Hacker News Reader

Comment by streb-lo | Hacker News Reader

We are all stochastic parrots. Currently, millions of years of evolutionary pressure have produced a better algorithm than a few years of optimization pressure but we’ll see how long that remains true!

View on HN

You are not just a stochastic parrot. Your speech is not just "next likely word", but a product of your internal model of the world.

Evidence suggests LLMs have internal models as well.

I would first like to ask, are you religious? Do you believe in souls? If the answer to both is "No", then I'd like to discuss further.

The answer to both is "No".

Ok, then let us discuss what you define stochastic parrot as.

Happy to go with the first paragraph from the Wikipedia entry: https://en.wikipedia.org/wiki/Stochastic_parrot

Suppose a "parrot" factors a product of two 2048 bit primes? Is it still an entity without understanding of the processes considering the search space size?

Does it have an internal model of what a prime is? If so then it's not parroting. If it doesn't then it is. The whole point is that either they have internal models of the concepts, from experience of those things, or they don't, and are purely working from ungrounded symbols which have no connection to the actual things they're referencing.

So, examples of "Look, it did this!" mean nothing to me. What matters is their internal systems.

What do you think "experience is"? The human brain and body stores experience, like anything else as physical data in itself. So why do you think an LLM can't have experiential knowledge if it is trained to or somehow naturally gains it during training?

An LLM, as the name implies, only has access to words. It can only model things about those words and the relationships between them. It never has access to the actual things the words represent.

So it absolutely has experiential knowledge of words but not of the actual things.

(You can absolutely train models on other inputs. I have no objections to those not being stochastic parrots)

Yes but I meant that for instance, how is experience stored in the humans? It is encoded as data in the brain and or rest of body. So I think its possible even a pure textual LLM may have imbibed experiential data somehow during its training.

But if your claim is at least that you are fine with thinking AI's that also include physical training are not stochastic parrots, then thats good enough for me.

Evolution selects for fit organisms that have to survive and reproduce in the world, not language models waiting on prompts. The temptation to analogize technology to biology is just top hard to resist I guess. Problem with analogies and metaphors is they are never accurate, they are only similar in some respects.

Yes, thank-you for informing me I was using an analogy.

I think it is worth asking whether anthropomorphization of LLMs does anything to contribute to a better understanding of them that isn't entirely post-hoc ("of course they were going to find CVEs, solve ARC-AGI...") or if it just ends up reinforcing simplified teleological views of neuroscience, evolution or psychology etc.

Sure, there are potential risks to anthropomorphizing language. There's also potential benefits, in terms of allowing us to talk about an emerging alien intelligence.

FWIIW, I don't think I was anthropomorphizing. Stretching an analogy, yes. Optimization pressure is producing something with at least characteristics of intelligence. Obviously in a very different way than evolution produced the human brain, but the similarities are there.

Reply on news.ycombinator.com