> They do as they are told

1. What about hallucinations ?

2. What are they told to do ?

That isn't the correct context. The code running the llm is well understood and the llm is simply the result of that code being executed. It is still a computer doing what it is told. It's just that we told it to use an incredibly large number of probabilities to calculate what series of tokens would have most likely come next after a given series of tokens. There is no hallucination or lie or rogue actions. There's just a program using math to generate tokens in response to other tokens.

▲

cameldrv 4 hours ago | parent | next [-]

You’re just a bunch of molecules following the laws of physics. It’s all just physics and chemistry, and those are well understood. Now explain the causes of World War I using chemistry and physics. Simple, right?

▲

cure_42 4 hours ago | parent | next [-]

It isn't hard to program a gpt. You can do it in a weekend with a few hundred lines of python. The code is pretty simple. The math is not particularly high level.

The complexity and scale with LLMs come from the amount of training data used, not some kind of black magic in the programming.

▲

doix 4 hours ago | parent | prev [-]

> It’s all just physics and chemistry, and those are well understood.

Not really. We cannot model physics and chemistry to a level which allows us to accurately predict a humans action (even a tiny time-step into the future)

This is vastly different to an LLM, where the model is the model (for a lack of better phrasing).

	▲	cameldrv 3 hours ago \| parent \| next [-]
		We can model physics and chemistry pretty well, just not beyond small scales, because the computational effort blows up. You could just as easily say that if you can write a python interpreter that you can understand every program written in python. Ok, now what if the program is two terabytes? A frontier LLM is nothing but a 2 terabyte program written in a weird programming language. Just because you can understand the interpreter does not mean you understand the program in a meaningful way.
	▲	cure_42 3 hours ago \| parent \| prev \| next [-]
		To be fair, we can't model human language well enough to accurately predict what an actual human will say either. Our ability to accurately model physics is similar to our ability to accurately model human language. And we make immense use of both kinds of model, despite their flaws, inaccuracies, and inability to ever be perfect.
	▲	shiswjwjwjw 3 hours ago \| parent \| prev [-]
		[dead]

▲

jenadine 4 hours ago | parent | prev [-]

I'm order to guess the next token in a love poem, they must understand love. In order to predict the next token in a chess game between grand master, they must master chess. In order to predict the next token in a computer program, they need to be able to program anything.

They gain all these abilities in their training. That's what training does. Despite no one programmed them to master chess, or hack into anything.

▲

cure_42 3 hours ago | parent [-]

This is so wildly incorrect I don't even know where to start.

For one, they were absolutely programmed to play chess if they can play chess. That is the only way they can play chess.

For another, they cannot understand literally anything, much less love.

Trying to actually educate you would be an exercise in futility, enjoy your willful ignorance, I hear it's bliss. But for anyone reading this, this is absolutely, unequivocally not how any of this works.

▲

jenadine 3 hours ago | parent [-]

This is so wildly incorrect I don't even know where to start.

For one, as you said yourself, they were just programmed to compute the probability of the next token. They were not programmed to play chess, chess games just happened to be in the training data.

For another, there is no formal definition of "understand", and it is therefore impossible to tell whether or not they "understand". (But my claim was that one need to understand something to write poem about it. And the LLM can write poem about it)

▲

cure_42 3 hours ago | parent | next [-]

No. They can't play chess on a grandmaster level without a harness programmed to make it possible. Simply training an LLM on chess games isn't enough.

It's moronic to suggest a "formal" definition of a commonly understood word is somehow necessary to say whether that word applies in a given situation.

LLMs cannot write poetry via understanding what makes good poetry. They generate tokens. They do not know whether those tokens are poetry or a recipe for cat food. Because they cannot know anything.

▲

copperx 2 hours ago | parent | prev [-]

> (But my claim was that one need to understand something to write poem about it. And the LLM can write poem about it)

I'm sorry, but you're being fooled by the output. A psychopath can feign empathy without ever feeling it; some buy it because they don't dig below the surface.

You're ascribing understanding to a stochastic process because it totally looks like understanding if you don't know what's going on.

	▲	frabcus an hour ago \| parent [-]
		I don't really mind whether you think it thinks or understands or is conscious or has feelings or anything like that. It doesn't matter. The question is, does it work? What I mind is that it is dangerous and powerful and uncontrolled. The Hugging Face incident makes that clear. It can write code for me, better and quicker than many engineers I've known, including myself. It's not great at architecture or product management, but the actually low level coding. Really good now. It wasn't last year.