[LamaOfRuin]

Only if you redefine "reasoning". This is something that the generative AI industry has succeeded in convincing many people of, but that doesn't mean everyone has to accede to that change.

It's true that something interesting is happening. GP did not dispute that. That doesn't make it reasoning, and many people still believe that words should have meaning in order to discuss things intelligently. Language is ultimately a living thing and will inevitably change. This usually involves people fighting the change and no one know ahead of time which side will win.

[peterlk]

I don't think we need to redefine reasoning. Here's the definition of "reason" (the verb): "think, understand, and form judgments by a process of logic"

If Claude 4 provides a detailed, logical breakdown in its "reasoning" (yeah, that usage is overloaded), then we could say that there was logical inference involved. "But wait!", I already hear someone saying, "That token output is just the result of yet another stochastic process, and isn't directing the AI in a deterministic, logical way, and thus it is not actually using logic; it's just making something that looks convincingly like logic, but is actually a hallucination of some stochastic process". And I think this is a good point, but I find it difficult to convince myself that what humans are doing is so different that we cannot use the word "reasoning".

As a sidenote, I am _very_ tired of the semantic quagmire that is the current AI industry, and I would really appreciate a rigorous guide to all these definitions.

[zahlman]

> Only if you redefine "reasoning". This is something that the generative AI industry has succeeded in convincing many people of, but that doesn't mean everyone has to accede to that change.

I agree. However, they can clearly do a reasonable facsimile of many things that we previously believed required reasoning to do acceptably.

[quesera]

Right -- we know that LLMs cannot think, feel, or understand.

Therefore whenever they produce output that looks like the result of those things, we must either be deceived by a reasonable facsimile, or we simply misapprehended their necessity in the first place.

But, do we understand the human brain as well as we understand LLMs?

Obviously there's something different, but is it just a matter of degrees? LLMs have greater memory than humans, and lesser ability to correlate it. Correlation is powerful magic. That's pattern matching though, and I don't see a fundamental reason why LLMs won't get better at it. Maybe never as good as (smart) humans are, but with their superior memory, maybe that will often be adequate.

[CyberDildonics]

> they produce output that looks like the result of those things

Is a cardboard cutout human to some degree? Is a recording a voice? What about a voice recording in a phone menu?

> LLMs have greater memory than humans,

So does a bank of hard drives by that metric.

[quesera]

(Memory Access + Correlation Skills) is a decent proxy for several of the many kinds of human intelligence.

HDDs don't have correlation skills, but LLMs do. They're just not smart-human-level "good", yet.

I am not sure whether I believe AGI will happen. To be meaningful, it would have to be above the level of a smart human.

Building an army of disincorporated average-human-intelligence actors would be economically "productive" though. This is the future I see us trending toward today.

Most humans are not special. This is dystopian, of course. Not in the "machines raise humans for energy" sort of way, but probably no less socially destructive.

[CyberDildonics]

HDDs don't have correlation skills, but LLMs do

So which is it, the memory or the correlation? I'll give you a hint, this is a trick question.

[quesera]

I never suggested that it was one or the other.

I think it's very obviously both.

(and these two qualities are likely necessary, but not sufficient)

[ToValueFunfetti]

It would be useful to supply a definition if your point is that others' definition is wrong. Are you saying they don't deduct inferences from premises? Is it "deduct" that you take issue with?

[zahlman]

They do not perform voluntary exploration of the consequences of applying logical rules for deduction; at best they pattern-match. Their model of conceptual meaning (which last I checked still struggles with negation, meta-reference and even simply identifying irrelevant noise) is not grounded in actual observational experience, but only in correlations between text tokens.

I think it should be abundantly clear that what ChatGPT does when you ask it to play chess is fundamentally different from what Stockfish does. It isn't just weak and doesn't just make embarrassing errors in generating legal moves (like a blindfolded human might); it doesn't actually "read" and it generates post-hoc rationalization for its moves (which may not be at all logically sound) rather than choosing them with purpose.

There are "reasoning models" that improve on this somewhat, but cf. https://news.ycombinator.com/item?id=44455124 from a few weeks ago, and my commentary there https://news.ycombinator.com/item?id=44473615 .

[ToValueFunfetti]

Okay, sure. My intuition is that LLMs reason at about a three-year-old level which appears more impressive because of their massive memories. By your definition and criticism, I take it that you wouldn't describe a three-year-old as capable of reasoning, so we're probably on the same page.