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

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

#61

Post by cscurmudgeon » Thu, Jun 10, 2021, 7:16 AM UTC

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…

There is another set of people. There are people with solid track records in AI and ML that disagree with DeepMind.

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

#62

Post by Clewza313 » Thu, Jun 10, 2021, 7:17 AM UTC

londons_explore wrote: post #58

ms1 wrote: post #49

Earlier quoted context omitted.

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.

Except that manmade solar cells are pretty bad at repairing or replicating themselves.

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

#63

Post by anyfoo » Thu, Jun 10, 2021, 7:17 AM 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,…

> but the sum total of what we know about intelligence says it probably isn't

Source? I am not aware of any other known process in the universe that could not be simulated by a Turing machine.

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

#64

Post by rich_sasha » Thu, Jun 10, 2021, 7:17 AM UTC

lvncelot wrote: post #56

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.

Is Deep Mind a “booming business”? They are achieving great things academically, but their business successes are either kept secret or mostly absent. All I know about is the Google data centre cooling scheduling, probably a big saving for Google but hardly an achievement that in its own professes their business success.

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

#65

Post by twhitmore » Thu, Jun 10, 2021, 7:18 AM UTC

While I am agnostic as to the advantages of RL over simpler ML, some basic structural concerns seem to apply here. The potential concern is that if RL works, the very best factors for learning are likely to be potentially dangerous.

The best environment for learning is the richest -- ie. unrestricted access to be the entire Internet/ or the world via the Internet.

The best reward function is quite likely to be reproduction. If the researcher allows moderate action but attempts to limit reward functions, at a certain level they may find the AI has found a better reward.

If the claimed advantages to learning of richer environments & richer action capability are even somewhat true, researchers are likely to be strongly incentivized to pursue such risky structures.

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

#66

Post by bildung » Thu, Jun 10, 2021, 7:18 AM UTC

londons_explore wrote: post #58

ms1 wrote: post #49

Earlier quoted context omitted.

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.

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

Only because we cheated, though: Houses can't sponantously grow more cells in place when more energy is needed.

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

#67

Post by sam-2727 » Thu, Jun 10, 2021, 7:18 AM UTC

By far not an expert in this area, but if (when?) we successfully generate true intelligence, I suspect it will be through some sort of "ensemble" model, where multiple agents are trained in parallel and interact with each other. Intelligence as we know it hasn't just resulted from an evolution of one agent in response to a cost function, but rather through the complex interactions of agents (humans and organisms in general) over time. I feel like the underlying journal article (https://www.sciencedirect.com/science/article/pii/S000437022...) is missing discussion of this.

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

#68

Post by wnkrshm » Thu, Jun 10, 2021, 7:19 AM UTC

londons_explore wrote: post #57

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.

energy supply and other constraints (material, robustness ...) are a good explanation though - an organism can't grow out of aluminum or steel

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

#69

Post by andyxor » Thu, Jun 10, 2021, 7:20 AM UTC

eeegnu wrote: post #29

andyxor wrote: post #9

good luck with that. DeepMind should sponsor B. F. Skinner award, to honor the father of their behaviorist theories of 'reward and punishment' as a sort of all-encompassing theory of everything related to cognition. At least now they are torturing GPUs and not some poor lab animals. on a serious note the only positive outcome of all this shameless PR is that the heavy investment in ML/RL might trickle down to actual…

> towards understanding natural intelligence, a prerequisite for creating an artificial one. I've thought about this before, and I'm not convinced it's really prerequisite. Naturally developed intelligence in my mind may actually be highly constrained and inefficient because it was limited to what was biologically feasible. i.e. There may be simpler ways of achieving comparable results. Natural intelligence does howe…

yes, it's a recurring thread, "do we really need to mimic the birds in order to build airplanes", etc.

I think someone serious about AI should treat it not as engineering problem but as a science, like physics, which starts with model of nature, and experiment to prove or disprove the theory. Nature provides the constraints by which theory is developed, which radically limits the "search space" of theories. Otherwise it's a bit like throwing things on the wall and see what sticks, which is the primary method of current AI research.

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

#70

Post by EE84M3i » Thu, Jun 10, 2021, 7:25 AM UTC

wnkrshm wrote: post #68

londons_explore wrote: post #57

Earlier quoted context omitted.

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.

energy supply and other constraints (material, robustness ...) are a good explanation though - an organism can't grow out of aluminum or steel

I suspect they're talking about this (sort of) simulation

https://rednuht.org/genetic_cars_2/