DeepMind says reinforcement learning is ‘enough’ to reach general AI • Hacker News Board

Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI

#301

Post by dr01d » Fri, Jun 11, 2021, 9:01 PM UTC

endtime wrote: post #192

burning_hamster wrote: post #137

Earlier quoted context omitted.

I think the title of the paper makes more sense if you consider that ten years ago, someone could have written a paper in a similar spirit with a different take on "what is enough". Back then, it would probably have been titled: "Backpropagation of errors is enough". The last ten years have shown that backpropagation -- while a crucial component -- is not enough. Personally, I would not be shocked to find out in the…

Why are forgetting and sleep relevant? If someone invented a pill that gave you a perfect memory and removed the need to sleep, would you stop being generally intelligent if you took it?

Probably because if we didn't have the ability to forget and turn off often we would pretty quickly kill ourselves to end the horror of our existence?

Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI

#302

Post by webmaven » Sat, Jun 12, 2021, 1:36 AM UTC

visarga wrote: post #293

webmaven wrote: post #34

Earlier quoted context omitted.

> RL + piggybacking on human culture might be enough, or evolution + RL for biological agents. Yes, but over what timeframe? Will there be any diminishing returns plateaus along the way?

We still have unknown unknowns but we also know a lot more about how neural nets deal with various tasks and dataset preparations. We know what kind of applications are good enough and where they still fail, which is much more than a decade ago. If you look at sci-fi movies with robots, they usually speak in a metallic voice but have good situational and language understanding. In reality it was the other way around,…

> If you look at sci-fi movies with robots, they usually speak in a metallic voice but have good situational and language understanding. In reality it was the other way around, it's much easier to do artificial voices than understand the topic. That kind of naive understanding seems silly now, and this is how we gradually advance.

I waffle a lot on whether that aspect of 1968's '2001: A Space Odyssey' is evidence of genius or just survivorship bias.

Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI

#303

Post by webmaven » Sat, Jun 12, 2021, 2:44 AM UTC

robrenaud wrote: post #225

xamuel wrote: post #213

Earlier quoted context omitted.

I thought it was a fun position paper, if not exactly groundbreaking. They did avoid one common pitfall at least. They are (intentionally?) vague about which number systems the rewards can come from, apparently leaving it open whether the rewards need be real-valued or whether they can be, say, hyperreals, surreals, computable ordinals, etc. This avoids a trap I've written about elsewhere [1]: traditionally, RL rewar…

There are more real numbers than programs. Computers cannot represent the vast majority of real numbers. AFAICT, it's not even clear that the universe is continuous rather than discrete. I really don't believe that using approximations of real numbers is going to be the bottleneck for AGI.

> AFAICT, it's not even clear that the universe is continuous rather than discrete.

I'm not sure that makes any difference (in either direction).

I mean, at the scale we care most about, the universe appears to be continuous, so an AGI has to be able to tackle continuous-appearing problems and use continuous-appearing representations.

OTOH, the universe is likely to actually be discrete, so an AGI has to be able to tackle actually-discrete problems, and use representations that are actually-discrete on a fundamental level.

There isn't much of a contradiction between these constraints, although the prospect of a continuous-appearing universe that is actually running on a discrete substrate seems to give a lot of people a brain cramp, and that same brain cramp gets elevated into 'proof' that current approaches cannot lead to AGI. Which is nonsense (there may be other limitations inherent in current approaches, but that can't be one of them).

One might as well claim that computers are digital and brains are analog and conclude that digital image representations cannot possibly be used to communicate information to analog brains.

Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI

#304

Post by webmaven » Sat, Jun 12, 2021, 3:26 AM UTC

rubatuga wrote: post #210

webmaven wrote: post #40

Earlier quoted context omitted.

> Afaik some biologists even think that the evolution of intelligence was rather unlikely and would not necessarily happen again under the same circumstances as on earth. Hmm. Can you provide a pointer to those biologists? AFAIK, high intelligence has arisen more than once on Earth (Hominoids, Cetaceans, Octopuses), so I'm somewhat skeptical of that claim, but perhaps they're construing intelligence more narrowly (ie…

Well that has a prior on life even existing in the first place

> Well that has a prior on life even existing in the first place

True, the question of intelligent life evolving can be construed as either:

"Given that life exists, what is the probability of intelligence evolving?"

Or:

"Given that the universe exists, what is the probability of life arising and evolving intelligence?"

Both are actually interesting and important questions (cf. the Drake Equation and Fermi Paradox), but I am pretty comfortable asserting that in the context of this conversation the former interpretation is more apropos.

Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI

#305

Post by seaorg » Sat, Jun 12, 2021, 8:33 AM UTC

helen___keller wrote: post #299

seaorg wrote: post #250

Earlier quoted context omitted.

The entire sandbox thing is foolish. Obviously it won’t be sandboxed anyway because of what you’ve pointed out. But any sandbox we design won’t be good enough. GTP is not supposed to exist according to computer science and machine learning experts in 2017. They were all wrong. Same with deep dream, styleGAN. And it will be true with the next thing. The mental gymnastics refers to the fact that people have cognitive d…

> But any sandbox we design won’t be good enough. This isn't really true. To bring back your analogy of searching the space of all programs - suppose we wanted to simulate every turing machine of at most N states for some sufficiently large N. One of these turing machines is going to encode an AGI. Nonetheless, it is still just an encoded turing machine being simulated by a turing machine simulator. No matter what gr…

Yes, programs are Turing machines that use memory. I’m a little insulted. No method of containment is good enough for something like that. But we agree that sandboxing as a basis to claim that AGI won’t be a danger is totally unsound.

Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI

#306

Post by samatoid » Sat, Jun 12, 2021, 1:57 PM UTC

ABCLAW wrote: post #220

endtime wrote: post #192

Earlier quoted context omitted.

Why are forgetting and sleep relevant? If someone invented a pill that gave you a perfect memory and removed the need to sleep, would you stop being generally intelligent if you took it?

One of the postulated reasons for why older people have worse reaction times and think slower than their younger counterparts is that the neural networks they use draw upon more stored information, thus making routine evaluations take longer. There's a sweet spot between knowing enough and knowing little enough so that you get the right answer and get it quickly enough.

I guess you haven't heard of ISRIB (integrated stress response inhibitor). This drug when given to mice causes both old and alzheimer's riddled mouse brains to return to near normal. Interestingly the patent is now held by an ALPHABET company. Not saying there isn't a limit to our storage capacity, but this drug make it clear that age alone doesn't prevent the brain from working well.

Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI

#307

Post by ramraj07 » Mon, Jun 14, 2021, 12:17 PM UTC

otabdeveloper4 wrote: post #52

ramraj07 wrote: post #5

Wait isn’t any Turing complete programming language sufficient to eventually reach general AI

You're assuming that intelligence is a computational process, but the sum total of what we know about intelligence says it probably isn't. (Unless you're making a more general reductionist statement that everything in the universe is a computational process - that kind of reductionism is understandable coming from people who work with computers for their job - but this is then a philosophical stance, not scientific,…

I’m not making the statement that everything in the universe is a computational process, only that everything in my brain is, at least the functionality of my brain towards creating my consciousness.

Now can I prove that ? Of course not. But it seems like a fairly solid working hypothesis (any other alternative hypothesis sounds far more quacky anyway? What quantum entanglement of mocrotubules?).

Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI

#308

Post by sbierwagen » Mon, Jun 14, 2021, 10:35 PM UTC

randcraw wrote: post #267

Animats wrote: post #48

"A sufficiently powerful and general reinforcement learning agent may ultimately give rise to intelligence and its associated abilities. ... We do not offer any theoretical guarantee on the sample efficiency of reinforcement learning agents." OK. This basically says "evolution works". But how fast? Biology took tens of millions of years to boot up. An related question is how much compute power does evolution, viewed…

> "why does robotic manipulation in unstructured situations still suck after 50 years" Great point. Until the promoters of RL can build us a robot that can 1) walk gracefully through a typical home that has stairs and closed doors, 2) cook a meal with pots and pans, and 3) get back up after it falls down -- I suggest we take their claims of impending Singularity with a big grain of salt.

Consider if Boston Dynamics released a video of a robot doing all three things tomorrow, and announced that they were taking orders for immediate shipment, priced at $100,000. How far away would you think the singularity would be then?

Separately, a hostile or indifferent AI could still cause a heck of a lot of trouble for human civilization without the first two things. Consider an autofactory clearing room for expansion with bulldozers, no need to navigate stairs there. Bullets or smart glide bombs don't need to understand doorknobs. Etc.

Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI

#309

Post by topologie » Tue, Jun 15, 2021, 2:04 PM UTC

endtime wrote: post #192

burning_hamster wrote: post #137

Earlier quoted context omitted.

I think the title of the paper makes more sense if you consider that ten years ago, someone could have written a paper in a similar spirit with a different take on "what is enough". Back then, it would probably have been titled: "Backpropagation of errors is enough". The last ten years have shown that backpropagation -- while a crucial component -- is not enough. Personally, I would not be shocked to find out in the…

Why are forgetting and sleep relevant? If someone invented a pill that gave you a perfect memory and removed the need to sleep, would you stop being generally intelligent if you took it?

Maybe not, Chesterton's Fence applies: https://wiki.lesswrong.com/wiki/Chesterton%27s_Fence

Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI

#310

Post by exporectomy » Tue, Jun 15, 2021, 11:47 PM UTC

6gvONxR4sf7o wrote: post #199

exporectomy wrote: post #129

Earlier quoted context omitted.

It can prove its intelligence by making testable predictions of the future better than us. As for whether it's "real" AGI or just acts like it, doesn't really matter. I think the Chinese room problem has been agreed on as not a problem, hasn't it?

I think proof of “real” intelligence by answering harder and harder questions is barking up the wrong tree. I think evidence and proof are a better way to denote varying levels of understanding. A deductive system can come with an answer and a proof of that answer, where proof is whatever counts as proof in that system. So the notion of “does it really understand it’s answers” gets punted off its Q&A abilities and on…

That's an interesting idea but it would exclude many humans who can make correct predictions using their experience and intuition but can't justify them correctly. Those people are still very useful.

What you're describing is what we do at school. We can't assess understanding so we assess justification of answers as well as other things like ability to do X (we don't care if they understood or not, just be capable).