[throw310822]

Jesus... This morning while I was drinking coffee and staring at the screen (it's Saturday) an agent did the equivalent of days of my work, reading code, understanding, hypothesizing, comparing, using tools, writing scripts, launching compilers and running tests, identifying problems and proposing solutions, and more. Only someone who hasn't spent a second reflecting about what it means to think and to be intelligent can claim that we miss a realistic path to intelligence. It's so damn clueless and stubborn and confidently wrong that it annoys me immensely, so sorry for the rant.

[davemp]

> clueless and stubborn and confidently wrong

Uhuh. I really shouldn’t be replying to this type of comment from a throwaway.

But the extremely powerful semantic search that we get from LLMs isn’t enough. I don’t think anyone is credibly arguing otherwise?

Agents already are a layer on top trying to bridge the gap. But they’re really just using LLMs as a heuristic to explore extremely NP problem spaces. The notable successes with agents so far are when we can provide them with a solid verifier and preferably additional context hints on the steps to take in the problem space. See the test oracle problem on where this gets us.

So forgive me if I think that it would be enough of a jump in computational complexity to remove those guard rails that it’s not feasible. But don’t say that I’m clueless, stubborn, or confidently wrong.

[27183]

> reading code, understanding, hypothesizing, comparing, ...identifying problems and proposing solutions, and more

Except it did none of those things, really, because that's not how it works. This might help, it's a good writeup: https://www.0xkato.xyz/how-llms-actually-work/

We know how these machines work, it's not mysterious, there's nothing "extra" happening.

[eihli]

Now show me a writeup that explains how the brain works so I can understand why the brain does those things.

> We know how these machines work, it's not mysterious, there's nothing "extra" happening.

It sounds like you're saying "We don't know how brains work, they're mysterious, there's something 'extra' happening", and using that as justification for why you're saying a computer, an AI, can't "understand".

I think most people on Hackernews now who would use the phrase "my AI worked overnight and hypothesized, compared, etc..." already know how an LLM works, and still chooses to use those words. So the issue isn't that they don't understand. It's that they understand and still use those words. So the disagreement is somewhere else.

[27183]

I'm not claiming there's something "extra" happening in brains. Merely that we just don't know how they work well enough to use that knowledge to do engineering. Neural nets are quite unlike brains, despite the unfortunate shared vocabulary.

OTOH we do know how neural nets work, and they definitely don't do "thinking" or "reasoning".

[fragmede]

First you have to give a specific definition for thinking and reasoning before determining if they definitely do or don't do such a thing.

[27183]

Fair enough, I should have phrased that less strongly. Until you show that your neural net does "thinking" or "reasoning" I'll disregard that and prefer to think about it in terms of what we actually know neural nets actually do. Does that work?

[fragmede]

Just start with your definition for reasoning. If A implies B, and B implies C, does A imply C? Does that count as reasoning?

[sudb]

This feels like a semantic disagreement to me? If an LLM got to an acceptable end result code-wise, what would you call the process that took place to get it there?

[27183]

I'd call it what it is: a good enough stochastic search result extracted from the model's embedding space.

[sudb]

Is an implication of this that models are incapable of producing entirely novel code?

Also, not to get too reductionist about this, but what do you posit is special about what is happening when humans think? Intelligence is hard to define so clearly, I reckon.

[27183]

> Is an implication of this that models are incapable of producing entirely novel code?

No, it does not imply that at all. Google "temperature in LLMs".

> what do you posit is special about what is happening when humans think?

I don't. And IIUC nobody knows, but I'm not a brain scientist. There have been some wild theories over the years (recall Penrose's). I don't really have a dog in the hunt, except that probably whatever is happening is physical. It doesn't really matter, except insofar as whatever is happening very probably isn't what LLMs are doing. We know enough about what an LLM does, and what a brain does, to be quite certain they don't work the same.

[sudb]

> Google "temperature in LLMs".

No need to condescend, I'm very aware of what temperature is for LLMs. But I'm going to push back - if you're claiming all LLMs simply do is a stochastic _search_, how can that produce novelty, in the conceptual sense? (I'm not, for example, talking about novel rearrangement of existing ideas and code)

> We know enough about what an LLM does, and what a brain does, to be quite certain they don't work the same.

