[andix]

Exactly. Just look at what they are really useful right now. Running LLMs in feedback-loops (agents) so they can try out random-ish approaches until some verification function passes (tests).

It's like the infinite monkeys on typewrighters that will type whatever you are looking for, given infinite time. LLMs are just tuned to much better odds than the monkeys are. But it's still a lot of randomness, with random results.

[roadside_picnic]

> It's like the infinite monkeys on typewrighters that will type whatever you are looking for, given infinite time.

In the monkey example the infinite time is doing a lot of work there. The fact that LLMs can search through semantic space and find reasonably correct paths in a reasonable time is directly tied to the reason why they are valuable.

Saying "these two things are similar except one can be useful and one can't" is not a great comparison.

For me the real lesson learned isn't how "smart" LLMs are, but rather how much human work is basically reducible to repeating past work with minor variation. Human's believe they are "reasoning" but so much code writen is just the human brain doing the same autocomplete style work that LLMs can do now.

[tikhonj]

The point is that it's the same process with—much—better priors.

This seems like a reasonable view to me. It's surprising just how much better priors matter and how we can develop those priors by training on a bunch of text. But it also explains, or at least hints at an explanation, for why LLM capabilities are so jagged, and in such inhuman ways.

[dpark]

> The point is that it's the same process

Except it’s not at all the same process. The fact that LLM are non deterministic is not the same as churning out random garbage.

[tsunamifury]

The literally churn out random garbage and are trained over time for that garbage to look more and more like an acceptable outcome to humans.

It’s training monkeys at typewriters through reinforcement.

[dpark]

> trained over time

So not random.

> acceptable outcome to humans

And not garbage.

It’s real weird to see people argue that LLM output is no different than random gibberish and then handwave over the fact that it’s clearly not with terms like “training”, as if a steam of random garbage is trainable.

[andix]

> but so much code writen is just the human brain doing the same autocomplete style work that LLMs can do now.

That's the part they are really good at. But they are really bad at taking complex decisions. Most of them are just guesses from a finite amount of solutions they were trained on, or from options they have in context.

[godwinson__4-8]

Indeed. Humans are well known for being good at "taking complex decisions" for which they have no "training", "options" or "context".

[andix]

Humans have a much bigger "context window". They remember many things they did an hour ago, a week ago, or even years ago.

[godwinson__4-8]

Yes, and your ability to remember a relatively few things that happened years ago is predicated on your ability to also forget most things that happen to you - like what you had for dinner last week. Good thing we have technology to fill in the gaps.

And nothing about this makes your initial comment any less goofy. Anyone who has ever had to make a difficult decision knows more than half the battle is preparation. Where do you think complex decisions come from? Have current events left you with the impression that people just waltz into idk say the Situation Room and just big brain their way through world events? That's how the current administration seems to think the world works, with quite predictable results.

Society is already algorithmic. To optimize for humans being dumb. AI is nothing more than another advance along this continuum. No one is impressed by your ability to remember something years ago, many if not most mammals have the same capability. Human recall is also notoriously bad in many cases - see numerous studies on the reliability of eye witnesses testimony.

AI is smart because most people are dumb. Come to terms with the fact that your anthropocentrism need not be based on a notion of intellectual supremacy and you'll be a far less tedious person to deal with.

[andix]

You didn't convince me, that I'm the tedious person to deal with here.

[nkrisc]

Humans also generally have the will to live.

[godwinson__4-8]

Indeed. It's almost like the LLM was the one that invented the "tactical" nuke in the first place.

[harry8]

>Saying "these two things are similar except one can be useful and one can't" is not a great comparison.

Launching a nuclear war is an interesting definition of "useful", not one I'd agree with and that exact scenario is what is being discussed.

So yes this is a perfectly valid and useful comparison in examining this particular, civilisation ending limitation.

[Folcon]

I mean to a point?

You do have to successfully write something the first time

We already acknowledge this to a degree, what is experience other than having done something similar before?

That first time though, you've got to figure something out that time

[mettamage]

Hmm saying it’s random-ish is doing it a disservice. I understand it’s a stochastic process but there’s definitely some level of understanding. Not at the level of lived experience but usually an LLM with vision capabilities can call a spade a spade and do something useful with it. And when a verification function shows how they are wrong then they usually come with a better and more informed approach.

So I can’t fully see how that’s related to the infinite monkeys. A typewriting monkey doesn’t have access to a verification function. And even if it did, it would not be the original concept anymore with infinite typewriting monkeys producing the works of Shakespeare.

Nevertheless, I upvoted your comment because it’s definitely insightful.

[dwattttt]

"understanding" is overstating it. Correlation between tokens embedded in the weights via training, yes.

[anon84873628]

Feedback loops certainly seem to give them some level of understanding.

Agent reads a skill file about how to use a CLI tool. It tries to use the tool but gets an error about the input format. It tries again with a different format based on the error message, and sees that command succeeded. It compares what worked to what was in the skill file and notes the difference. On future invocations it continues to use the new format.

Is that not "understanding" how to use the tool?

[hgoel]

What exactly would you call understanding? It's a correlation matrix of concepts.

[mountainriver]

What’s the difference? It’s clearly processing information and coming up with the right answer

[varjag]

Training is a loan word used to describe human learning process. For a reason.

[andix]

Humans learn on the job. LLMs don't. Very important difference.