Clojurians Log v2

Not me. Trump’s opinion on something doesn’t tell me anything about reality, either way. If something is 50/50 likely to be true, it remains 50/50 after he’s weighed in, regardless of his stance on the matter.

At most, it might tell you that he’s got personal vested interest in you thinking one way or the other.

I wish they'd stop calling it AI, that the problem, there is no Intelligence in LLM's. They are just machines, there is no emergent behaviour.

I don't think it's fair to say there is no emergent behavior. Neural networks have behaviors that an individual “neuron” inside them does not. I do wish the news would separate LLMs from other forms of machine learning, but I think the differences would be lost on a nontechnical audience.

the question is, can the study of the individual "neuron" allow predications about the (novel) behaviour of the whole... (my assumption is yes).

All behaviour is contained in the training data.

The answer to that question doesn't change whether or not the system exhibits emergent behavior, but looking at a single neuron does not allow you to predict the behavior of the whole.

The issue is that we're communicating it to normal people. The word emergent in science has nothing extraordinary to it. But to normal people it does. In science, a cake exhibits emergent properties. It's not a very special property, as tons of mundane things show emergent properties all the time.

What term should they use instead?

Nothing. Just drop the word and talk directly to what the model can do.

And how should the news refer to the models?

They have to describe them somehow

Why do people not say emergent when talking about baking? It's simply not a quality that matters to anything discussed. Instead of: Almost by magic, full thoughts and sentences emerged from the models, you could see them think and reason, solving complex problems. You can say: As hypothesized, the model architecture generated reasoning traces and managed to put together coherent sentences.

Sorry, I think I've misunderstood you. I agree that "emergent" is not a useful descriptor for news about LLMs to use.

I meant to ask what they word they should use instead of AI.

Oh, that I'm not sure. I think I prefer just calling them models, you can say transformer model, diffusion model, large language model, etc. But the ship has probably sailed on not calling it AI. I maybe would have preferred the whole AI field being called II for Intelligence Imitation or something less anthropomorphizing.

And ya, maybe I didn't express myself all clearly. I think emergent as a word exists colloquially and in science in a different interpretation. So saying emergent to a layperson reinforces the human-like quality of consciousness, alive, unique, etc. Same as saying intelligent, AGI, ASI, all these give the impression that something that is basically human will appear in front of you, or even god-like in the case of ASI. Whereas in the technical sense, all these imply a lot less. Emergent is a non-intersecting property that ice cubes and traffic jams exhibit as well. It doesn't have anything to do with human consciousness. Intelligence doesn't imply human thinking, but just behavior that seems to deliver on the desired goal outcome, or choosing the better option out of many, basically anything that does better than random luck at a task is "intelligent" in science. Similarly AGI and ASI mean that for any task of the same modality (so say tasks you can do with a computer) the model can do it as well or better than a human (and it's unclear if we mean average human or expert at that task human). Without any constraints of time/energy efficiencies.

for clarity, I'm using "emergence" as a (potential) property of a complex system, that is a system which is disordered/unpredictable/non-linear, such as people and society, here experts are less useful as the past is not a predictor of the future. In contrast to complicated systems which are ordered/predictable/linear, such as software (including LLM's), here experts are useful, their experience tells them which recipe/pattern to apply for the problem at hand (which they've likely tackled in the past, or something very similar).

Im gonna say II from now on :D everyone is probably going to think I am trying to do some Kind of british accent, which makes this even funnier:smile:

Or i just call it Aye-aye, based on that llms often send signals indicating willingness to follow a command

The thing I'd worry about is the damage it’s going right now. If these companies actually believed the work they were doing could lead to extinction they would stop now. It's clearly a lie to mollify “doomers.”