OpenAI Chief Scientist Says Advanced AI May Already Be Conscious | Hacker News Reader

OpenAI Chief Scientist Says Advanced AI May Already Be Conscious

futurism.com

2 pointsCharlesW9 comments

Not sure why everyone is being so negative. Taken charitably the scientist is saying

"We don't currently have a mathematical definition for what consciousness is, but if we do develop one in the future, I would not be totally surprised if we learned that todays large neural networks (I assume here he is talking mostly about the recent gigantic language models) turn out to be conscious in some sense."

Maybe that's overly charitable, but that statement doesn't seem too unreasonable.

Because by muddying the waters about what counts as intelligence and consciousness AI researchers are making it harder to actually discuss these topics in a coherent way. It's a self serving agenda that serves no purpose beyond aggrandizing some very basic mathematics. No one has any coherent explanation for why a bunch of matrix multiplications and very basic non-linearities like ReLU are sufficient for creating conscious agents. It's obvious that "1+1" is not an expression that denotes anything about consciousness and yet AI researchers would have us believe that if we just add more 1s and a few more mathematical operations then we will magically get conscious agents. It's a preposterous claim and as I said previously serves no purpose other than confusing and befuddling the public in order to sell them silly toys.

All the ironic comments let clear that us humans have full biased understanding of what consciousness is: Most probably a talking head attached at a walking capable body.

Spoiler alert: being able to speak isn't the definitive demonstration of self-consciousness, even for humans.

The things could actually be conscious right now and we'll know in maybe 600 years.

Just like what happened with Leonardo's flying machines: yes, we actually could have flied a lot earlier, but most people just could not believe it was even possible.

OpenAI software may be "conscious" in the same way that a flatworm is conscious.

If OpenAI's code is conscious then it's definitely insane which should make people wonder about what exactly they're saying when they say things like "Neural networks are conscious". AI researchers constantly use words without defining them and then wonder why they're confused about what counts as intelligence and, more generally, consciousness.

In fields like AI there really is no room for confused language like that, it's a mathematical field so no one should be confused about function approximators being conscious or not.

"If OpenAI's code is conscious then it's definitely insane.."

As a kid, I spend some time in my aunt's house, close to psychiatric clinic. They have there a lot of "insane" people, most of them not dangerous at all, so, many could walk out, most ended talking with neighbors around the clinic.

This experience opened my mind about that what humans say aloud could be a thousand years aways from rational thinking, if self-consciusness could be scientifically defined I think you could diagnose many people as not self-conscious.

That said, what we currently can read written by NLP models looks for me a lot like the "psychiatric talking" I listened a lot time ago.

I always wondered if that people from the psychiatric clinic was really self-conscious, maybe they were "part-time self-conscious", and sometimes they just did "crazy talk", because they were not in "self-conscious mode".

Now I wonder if the advanced NLP models could have some conscious thing just popping up there in the words they push out of their internal stuff, in the flow of the equations running the calculations. Somewhere, currently out of reach of the state of the art science.

Consciousness is not reducible to symbol shuffling is my point. Computers at the moment can only perform mathematical operations, i.e. they can only shuffle abstract symbols. Neither intelligence nor consciousness can be defined purely in terms of symbol shuffling unless we vacuously define intelligence and consciousness to be abstract symbol shuffling.

Talking seriously, there are conversations about how the human brain works, what could exactly be the "main loop" running the "consciuosness" yet undefined thing (biologically located in our heads, but it seems a bit of "consciusness" could be located right into our intestines, it even could have enough chemical bandwith for regular - close to realtime - comms with its upper buddy in the head). And it's not going well for us - our brains mainly- supposely being "exceptional stuff" among other living things, even not living things (crystals).

Most of the stuff currently available point to, that us humans are just somewhat very simple biological algorithms, possibly not even fully rational, but - just A LOT like neural networks - just able to better "match" - i.e. "a closed door needs to be unlocked, opened, before you can walk through it" - internal models of the reality with the actual reality around us, than other living or non living things.

We could even have hardcoded objectives deep inside the DNA, not a lot different than GPT3 and its fellow advanced descendants. Also not a lot alike neural networks, we could be just talking heads, and the actual calculations, rationality - the algorithm - is deeply buried well below the "conscious" level we experiment everyday. There's a chance that the evolution driven biological algorithm lying in our DNA - having evolved through a couple of million of years - it is just better, just matches the reality just a lot better than our 5 years old algorithms.

Think about, every living thing in the planet has biologically hardcoded instructions, why humans would be the only "special" thing not having pre-defined DNA based set of objectives. i.e. "you" get to pick the wall colors, but you didn't designed the "house".

The extra stuff there, eyes, legs, mouth, language could be just very sophisticated tools, not quite a lot different than feathers, tails, wings, etc.

So, back to our recently constructed structures, the hiper-complex neural networks could have some "patterns" quite a lot alike what we have in our heads/intestines. Those "patterns" one day, maybe in 600 years could technically be defined as "yep, that was consciousness"

With no support whatsoever for that claim.

Reply on news.ycombinator.com