Remix clone Hacker News

Most people, as far as I'm aware, don't have an issue with the idea that LLMs are producing behaviour which gives the appearance of reasoning as far as we understand it today. Which essentially means, it makes sentences that are gramatical, responsive and contextual based on what you said (quite often). It's at least pretty cool that we've got machines to do that, most people seem to think.

The issue is that there might be more to reason than appearing to reason. We just don't know. I'm not sure how it's apparently so unknown or unappreciated by people in the computer world, but there are major unresolved questions in science and philosophy around things like thinking, reasoning, language, consciousness, and the mind. No amount of techno-optimism can change this fact.

The issue is we have not gotten further than more or less educated guesses as to what those words mean. LLMs bring that interesting fact to light, even providing humanity with a wonderful nudge to keep grappling with these unsolved questions, and perhaps make some progress.

To be clear, they certainly are sometimes passably good when it comes to summarising selectively and responsively the terabytes and terabytes of data they've been trained on, don't get me wrong, and I am enjoying that new thing in the world. And if you want to define reason like that, feel free.

▲

og_kalu 3 days ago | parent | next [-]

If it displays the outwards appearances of reasoning then it is reasoning. We don't evaluate humans any differently. There's no magic intell-o-meter that can detect the amount of intelligence flowing through a brain.

Anything else is just an argument of semantics. The idea that there is "true" reasoning and "fake" reasoning but that we can't tell the latter apart from the former is ridiculous.

You can't eat your cake and have it. Either "fake reasoning" is a thing and can be distinguished or it can't and it's just a made up distinction.

▲

suddenlybananas 3 days ago | parent [-]

If I have a calculator with a look-up table of all additions of natural numbers under 100, the calculator can "appear" to be adding despite the fact it is not.

▲

sourcepluck 3 days ago | parent | next [-]

Yes, indeed. Bullets know how to fly, and my kettle somehow knows that water boils at 373.15K! There's been an explosion of intelligence since the LLMs came about :D

	▲	og_kalu 3 days ago \| parent [-]
		Bullets don't have the outward appearance of flight. They follow the motion of projectiles and look it. Finding the distinction is trivial. The look up table is the same. It will fall apart with numbers above 100. That's the distinction. People need to start bringing up the supposed distinction that exists with LLMs instead of nonsense examples that don't even pass the test outlined.

▲

int_19h 3 days ago | parent | prev | next [-]

This argument would hold up if LMs were large enough to hold a look-up table of all possible valid inputs that they can correctly respond to. They're not.

▲

og_kalu 3 days ago | parent | prev [-]

Until you ask it to add number above 100 and it falls apart. That is the point here. You found a distinction. If you can't find one then you're arguing semantics. People who say LLMs can't reason are yet to find a distinction that doesn't also disqualify a bunch of humans.

▲

atemerev 4 days ago | parent | prev [-]

LLMs can _play chess_. With the game positions previously unseen. How’s that not actual logical reasoning?

▲

sourcepluck 4 days ago | parent [-]

I guess you don't follow TCEC, or computer chess generally[0]. Chess engines have been _playing chess_ at superhuman levels using neural networks for years now, it was a revolution in the space. AlphaZero, Lc0, Stockfish NNUE. I don't recall yards of commentary arguing that they were reasoning.

Look, you can put as many underscores as you like, the question of whether these machines are really reasoning or emulating reason is not a solved problem. We don't know what reasoning is! We don't know if we are really reasoning, because we have major unresolved questions regarding the mind and consciousness[1].

These may not be intractable problems either, there's reason for hope. In particular, studying brains with more precision is obviously exciting there. More computational experiments, including the recent explosion in LLM research, is also great.

Still, reflexively believing in the computational theory of the mind[2] without engaging in the actual difficulty of those questions, though commonplace, is not reasonable.

[0] Jozarov on YT has great commentary of top engine games, worth checking out.

[1] https://plato.stanford.edu/entries/consciousness/

[2] https://plato.stanford.edu/entries/computational-mind/

	▲	atemerev 3 days ago \| parent [-]
		I am not implying that LLMs are conscious or something. Just that they can reason, i.e. draw logical conclusions from observations (or, in their case, textual inputs), and make generalizations. This is a much weaker requirement. Chess engines can reason about chess (they can even explain their reasoning). LLMs can reason about many other things, with varied efficiency. What everyone is currently trying to build is something like AlphaZero (adversarial self-improvement for superhuman performance) with the state space of LLMs (general enough to be useful for most tasks). When we’ll have this, we’ll have AGI.