is 5 y.o., which means because of kindergarten, I spend plenty of time about kids this age. A random kid this age would absolutely fail your test. They may remember the word after one exposure, but I doubt they'll remember any of the distinctive features.

Hell, many adults would fail it. I'm not sure if I could pass such test - in my experience, you remember the important details only after first experiencing a test and realizing what exactly it is that would be useful in distinguishing the two animals.

It's a massive amount of data. Not just video, but all senses, continuously. Also, I think an important aspect is that the training data is interactive. From day one, an infant sees his own hands moving and begins testing cause and effect. They aren't passively absorbing.

If the model is first pre-trained on unlabeled images, then it takes about 10 labeled images of cats and 10 labeled images of dogs to train a (possibly strong) classifier (example: DINOv2), I doubt humans will do better.

That's a good point; when comparing performance with humans, one has to remember that a human spends years of training on unlabeled images in form of continuous video stream, on top of audio streams and other senses, before they're able to communicate with you so you could test them.

Even language is not sequential.

Tell me more?

Language is only sequential in the form it is transmitted (verbally). There is no reason that sequential statements are generated sequentially in the brain. Quite the opposite, really, if you consider rules of grammar.

I really wish most of the LLM folks just took a few courses in linguistics. It would avoid a lot of noise.

Since LLMs seem to follow the rules of grammar quite well, and this is quite the opposite of "sequential", I guess LLMs do not process informations very sequentially.

Any pointers for where to start? Linguistics is dense, the terminology is a slog, it seems huge, and I was never clear that there was real value to be gained. But it's always useful to be more informed.

LLMs don't generate their language sequentially either, they just output it sequentially token by token.

Whoever cares about "rules of grammar"?

Grammar is descriptive, it formalizes the language so it doesn't break down into regional dialects too fast, and otherwise is just a crutch for people learning the language, especially if it's not their first language. The way you acquired your first language is the same way LLMs learned to utter grammatically correct sentences: by being exposed to lots and lots of examples, and eventually getting a feel for it. Similarly, if you're fluent in a language, you don't even think of grammar when using it - the right phrases in correct forms just come to you.

We don't know how we are different, but its pretty clear that we are fundamentally different. Modern machines calculate much faster than humans, if human thoughts and reasoning would work like LLM, we would already sit in front of general purpose hard AI. But we don't, so obviously there is something missing. Unfortunately its quite hard to know what is missing, because we don't know how conscience and reasoning actually work and we can't formularize them. Thats why its fair to assume we won't see general purpose AI that fast (probably not in our lifetime), which a lot of AI researchers seem to support, especially those who are not working for companies who try to sell AI.

Reply on news.ycombinator.com