Hacker News Reader

Comment by aatd86 | Hacker News Reader

The difference is not really between speaking and thinking but processing data with and without experience.

An AI doesn't have experience so doesn't encode things such as empathy very well.

Because it doesn't infer rules based on experience from interacting with other things. (doesn't know what pain feels like so cannot react and decide about some things on its own for instance)

It's merely a data processor with statically coded rules. It doesn't handle the unknown very well.

Still quite impressive what can be done but there is of course some room for improvement.

View on HN

On the topic of "still quite impressive what can be done", I can't more emphatically recommend (the now quite old!) lecture that Peter Norvig gave which is available on YouTube called the unreasonable effectiveness of data. He does a great job showing some toy and small examples about just how nuanced this point is, and that it is easy to underestimate the impact that simply more data can have. We tend to think of it as a quantitative impact and make all sorts of assumptions as to the second derivative of what the gains can be, but he does a great job showing that there might be all sorts of interesting inflection points and that simply more data can actually have significant qualitative differences.

I think I'd argue that AI does have experience. That's exactly what training an AI gives it. It is however true that that experience is currently less wide ranging than that of humans.

That's not really self experience though.

The same way, I can read about historical events but I probably haven't experienced these events.

So the AI might not be able to realize/derive new causations, mostly able to find out correlations at best. That's what may make it "hallucinate".

AIs can't experience everything, but that doesn't mean they can't experience anything. You can have the experience of reading a book, no? And that might well teach you something about the world.

Of note to me is that there is no theoretical reason why an AI couldn't have experiences of the real world. It's just that currently nobody has built interfaces that allow them to have that connection yet (for best learning you'd want sensors that allow it to sense the world and actuators that allow it to influnce the world - although you quickly get into considerations about whether it's a good idea to give AI's this power)

> mostly able to find out correlations at best

It may well be that human learning is just sophisticated iterated correlation finding.

Yes, the issue is that they don't actually pick up the book. Their experience of the real world is too limited to have valid inferences.

The same way a human being born without any sense but hearing wouldn't be able to interact fully with the physical world and infer things about it.

Basically, what I mean is that for now, unsupervised data acquisition is limited for AI. So it lacks "common sense" if you will.

I don't think we really disagree anyway, that's probably just semantics.

I can watch a war movie and get into the mindset of being shot at even if I have zero experience of it. I guess those emotions are connected in some way to the experiences I had myself. But most of them are inferred from the expressions and words in the movie.

You could create signal of those emotions in the NN, it could predict them accurately. You could create a memory/association of emotions with subjects, some persistence in time.

Then add a simulation/dream of the world (like our consciousness)

At some point I don’t see much missing.

Show me an interaction between two dolphins and I'll be very unlikely to simulate what they feel/experience.

For an AI, that would probably be the same except we are the dolphins in that case.

Besides, as you hint, you still have to have had an experience of feeling pain before and have a mechanism to avoid it. Which means, having self-preservation encoded within your brain processi. And if you don't get PTSD from the movie, you won't know what PTSD feels like (let's imagine this is the only way one experiences stress). There is a need for more correlation processing between sources of data that would be similar to what a living-being may capture.

We're not there yet.

Yet. GPT-3 and their ilk, for all their novelty, are just the infancy of a new wave of AI tools.

Right now ChatGPT is technically learning from experience, although that learning is delayed, as it comes from the OpenAI team retraining it with all its new conversations. They could let it learn freely, but we've seen where that ends. Maybe once their model is perfected we will see it learning in real time.

Reply on news.ycombinator.com