Hacker News Reader

Comment by gibsonf1 | Hacker News Reader

How could a statistical system trained on data display intelligence let alone innovation? It's just very good statistics at the end of the day. Artificial Perception is about as far as you can get with ml/dl tech if you point the sensors at the world of space-time. If you point it at words, as with LLM, you get statistics about words - that is, no actual understanding of what the words model in the minds of the original authors of those words. So utterly unreliable for any kind of mission critical or autonomous use.

View on HN

I'd argue that intelligence is precisely the ability to predict (and therefore also plan) outcomes based on prior experience. The ability to perceive and predict what happens next; the ability to predict/plan/act and be mostly correct about what you predicted would happen as a result of your actions.

With this definition, LLMs in their ability to predict can reasonably be considered to display some basic form of intelligence, even if only in their own world of words rather than the world at large. If we build embodied robots with a similar ability to predict, and also the ability to continually update their predictions based on prediction success/failure (the closed loop that is missing from LLMs), then we'll have something much more recognizably close to animal intelligence.

LLMs aren't statistical systems in any substantive sense. They are deterministic programs over the input sequence. They capture statistical relationships about words, but so do human minds. That they are sensitive to statistical relationships does not discount their ability to understand.

> That they are sensitive to statistical relationships does not discount their ability to understand.

The claim is that nothing but sensitivity to statistical relationships somehow leads to an ability to understand?

I am not going to believe it.

How do you think our brains work?

I don't know how our brains work.

If you don't just want to be snarky: I think the difference between acquisition and learning is important, I think the role of instruction might be important, I think the role of practice is important.

If the stream of data that is our 5 senses of reality is devoid of statistical significance of any kind, how do you propose our brains can learn anything?

Of course, brains as we know it (or complex systems of any sort for that matter) would not exist in such a reality.

In other words it is first and foremost the presence of complex statistical relationships inherent to data/reality that facilitates emergence of other complex systems. Such as the notion of "intelligence" and brains.

We need statistical significance.

is not the same as

Statistical significance is enough.

These are two different things.

In case you are interested, I wrote a piece in defense of understanding in LLMs[1]. The section under the heading "Why think LLMs have understanding at all" is the most relevant.

[1] https://www.reddit.com/r/naturalism/comments/1236vzf

If you know everything, you can connect more dots, and come up with innovative stuff.

Also I’ve just gave instructions to gpt4 rot13 encoded, and it followed the instructions. I don’t really care if “it” understood what it was doing, but the responses were good enough for me to be impressed.

Also I’ve got quite a lot of use cases where I get reliable value out of it, what were you trying to do that made you conclude it is “utterly unreliable”?

As a counter point, I cannot be sure that the words you are speaking mean the same thing to you in your mind as they do to me when I hear you.

Reply on news.ycombinator.com