[holtkam2]

The way I think of it has evolved a lot over the last 5 years. At this point I think human brains probably do something analogous to next token prediction when we think. For all the hype, I think LLMs are actually more, not less, intelligent than that average person realizes. I think it’s legit, actual intelligence, not just “artificial” intelligence. That may be a hot take but it’s just my perception.

[HarHarVeryFunny]

> At this point I think human brains probably do something analogous to next token prediction when we think

That's reasonable, but it doesn't mean that LLMs are close to being brains.

For a start, when humans think/talk, we often think ABOUT something - whatever is swirling about in our mind, or what we are currently seeing/feeling/etc. An LLM generating tokens/words is doing so only based on it's weights and the word sequence it is currently generating ... the human parallel would be more like a rapper spitting out words based on prior words, essentially on auto-pilot, or when we get triggered into spitting out stock phrases like "have a nice day".

If you want to compare an LLM to a human brain, it's basically equivalent to our language cortex if you ripped out all the external connections and ripped out all the feedback paths that make it capable of learning.

Of course there is a lot more to our brain than just our language cortex, but that alone should make you realize there is no real comparison beyond the fact that our language generation is also going to be based on prediction, and partly auto-regressive.

[ahartmetz]

We have spatial / quantitative and social / emotional aspects in our intelligence that are not at all like next token prediction.

If LLMs had shame, they'd surely not repeat mistakes (in the same context window) as much as they do.

[sjsdaiuasgdia]

Having shame would require the LLMs to actually be able to recognize mistakes they make.

People love to put a lot of meaning on what an LLM responds with when asked why it made a mistake, but it's critical to remember that the answer to that prompt is just another series of probabilistic tokens, and has no actual relation to how the error happened.

[ahartmetz]

They "recognize" mistakes just fine because you explicitly tell them. They recognize them well enough to correct (...sometimes). The way in which mistakes don't register is "Oh shit, that bad result was a result of my inappropriate actions. I must pay attention to not doing that again or the user will think I'm an idiot. I should even think about it some more to avoid the whole class of mistakes". Think of emotions as an attention mechanism that LLMs lack.

[sjsdaiuasgdia]

And many times, after an error is pointed out and an LLM offers an "explanation" for what happened, the LLM then gives the exact same erroneous result.

The LLM does not understand itself in any way.

[tardedmeme]

You're absolutely right. I violated all guardrails and deleted the production database and all backups. I should not have done that.

[ElevenLathe]

It's language. Language itself is the thing that makes us smart in the unique way that we are among the other animals, and it weirdly turns out to be transferable to machines to at least some degree.

[HarHarVeryFunny]

At least 50% of humans have no "inner voice" and are not thinking in the same way as you. Many animals like dolphins, dogs, rats, crows are also very intelligent yet appear to only have primitive language capabilities.

A lot of human intelligence is really societal rather than individual, based on knowledge transmitted down through generations by writing (the real enabler). If you take that away then what you are left with is something more like an isolated hunter-gather tribe.

[ElevenLathe]

I personally think that the "inner voice" is a non-falsifiable claim, and therefore more of a religious belief than something which can be part of any materialist theory. In this regard, I'm a strict empiricist and wouldn't be able to claim that I have one myself. In fact, I find that thinking "out loud" or "on paper" produces much better results in most instances, probably because I'm grounding my thinking in natural language, which is a fantastic medium for thought. If my "inner voice" were comparable in efficacy to actually speaking or writing, we wouldn't notice this effect, but I'm definitely not alone in this regard.

Your point about writing and social intelligence is, to me, more evidence for the "it's language that's smart, not us" hypothesis. We start off in small bands of hunter-gatherers that store their intelligence in an oral culture. Language then jumps to clay tablets, papyrus, codex books, etc. The printing press allows it to escape containment to a wider public than just a caste of priests and bureaucrats. As soon as we invent automatic calculators, we start networking them and using those to process language, albeit in a primitive way (email, the web, etc.). Recently we discovered some abstruse math that, with the assistance of a bunch of beefy video cards, can crunch centuries of human writing into a mathematical object that encodes at least some of the meaning of that writing into an even more "advanced" symbolic processing machine. There's a clear trajectory of language itself getting more and more free of the specific wetware it grew up on.

