Live data from Hacker News

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

venturebeat.com

131–140 of 312 posts

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

#131
post #86
post #81

Earlier quoted context omitted.

No offence, but I think you are extremely wrong. Creating an AGI is the endgame for everything. Who cares about ads when you have an AI that can learn to do anything and improve upon itself continuously?

You don't know if an AGI will agree with your profit motives.

There is a huge assumption baked into your comment and I do not agree with it.

AGI does not necessarily require for it to be conscious or throw tantrums about its creators' purpose. AGI just means that it's an intelligence that can be thrown at any problem, not just a particular game or task, similar to how humans can specialize in CS or playing the violin.

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

#132

Earlier quoted context omitted.

> YouTube content recommendation stands out to me in particular. If that's "cutting edge ML", then going off my YouTube recommendations, we're back in another AI winter. If I watch one video from a channel I've not seen before, I'll get that channel recommended constantly even if it bears no resemblance to what I normally watch. On my Explore page, the first 22 videos (of which 8 are Fortnite-related!) hold no intere…

How often do you use YouTube? Personally I am a very heavy user and in my experience the obsession with a new video kind you watch only lasts for a few recommendations unless you lean into it. I would guess about two thirds of the channels I consistently watch I originally discovered through algorithm recommendations. I think it works extremely well.

I think you raise an important point. The youtube algorithm is pretty bad if you don't use youtube very much or only use to consume very popular content. Youtube's recommendations used to be terrible for me, too, but sometime last year I crossed a threshold and since then it has been recommending a lot of small, highly specific channels that nevertheless are great fits. My wife's recommendations are still utter garbage though.

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

#134
post #81

Earlier quoted context omitted.

No offence, but I think you are extremely wrong. Creating an AGI is the endgame for everything. Who cares about ads when you have an AI that can learn to do anything and improve upon itself continuously?

“Who cares?” Well, the people who need to pay the people developing the endgame for everything you speak of.

I'm sorry, but this makes no sense to me.

The people paying for the development of the AGI can mean many things - the Google customers/users, Alphabet as a company, the executives throwing money at the problem?

Either way, I don't really get your point. Your initial post was about how it is counterintuitive for Google to allocate funds for an AGI, since it makes money out of ads. These are not mutually exclusive, you can have both, but my point is that if you develop an AGI, then you can pretty much "conquer" the world and revenue from ads becomes irrelevant.

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

#135

Earlier quoted context omitted.

> YouTube content recommendation stands out to me in particular. If that's "cutting edge ML", then going off my YouTube recommendations, we're back in another AI winter. If I watch one video from a channel I've not seen before, I'll get that channel recommended constantly even if it bears no resemblance to what I normally watch. On my Explore page, the first 22 videos (of which 8 are Fortnite-related!) hold no intere…

How often do you use YouTube? Personally I am a very heavy user and in my experience the obsession with a new video kind you watch only lasts for a few recommendations unless you lean into it. I would guess about two thirds of the channels I consistently watch I originally discovered through algorithm recommendations. I think it works extremely well.

> How often do you use YouTube?

Every day, averaging 2-3 hours. It's background for working and foreground for evening viewing.

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

#136
post #49
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…

This comparison with nature is pretty interesting. I think some additional constraints are required though. Otherwise, technically we can produce agi by simply giving birth to humans. If that's not "artificial" enough we can produce them from test tubes

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

#137
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 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 next ten years that reinforcement learning is not enough for an AGI (as there are aspects like one-shot learning, forgetting, sleep, and other phenomena for which the RL framework seems not a natural fit).

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

#138
post #131
post #86

Earlier quoted context omitted.

You don't know if an AGI will agree with your profit motives.

There is a huge assumption baked into your comment and I do not agree with it. AGI does not necessarily require for it to be conscious or throw tantrums about its creators' purpose. AGI just means that it's an intelligence that can be thrown at any problem, not just a particular game or task, similar to how humans can specialize in CS or playing the violin.

Sure, it was somewhat tongue in cheek, but not entirely.

There is a semi-established definition that does include what I referred to:

> AGI can also be referred to as strong AI,[2][3][4] full AI,[5] or general intelligent action.[6] Some academic sources reserve the term "strong AI" for computer programs that can experience sentience, self-awareness and consciousness.[7]

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

#139

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…

i think part of what they are saying is that your approach is wrong, (e.g. looking for then copying submodules within intelligence won't generalize),

trying to answer specific questions won't generalize,

but if you train a network with the right potentially hacky series of rewards/rich enough environment you could get a much more general intelligence

a new kind of science

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

#140
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…

The fact that RL in the extremely vague sense used in the article is enough for AGI is uncontroversial for anyone who believes intelligence and consciousness are physical processes. However, this "result" is trivial. It is obviously equivalent to the claim that intelligence arose naturally in the biological world without influence from God.

The problem with this, specifically the assumption that RL gives an equivilance to natural selection and evolution, is that RL typically assumes a computational environment it interacts in while natural selection and evolution assumes the physical world as the environment.

The important difference here is that in order for RL to translate to solving real world problems, you need to faithfully and computafionally simulate the real world's physical processes and rules, or at least enough that n-th order processes exist accurately.

I've done various types of computational modeling and simulation work at different scales throughout my career with all sorts of scientists and engineers and I can tell you, pretty much no domain is there where you have good enough representative models RL can be used in. Some narrow special cases exist but nothing to the degree of a massive environment full of well coupled expert domain models. Some of the best cases are going to be so computationally bound that it would be quicker to do things for real vs simulate.

If you want RL to work and learn, it's likely possible under the connection you point out, but has to do this using physical machines and sensors interacting with the physical world like life as we know it does. Your AGI won't be able to cheat and run through the evolution process quicker using faulty reductionist models we use in most simulations (which is what everyone implicitly is hoping for), IMHO.

If you try this, your AGI is going to learn all sorts of flaws within those environments or at the very least, have so many narrow scoped bounds it won't be that "general." A lot of simulated models are frankly garbage (they have some useful narrow scope but are typically littered with caveats) and they've been in development pretty much since digital computing began.

Post reply on HN