I don't think the claim is that LLMs do what brains do - I think the correct form of the counterargument is that _whatever LLMs seem to be doing_ produces end results that were previously only possible through the application of human intelligence, so there must be some axis of however you define human intelligence that LLMs currently seem to display as an emergent behaviour.

[0xDEAFBEAD]

Just a few comments ago you argued that we don't know how to build superintelligence. Now you're saying we know how the (unevenly superintelligent) Fable system works.

It doesn't seem like you're being consistent here. I'm concerned there might be some motivated cognition going on.

"What is true is already so. Owning up to it doesn't make it worse. Not being open about it doesn't make it go away. And because it's true, it is what is there to be interacted with. Anything untrue isn't there to be lived. People can stand what is true, for they are already enduring it."

[27183]

I'm not following the "Fable" stuff, what is that? [edit: ah it's a new language model from Anthropic.. yeah still don't see how it's relevant here]

[throw310822]

> We know how these machines work, it's not mysterious, there's nothing "extra" happening.

Lol. This is more telling about your implicit unscientific preconceptions that you wanted to reveal. Of course there isn't anything "extra". Where do you think intelligence comes from, some mysterious realm? It's physical, computational. The fact that at the bottom we produced it via matrix multiplication is irrelevant. Maybe humbling. You are denying a visible fact (a machine performs tasks that require flexible analytical and cognitive skills) precisely because there is no magic happening anywhere.

[27183]

> Where do you think intelligence comes from, some mysterious realm? It's physical, computational.

Well, no. I don't think it comes from some mysterious realm. I think that which is not physical does not exist [edit: and if you like I'll follow that one right down the rabbit hole--continuity and infinity are useful delusions]. But that eminently does not mean we know what intelligence is, let alone how to build one.

> The fact that at the bottom we produced it via matrix multiplication is irrelevant.

Huh? We don't even know what "it" is. How can you say you produced it?

> a machine performs tasks that require flexible analytical and cognitive skills

You see that, I see a lucky stochastic search result. Don't underestimate the "creativity" of random algorithms! They can do some wild shit! This is nothing new, we've been playing with these toys for like 70 fucking years. It's only recently that they started spewing words and everyone lost their minds over it.

[throw310822]

You do realise then that "we know how it works, there is no "extra" there" is an argument that can be used against any artificial intelligence, now or in a thousand years, as well as (at some level) against human intelligence (no magic, it's all physics, just dumb cells exchanging signals). This should be enough to give you pause- you immediately reached for an argument that is entirely empty.

> I see a lucky stochastic search result

Again you're reaching for a mechanistic explanation of some kind (let's leave for the moment whether it makes sense or not) as if having an explanation somehow contradicted a display of intelligence. It doesn't. Yes of course we made it, we know how it works (ar some level) and there is no magic. But what matters is the result- this machine, matrix multiplier, stochastic parrot, consistently displays intelligence, to the point of being able to perform very complex, open-ended tasks that integrate discovery, planning, tool usage, decision and even some aesthetic sense, understanding and using natural language, context awareness, you name it.

> This is nothing new, we've been playing with these toys for like 70 fucking years

Lol no. For god's sake. Hundreds of billions of parameters organised in a specific architecture and trained with unimaginable amounts of data and compute? Unless by "these toys" you mean "any computer program vaguely AI-related".

[27183]

> But what matters is the result- this machine, matrix multiplier, stochastic parrot, consistently displays intelligence, to the point of being able to perform very complex, open-ended tasks that integrate discovery, planning, tool usage, decision and even some aesthetic sense, understanding and using natural language, context awareness, you name it.

IDK, it doesn't seem like they actually do any of that. To me it seems like they have good enough semantic embeddings that they can kind of approximate those things, sometimes, well enough if you don't look too hard. This is enough to fool people. Of course there's gold in them hills--some recent mathematical results were found there. But to say that's "intellgence" is to say that lossy compression is intelligence. It's static. It does not learn. It does not adapt.

> Unless by "these toys" you mean "any computer program vaguely AI-related".

Not "vaguely AI related". I mean stochastic computer programs that can do things that look awful thinky. They've existed for a long time, but only recently (due to word2vec and other advances) have the results been words that mostly go together well instead of numbers. For some reason people seem to think a lot less critically when the output is words. IDGI but it's a whole thing.