Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

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

venturebeat.com

Post reply on HN
Print view View original

51–60 of 312 posts

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

#51

Post by 0xB31B1B »

maiodude wrote: post #43
hervature wrote: post #6

I guess things are slowing down at DeepMind. I have tremendous respect for David Silver and his work on AlphaZero and Richard Sutton as a pioneer in RL. But the cynic in me is that this paper is just a result of Goodhart's law with publishing count as a metric. Any proof of the type of emergent behaviors that they mention from RL with an actual RL experiment would go a long way. Showing an RL agent developing a langu…

They just "solved" protein folding late last year. How can you say things are slowing down? Do you honestly expect life-changing discoveries every other week?

Protein folding is a well modeled math problem. The alpha fold solution is extremely good at pruning (aka guessing) folding chain structure possibilities. I am Impressed and this is a difficult problem but this is extremely different from AGI as this is a well scoped easily modelable problem that is basically a chain of 26 inputs types of links of arbitrary length. I am not trying to take away that the protein folding is incredible but AGI is extremely different. AGI is literally having a model that can both do alpha fold and self driving cars, as well as the ability to generate novel models to solve new well scopes problems. RL can do 0 to 1, the 1 to n (generalizeability) is the extremely difficult part.

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

#52

Post by otabdeveloper4 »

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, and frankly a very strange one.)

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

#53

Post by unishark »

jjjdjjddddsfsd wrote: post #25

Are the just reformulating the principles of evolution in digital terms, and essentially not providing any new insights at all? Yes, intelligence has been created by evolution. That doesn't imply that any system that is subject to evolutionary forces will lead to the creation of intelligence (and not within a reasonable timeframe, either). The challenge is to create a system that is capable of evolving intelligence.…

I'd say it's even less than that. They seem to be summarizing the ways the problem of teaching an agent to do anything (including be generally intelligent) can be formulated as a problem of maximizing a reward (hence the title).

Another way to look at it is, if we had a good enough function (e.g. a universal approximator) it can be made to model any behavior using numerical optimization. Which I think isn't very surprising, but apparently there is some arguments about it.

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

#54

Post by uyt »

hervature wrote: post #6

I guess things are slowing down at DeepMind. I have tremendous respect for David Silver and his work on AlphaZero and Richard Sutton as a pioneer in RL. But the cynic in me is that this paper is just a result of Goodhart's law with publishing count as a metric. Any proof of the type of emergent behaviors that they mention from RL with an actual RL experiment would go a long way. Showing an RL agent developing a langu…

It might be a stretch but some people say that the weights learned by a neural network is somewhat like a language. For example if you look at the weights of a random middle layer it would seem like gibberish. Much like how aliens would react when looking at humans making gibberish noises (aka talking) to each other. In both cases they are just compressing signals based on learned primitives.

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

#55

Post by Barrin92 »

Sorry but where is actual scientific content in that paper? I'm concerned with the state of AI. saying that "reinforcement is all you need", when reinforcement learning is defined as abstract as "agent does something, adapts to environment and rewards, then does another thing" is borderline tautological.

The actual scientific question is, what are the mechanisms that make agents work, what are the fundamental modules within intelligent systems, is there a distinction between digital and biochemical systems, what costs are there in terms of resources and energy to get to a certain level of intelligence, and so on. Real questions with specific answers. For all the advances coming from just upping the amount of data and GPU hours, there is so little progress on trying to have a model of the structures that underpin intelligence.

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

#56

Post by lvncelot »

harry8 wrote: post #18

Some Bozo who has heard all this many times before is suspicious of claims from places like Deep Mind who have a financial incentive to make them (keep funding) where there aren't working machines to back that claim up. Some Bozo has no credentials, no reputation, no track record of publications and barely supports the claim they're making with anything much. Some Bozo has no financial incentives or otherwise to opin…

Some cynic remarks that during the first AI golden years, claims of imminent success seemed to come from a place of hopeful naïveté of a fledgling science, whereas those same claims nowadays seem to come from a place of cold calculation of a booming business.

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

#57

Post by londons_explore »

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…

In some cases biological 'genetic algorithm' hill climbing can be remarkably ineffective.

For example, the classic "design a car that can drive over this terrain" problem, even after a billion generations (~ the same number as life on earth), shows no substantial performance improvement.

That makes me suspect something is missing from our biological genetics model.

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

#58

Post by londons_explore »

ms1 wrote: post #49
webmaven wrote: post #3

Saying RL is sufficient to (eventually) achieve AGI is a bit misleading. One might similarly state that biological evolution is sufficient to (eventually) achieve biological general intelligence. Both statements are probably true, but the parenthetical (eventually) is doing an awful lot of heavy lifting.

I think, in really broad terms, in order to get AGI actually we would need to do better than nature. If our metric is (intelligence)/(joule), nature seems pretty bad at a first glance: it took many trillions of lifetimes to achieve "general intelligence" * But then again, on the big stuff like this, have we ever really beat nature? That asterisk is there because, sure, turning the earth's biosphere into computers wou…

Nature has a massive incentive to make good use of energy from light through photosynthesis. Billions of plants compete, and whoever can get most out of the sun will win out.

Yet manmade solar cells are more efficient by nearly all measures.

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

#59

Post by littlestymaar »

Being “enough” isn't practical though and whether it's sufficiently efficient is still an open question.

We already know that evolutionary trial and error is “enough” to create flying animals, yet with a little bit of domain knowledge we've been able to build planes way faster than nature has built flying insects.

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

#60

Post by 29athrowaway »

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…

We just want AI to be able to think. We do not need an AI with an autonomic nervous system, or many of the functions in the central nervous system. We do not need AI to be very power efficient. If it takes several megawatts of electricity to get our first strong AI working, so be it.

So, we do not have as many constrains as life did.

Post reply on HN