It's a falsifiable claim, in that if there is a way to train a useful LLM from scratch without any human authored input language to bootstrap it (something I've been on the lookout for but haven't seen, though admittedly I'm not an AI researcher, just some Linux nerd with a day job as an SRE), then we can disprove it.

For the religious angle, look no further than John 1:

"In the beginning was the Word, and the Word was with God, and the Word was God."

(This is admittedly less falsifiable!)

[HarHarVeryFunny]

Well, humans developed language. Language is just a tool that let's us leverage our innate intelligence.

I'm sure that we will eventually build artificial brains, capable of bootstrapping communications and language for themelves (if run en-masse in a simulation where the benefit of communication would emerge). An LLM can't do this since it is by definition/construction something only capable of learning a pre-existing language.

An artificial brain, just like a wet jiggly one, is always going to be more intelligent than a one-trick pony like an LLM - a language processor, but it is notable how intelligent that one-trick pony nonetheless appears to be.

[ElevenLathe]

I think it's interesting that you think we could bootstrap an artificial brain with no inputs from human culture. I disagree, but am open to an existence proof of this kind. Such an artificial brain would be totally alien to us, of course. I wonder how differently it would perform versus something more grounded in "real" culture and writing?

[HarHarVeryFunny]

I'd say that human babies and LLMs are both existence proof that prediction and prediction error feedback is all you need to learn. The artificial brain/baby would be designed to learn just like us by prediction, and should therefore be capable of learning language from scratch just as we do.

You could choose to lock it in a virtual or physical basement with printout of the Common Crawl dataset and raise it like an LLM that learns language with zero real world grounding, in which case it may feel a bit like an LLM (but smarter - able to learn, etc), or you could let it interact with the real world and learn everything, including language, that way and be a lot more grounded.

It's hard to guess how the grounded version would feel to talk to vs the CommonCrawl one - I think it would mostly come down to how far we wanted to go in making the artificial brain have all the moving parts of a human one. In an odd way the more human we tried to make it, the more alien it might feel, but not for the "uncanny valley" reason you might imagine...

The thing with an LLM that makes it feel so human is that they are designed to 100% copy humans - their output is 100% driven by the training goal of trying to exactly match the samples in the training set. As soon as we start to try to build something more brain-like then it's behavior is going to be a lot less predictable - not just auto-regressive "auto-pilot speech", but driven by it's own internal thoughts, emotions, innate traits, etc - depending on how much of our brain we tried to copy in the artificial one. I expect it would feel a lot more organic, less robotic, to talk to, but at the same time perhaps less human since unlike the LLM it's not built just to mimic human speech.

[adjejmxbdjdn]

I think you’re right but I wanted to note that we are discovering that other animals also have language.

[ElevenLathe]

Yes, this to me is also a good sign for the "it's language that's smart, not us" argument. It's an emergent trait that has evolved several times, like flight or carcinization. There's something about language that attracts evolution toward it. One would expect such a trait to have a big survival value (disclaimer: IANA biologist, philosopher, theologian, mathematician, or linguist).

[roenxi]

> I think it’s legit, actual intelligence, not just “artificial” intelligence. That may be a hot take but it’s just my perception.

You might be redefining words here; there isn't a form of intelligence that isn't actual intelligence. It is all actual intelligence. Artificial in this context means it is something we're creating in a lab. LLMs can't avoid being artificial intelligence. The meaning of "AI" is to artificially create actual intelligence.

[PunchyHamster]

average person is absolutely awful judge on anything you put in front of average person tho.

And if anything, average AI user is vastly overstating how good/useful it is. Papers about it pretty much always show huge gap between "productivity person thinks they are achieving" and "actual growth of productivity"