A. H. Zakai

Is there any evidence which *could* convince you that AlphaZero (or any other machine) is smart at chess?

To put it another way: you could search inside older chess engines but would not find such complexity.

When AlphaZero came out, it beat all previous engines by a lot. Since then, all top chess engines use neural network components, because they can reason incredibly well about chessboards.

You're right that at a high level, big data finds stuff.

But the amazing thing is that there is stuff to find! Finding them is only how we confirm that this machine computes complex concepts in chess, both ones humans already did, and new ones we can learn from.

Being able to look inside a chess engine and find new concepts that humans can learn from - that is an actually remarkable scientific result!

And it is not unique in this field.

+ "This work gives an end-to-end example of unearthing machine-unique knowledge in the domain of chess. We obtain machine-unique knowledge from an AI system (AlphaZero) by a method that finds novel yet teachable concepts and show that it can be transferred to human experts (grandmasters)."

Bridging the human–AI knowledge gap through concept discovery and transfer in AlphaZero | PNAS AI systems have attained superhuman performance across various domains. If the hidden knowledge encoded in these highly capable systems can be leve...

I agree. But that paper, and related work, *are* extraordinary evidence. This is a top-tier journal, which describes the work so:

www.pnas.org/doi/10.1073/...

"As AI systems become more capable, they may internally represent concepts outside the sphere of human knowledge." +

What could convince you that a machine was playing chess intelligently?

I ask because that paper & related ones are what convinced me:

1. a machine that beats humans,
2. inside it we find human-like tactics
3. inside it we find new tactics that surprise us

What additional evidence do you need?

And even if there *was* some conspiracy against rock, it would be from the record companies, not the White House or the CIA...

Even if so, the idea that a Republican administration fought to replace rock with... rap, is just laughable. If anything they would've tried country.

It sounds like your position is that no software - no machine - can possibly be intelligent?

What makes it possible for organic molecules to be intelligent but not machines?

Maybe look at

bsky.app/profile/phin...

To be fair, he talks about rock as a genre being pushed away, not his own band

But his theory is apparently that George W Bush (or the CIA) decided to replace rock with rap... so he's making an even stupider point than what you said

Some examples:

emscripten.org/docs/porting...

Safari Technology Preview 238 Release Notes | Apple Developer Documentation Learn about the latest web technology updates in Safari Technology Preview: Animations, CSS, Editing, Forms, Networking, Rendering, SVG, Scrolling, Web API, Web Inspector, WebAssembly, and WebRTC.

Safari Preview 238 adds WebAssembly support for JSPI:

developer.apple.com/documentatio...

JSPI lets wasm programs use async Web APIs in the sync ways they expect

Look, yes, I'm not going to argue that the majority of work in this area is good :)

But there *are* serious papers coming out that do try to make progress on the question of LLM intelligence. Good papers, arguing in different directions - so this is just not settled yet.

Doesn't that suggest the machine is intelligent about chess?

It isn't just doing some statistics on boards it saw, but analyzing them methodically. I would say it is reasoning about chessboards.

But the machine does analyze chessboards well enough to beat any human player, and this paper shows we can learn at least a bit about *how* it does that - and the things it does are quite neat, actually. It's not using brute force!

Anyhow, I am just not the right person to help out here, sorry.

I have no guess one way or the other whether LLMs can do this task.

I am used to reading academic articles (in math and related fields) in English, but this distinction was never necessary for my career, I guess.

Thanks for your patience here!

I follow the winnie-the-pooh example, I think, but struggle with the legal brief.

As context, I didn't go to school in an English-speaking country, so that might be my source of difficulty here. This task is quite unfamiliar to me.

I mean that "Chapter 4 of X" may be a fragment of a sentence that ends up connecting an argument to X (or it might not).

Or was the specific value of X here the reason it is a reference and not a citation?

citation

I am trying to use this definition I found - is it right?

"A citation is a reference to a source that supports a statement or is otherwise related to it."

www.law.cornell.edu/wex/citation

If so, "Chapter 4 of X" seems insufficient to tell either way?

Still, it is entirely possible that neural networks are intelligent in many spheres, but not when it comes to language. (It's much harder to measure there, at the very least!)

Researchers are debating LLM intelligence right now. It is literally an open question.

I think we have clear evidence that neural networks, the foundation of LLMs, are more than just statistical models. Those systems show verifiable, intelligent behavior on a wide variety of tasks. For example, in chess they have taught the best human players new tricks:

www.pnas.org/doi/pdf/10.1...

Other parts of our bodies can influence the brain, of course (like via hormones), but it is indirect. According to all the scientific evidence I've seen, we are only directly conscious of things in the brain (and not even the entire brain)?

> as if brains were standalone entities

I agree with your general points, but maybe not about this. For consciousness, at least, the brain does seem "standalone". Scientists working on the neural correlates of consciousness find stuff only in the brain, afaik.

Related to this, this is an interesting post by a former colleague: john.regehr.org/writing/clau...

where he says that the KINDS of bugs that CCC introduces are just very weird.

Sorry for taking you too literally (story of my life...)

At the risk of doing so again: I don't think we know the answer here. There are plenty of peer-reviewed studies showing cases where these machines can reason, and where they can't. I'm still curious about the answer for this specific task!

Curious what your thoughts are on Integrated Information Theory? That looks not at the number of components but at their causal connectivity:

en.wikipedia.org/wiki/Integra...

It predicts humans are conscious while (current) computers are not.