Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Large Language Models Are Neurosymbolic Reasoners

arxiv.org

Post reply on HN
Print view View original

151–160 of 172 posts

Re: Large Language Models Are Neurosymbolic Reasoners

#151

Post by eru »

klyrs wrote: post #125
still_grokking wrote: post #54

Earlier quoted context omitted.

Is it actually possible to beat Nethack without reading up some "spoilers" upfront? I've never heard of anybody who managed to do that. Even when you read up all kinds of info about the game before you attempt a run it's extremely hard to reach higher levels, yet beat the game. (I myself never reached any later levels despite I know some tricks by now. Tricks impossible to infer from just playing the game; you need t…

> Is it actually possible to beat Nethack without reading up some "spoilers" upfront? If people are playing nethack without reading the source, they're needlessly hobbling themselves.

Human readable spoilers are enough, no need to dive into the C source. People have already extracted most of the information and tables you need.

Especially if you only care about winning, and not about doing it eg in the fewest number of moves.

Re: Large Language Models Are Neurosymbolic Reasoners

#152

Post by eru »

edflsafoiewq wrote: post #28
clooper wrote: post #5

I was recently thinking how every neural network is equivalent to a lookup table where the input is all numbers up to what can be expressed within the context window and the output is the result of the arithmetic operations applied to that number. So every neural network is equivalent to T = {(i, f(i)) : i The main practical issue is the size of the table but I don't see any theoretical reasons why this is incorrect.…

This reminds me of the classic problem in computation, where the simplest form of computation, the lookup table, input -> output, is limited to a finite domain. Turing modified the computation to have a finite internal state and infinite external environment (tape), so it becomes a transition function (state, stimulus) -> (new state, response), applied recursively in a feedback loop, allowing it to operate on infinit…

Have a look at Post's correspondence problem for even crazier universal models of computation. Or at Fractran.

Simplified for Post's correspondence problem, you have a set of playing cards with text written on the front and back. (You can make copies of cards in your set.)

The question is, can you arrange your cards in such a way, that they spell out the same total text on the front and back?

As an example your cards might be: [1] (a, baa), [2] (ab, aa), and [3] (bba, bb). One solution would be (3, 2, 3, 1) which spells out bbaabbbaa on both sides.

Figuring out whether a set of cards has a solution is Turing complete.

Re: Large Language Models Are Neurosymbolic Reasoners

#153

Post by eru »

still_grokking wrote: post #144
neurallambda wrote: post #128

Earlier quoted context omitted.

"AI heats the planet"... really? You mean marginally? I'll assume you're asking in good faith. Using NNs allows this project to stand on the shoulders of giants: philosophically, mathematically, programmatically, but also I expect this to plug in to OSS LLMs, and leverage their knowledge, similarly to how a human child learns in a Pavlovian/intuitive response, and only later starts to learn to reason. Wrt inefficienc…

> Wrt inefficiency, training will be inefficient That's the "heating the planet part" I was referring to. :-) > but the programs can be extracted to CPU instructions / CUDA kernels during inference This just makes my original question more pressing: What are the NNs good for if the result will be normal computer programs? (Just created with astronomical overhead!) > Cyc looks cool, but from my cursory glance, is it c…

> This just makes my original question more pressing: What are the NNs good for if the result will be normal computer programs? (Just created with astronomical overhead!)

You know how expensive it is to pay humans to write 'normal' computer programs? In terms of both dollars and CO2.

Re: Large Language Models Are Neurosymbolic Reasoners

#154

Post by eru »

still_grokking wrote: post #77
neurallambda wrote: post #65

I'm trying to tackle this problem more head-on, by outfitting LLMs with lambda calculus, stacks, queues, etc. directly in their internals, operating over their latent space. [1] I'll read your paper, but, LLMs famously fail horribly at "multi jump" reasoning, which to me means they can't reason at all. They can merely output a reflection of the human reasoning that was baked into the training data, and they can also…

May I ask why there are NNs in your project at all? Just to heat up the planet and make Nvidia share holders even more happy? :-) I mean, what I've seen in the Readme makes sense. But doing basic computer stuff with NNs just makes the resource usage go brr by astronomical factors, for imho no reason, while making the results brittle, random, and often just completely made up. Also: Do you know about the (already 40 y…

> May I ask why there are NNs in your project at all? Just to heat up the planet [...]

> Also: Do you know about the (already 40 year old!) project Cyc?

Has Cyc accomplished anything so far? Or is it just to heat up the planet? The Wikipedia page makes it sounds pretty hopeless:

> Typical pieces of knowledge represented in the Cyc knowledge base are "Every tree is a plant" and "Plants die eventually". When asked whether trees die, the inference engine can draw the obvious conclusion and answer the question correctly.

