[27183]

> there seems to be some emergent behaviour that is not simply just next-token-prediction, or that the ability to do accurate next-token-prediction requires something "extra" that LLMs have.

The next step then would be to design and conduct experiments that isolate this effect. Figure out how to make it happen reliably and in such a way that you know it's actually happening as opposed to just something you're imagining. Isolate it or distill it so it can be studied directly. Until then, it's easiest to dismiss it as imaginary.

[sudb]

> Figure out how to make it happen reliably

And you're happy that the replication of LLMs across many foundation model companies is insufficiently reliable?

> just something you're imagining

So the alternative explanation you're suggesting to emergent LLM behaviour is mass independently-corroborated human hallucination. Which is more likely?

Also it really does seem like you've moved the goalposts a lot here without really giving me a substantive response.

[27183]

To say that LLMs' existence is evidence for emergent phenomena in LLMs is tautological. I'm merely suggesting if you want to make a claim about emergence it would be best, especially in absence of a convincing theory, to demonstrate it experimentally. Otherwise probably better not to claim it's actually happening.

[sudb]

> To say that LLMs' existence is evidence for emergent phenomena in LLMs is tautological.

This is not at all what I was saying. I think you've already conceded that LLMs demonstrate emergent behaviour but you dismissed it as a "close enough facsimile to intelligence". I was saying that the emergent behaviour is reliably replicable, in response to your following statement:

> Figure out how to make it happen reliably and in such a way that you know it's actually happening as opposed to just something you're imagining.

I think there is real work underway in the area of interpretability. In the meantime, there appears to be plenty of empirical evidence for the claim that LLMs exhibit some sort "intelligence" in the enormous penetration that agentic coding has achieved in software development? Do you deny the usefulness of LLMs here, or are you going to assert that actually software development requires no intelligence of any sort?

[27183]

> I think you've already conceded that LLMs demonstrate emergent behaviour

No. Please don't put words in my mouth. What I said is that an LLM compresses a bunch of information into a semantic embedding space and then does sort of a stochastic search in that embedding space. Any similarity to "intelligence" is accidental. You may look at the results of that process and "see" thinking or reasoning or something, but it ain't there.

> "intelligence" ... agentic ... usefulness

I don't think LLMs need to be intelligent to be (at least narrowly) useful. No more than random forests or genetic algorithms do at least.

[edit] Look, this has devolved to the point where it's no longer productive to continue. If you're going to state things like this as fact, there's really nothing more I can do here:

> emergent behaviour is reliably replicable

Go collect your Nobel prize then! This is no longer a discussion grounded in reality.

[sudb]

> but it ain't there

On what grounds? I don't think you've provided any evidence other than LLMs can't "adapt" or "learn" to show that LLMs do not show intelligence in any way. I think it's clear that there must be some emergent form of intelligence over words from just the agentic coding ability alone. I am not claiming that LLMs are intelligent, only that they display aspects of what we understand as intelligence.

> I don't think LLMs need to be intelligent to be (at least narrowly) useful

I agree! But they are more than narrowly useful, and they absolutely do not belong in the same category as random forests or genetic algorithms!

> Go collect your Nobel prize then! This is no longer a discussion grounded in reality

Once again you are being condescending while misrepresenting my position. The emergent aspects of "intelligence" have been replicated by virtue of independent LLM vendors training their own models - I am not making a stronger claim, you have misunderstood me.

Thanks for participating.

[layla5alive]

@sudb, I salute your effort. It is unfortunate how little this person took from your attempts at a genuine exchange.

Without moving goal posts: @27183: the word intelligence is defined as "noun. The ability to acquire, understand, and use knowledge."

LLMs can absolutely do this via in-context learning. They can even persist this as mementos on disk and re-attend to it later if its been purged from their context window.

You might try to make some comment like "they're not learning in the way humans do, they're not changing their weights, and they predict token by token!" Who cares? That's semantics - LLM attention is able to accumulate contributions of tokens and weights mixed together head by head and layer by layer with non-linear activations which functionally quack like intelligence.

They may be far different from us, but they're functionally intelligent (at least in several subsets of intelligence. And frankly, in some cases, they're more functionally capable than many humans).

Birds, planes, helicopters, rockets and gliders don't achieve flight in the same way, or have remotely the same capabilities in the air - but all are nonetheless forms of flight.