> Most of Cyc's knowledge, outside math, is only true by default. For example, Cyc knows that as a default parents love their children, when you're made happy you smile, taking your first step is a big accomplishment, when someone you love has a big accomplishment that makes you happy, and only adults have children. When asked whether a picture captioned "Someone watching his daughter take her first step" contains a smiling adult person, Cyc can logically infer that the answer is Yes, and "show its work" by presenting the step-by-step logical argument using those five pieces of knowledge from its knowledge base.

Re: Large Language Models Are Neurosymbolic Reasoners

#155

Post by FrustratedMonky »

mjburgess wrote: post #136
calf wrote: post #131

Earlier quoted context omitted.

So I'm trying to understand your argument here, but why isn't "Reasoning is a cognitive process" circular logic? AlphaGo wasn't reasoning, how? I think reasonably AlphaGo has modeled a world, and it is by design subject to the bounded rationality of a game-theoretic optimization problem. So two of your criteria are satisfied. So - I'm just reading your definition here - AlphaGo wasn't reasoning because it is not a co…

> Reasoning is a cognitive process You could have a non-cognitive view of reasoning, or an embodied one. NNs are a-cognitive systems, they do not engage in reasoning of any form. Reasoning concerns inference across truth-apt propositions (eg., A->B, A thef. B). NNs have no propositions, nor are any parts truth-apt, true or false. NNs are statistical systems which select answers by weights found from optimisation. No…

sidebar

"If reasoning is an abstract pattern, then rice falling to the ground is likewise "reasoning"."

Technically, I think some philosophers over the ages have made that point, and argued that 'rice falling' is reasoning, and the rice is 'wanting' to be closer to the 'earth' or some such thing. It does sound wacky. Think Leibniz and monads made some argument like that.

And the various takes on 'Will to Power', Schopenhauer argued the will is a blind force. We don't have control of our own thoughts.

I think you are still giving humans too much credit for this aspect of 'cognition'.

"A man can do what he wills, but he cannot will what he wills." Schopenhauer

Re: Large Language Models Are Neurosymbolic Reasoners

#156

Post by FrustratedMonky »

mjburgess wrote: post #149
FrustratedMonky wrote: post #139

Earlier quoted context omitted.

This is kind of ignoring other NN's. You're very focused on LLM's as the example. AlphaGo learned by playing itself. And is able to anticipate multiple moves ahead. Then, that same 'engine', was able to be applied to Chess, and learned how to beat a master from scratch, by playing itself, in just a few hours. There was no lookups, or zip'ing of aggregated data. A lot of what you are postulating as cognition, humans d…

Yes, it is impossible for anyone to figure out gravity by induction. That's the problem with AI. The way we build explanations is largely by reasoning-by-analogy. We build physical models with our hands, to resolve ambiguities in our environment, ever more complexly -- and then, eventually, land upon the right analogy that then falls away. Prior to gravity we had crystal spheres -- reasoning by ananlogy with such thi…

Guess this is the crux:

"The full specification of a "world" in mathematical terms doesnt require intelligence. At that point you can use the dumb strategies of alphago."

I fall in camp that humans are just glorified amoeba, twitching at stimuli.

Eventually an AI could model 'us' with dumb strategies, because really baked into the human brain/body are just dumb strategies.

Re: Large Language Models Are Neurosymbolic Reasoners

#157

Post by mjburgess »

naasking wrote: post #138
mjburgess wrote: post #137

Earlier quoted context omitted.

Thermodynamic entropy isnt information in the relevant sense. There's a wide class of computational mysticism born of people going around and equivocating between "information" as it means radically different things where it is used. thermodynamic entropy (a real number) != information theory entropy (bits) != information in csi != information in stat mech' != information in QM != information in a turing machien !=.…

> but QM requires the state be real-valued (having infinite information in the computability sense). The unobservable state, which is merely a physical model that may have little resemblance to reality. All observable states necessarily have finite precision and beyond 60-70 digits are effectively undefined due to the uncertainty principle, which is yet another reason why people suggest physics is effectively computa…

No one here disagrees that measurement produces finite information; that is obvious and a necessary --- if it wasnt, we would never be able to know anything. Knowing requires an "early termination".

The issue is that there's no evidence this property of measurement is a property of reality, and all the methods, premises, etc. of physics attribute the opposite to reality.

Here, it is absolutely necessary for QM to work that the unmeasured state is real-valued,.

I'd also say that since measurement is finite in this manner, it then follows large swathes of reality are unknowable.. and this makes it clear why we cannot obtain the latent state of a QM system.

Re: Large Language Models Are Neurosymbolic Reasoners

#158

Post by mjburgess »

naasking wrote: post #135
mjburgess wrote: post #109

Earlier quoted context omitted.

Organisms are mechanistic, there's nothing "not mechanistic" about the mind. Its insane to suppose that somatosensory representation building, which requires organic neruoplasticity to be connected to the organically adaptive neuromotor system, etc. etc. etc. can just be instantiated in a bit of sand. This is deeply mystical, pseudoscience. You're credulously throwing away any kind of empirical analysis of the world…

> Its insane to suppose that somatosensory representation building, which requires organic neruoplasticity to be connected to the organically adaptive neuromotor system, etc. etc. etc. can just be instantiated in a bit of sand. Why? That's just an assertion, not an argument. All of those big words and sophisticated concepts you used all reduce to field interactions that we've mostly captured mathematically. Unless yo…

I'm just driven mad by how people thinking physical objects with specific properties is "magic", but reducing reality down to abstract platonic mathematical forms is "science".

This is the opposite.

Gold isnt lead. Lead isnt wood. Cells are not bits of metal. Bits of metal do not polymerise. Bits of sand do not form weak covalent bonds...

It's kinda exhausting that this mathematical superstition is so prevalent in people who otherwise believe they are somehow anti magical thinking.

There is no more magical thinking in supposing that you can ignore reality, describe it in a formula, reinterpet that formula against some other reality, and it'll all work out.

As if, "2 + 2 = 4" means the same thing when it's "2 drivers + 2 drivers = 4 deaths" vs. "2 cookies + 2 cookies = 4 happy children"

The idea that reality is essentially mathematical and not essentially physical is pythagorean magical thinking. Science says the opposite.

All the mathematics in explanatory scientific laws are just paraphrases of descriptions of the physical properties of systems. None of it is actually mathematics.

Physics does not study "2". It studies there being earth and the sun and a force between them, summarised as "2 masses" etc.

Re: Large Language Models Are Neurosymbolic Reasoners

#159

Post by mjburgess »

FrustratedMonky wrote: post #156
mjburgess wrote: post #149

Earlier quoted context omitted.

Yes, it is impossible for anyone to figure out gravity by induction. That's the problem with AI. The way we build explanations is largely by reasoning-by-analogy. We build physical models with our hands, to resolve ambiguities in our environment, ever more complexly -- and then, eventually, land upon the right analogy that then falls away. Prior to gravity we had crystal spheres -- reasoning by ananlogy with such thi…

Guess this is the crux: "The full specification of a "world" in mathematical terms doesnt require intelligence. At that point you can use the dumb strategies of alphago." I fall in camp that humans are just glorified amoeba, twitching at stimuli. Eventually an AI could model 'us' with dumb strategies, because really baked into the human brain/body are just dumb strategies.

Dumb is the wrong word, lets say, cheating.

All AI at the moment is just cheating with a fake UI.

We don't learn about the world by first being told what it is like. If you can fully specify an abstract system in mathematics, or use a historical corpus to answer questions --- you're nothing more than a kid cheating.

You seem to think that cynicism requires believing that animals are not, by construction, any different to incredibly absurdly dumb engineered works of our most over hyped morons.

This isnt cynicism, or scepticism, or erudition or sophistication. It's meglomania.

The whole history of evolution has not produced, in us and most animals, the most complex object (, likely,) in the entire universe to do something that alphago is doing. The level of ego here is off the charts.

This view is only a product of a pround ignorance of zoology (and so on) -- and a deep deep anti-intellectualism which says, "reality is easy to know, just build a computer program"

Re: Large Language Models Are Neurosymbolic Reasoners

#160

Post by mjburgess »

FrustratedMonky wrote: post #155
mjburgess wrote: post #136

Earlier quoted context omitted.

> Reasoning is a cognitive process You could have a non-cognitive view of reasoning, or an embodied one. NNs are a-cognitive systems, they do not engage in reasoning of any form. Reasoning concerns inference across truth-apt propositions (eg., A->B, A thef. B). NNs have no propositions, nor are any parts truth-apt, true or false. NNs are statistical systems which select answers by weights found from optimisation. No…

sidebar "If reasoning is an abstract pattern, then rice falling to the ground is likewise "reasoning"." Technically, I think some philosophers over the ages have made that point, and argued that 'rice falling' is reasoning, and the rice is 'wanting' to be closer to the 'earth' or some such thing. It does sound wacky. Think Leibniz and monads made some argument like that. And the various takes on 'Will to Power', Scho…

schopenhaur's idealism is indeed very similar, as is all idealism to this computational mysticism.

What none in this tradition considered possible is that the world exists; each reduced it down to a purely formal pattern one way or another.

Thankfully today we treat mental illness, and derealisation and depersonalisation and raised to this status.

I operate in a framework where there's a world and we're in it, and it is one way and not another, and the way it is arises from spatio-temporal properties that arent equivalent because the symbols we use in models of them are isomorphic.

In otherwords, I have passed through my phase of insanity and arrived back into the world where the grass is green because it is green; and the chair heavy, because it is massive,.

Post reply on HN