Less Wrong is a community blog devoted to refining the art of human rationality. Please visit our About page for more information.

Lumifer comments on The Critical Rationalist View on Artificial Intelligence - Less Wrong Discussion

0 Post author: Fallibilist 06 December 2017 05:26PM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Comments (59)

You are viewing a single comment's thread.

Comment author: Lumifer 07 December 2017 08:49:49PM *  1 point [-]

AlphaGo is a remarkable algorithm, but it cannot create knowledge

Funny you should mention that. AlphaGo has a successor, AlphaZero. Let me quote:

The game of chess is the most widely-studied domain in the history of artificial intelligence. The strongest programs are based on a combination of sophisticated search techniques, domain-specific adaptations, and handcrafted evaluation functions that have been refined by human experts over several decades. In contrast, the AlphaGo Zero program recently achieved superhuman performance in the game of Go, by tabula rasa reinforcement learning from games of self-play. In this paper, we generalise this approach into a single AlphaZero algorithm that can achieve, tabula rasa, superhuman performance in many challenging domains. Starting from random play, and given no domain knowledge except the game rules, AlphaZero achieved within 24 hours a superhuman level of play in the games of chess and shogi (Japanese chess) as well as Go, and convincingly defeated a world-champion program in each case.

Note: "given no domain knowledge except the game rules"

Comment author: curi 08 December 2017 05:17:19AM *  0 points [-]

They chose a limited domain and then designed and used an algorithm that works in that domain – which constitutes domain knowledge. The paper's claim is blatantly false; you are gullible and appealing to authority.

Comment author: Lumifer 08 December 2017 06:02:35AM 0 points [-]

<shrug> You sound less and less reasonable with every comment.

It doesn't look like you conversion attempts are working well. Why do you think this is so?

Comment author: Fallibilist 08 December 2017 09:54:23AM 0 points [-]

Unreason is accepting the claims of a paper at face value, appealing to its authority, and, then, when this is pointed out to you, claiming the other party is unreasonable.

I was aware of AlphaGo Zero before I posted -- check out my link. Note that it can't even learn the rules of the game. Humans can. They can learn the rules of all kinds of games. They have a game-rule learning universality. That AlphaGo Zero can't learn the rules of one game is indicative of how much domain knowledge the developers actually put into it. They are fooling themselves if they think AlphaGo Zero has superhuman learning ability and to be progress towards AI.

Comment author: Lumifer 08 December 2017 03:55:23PM *  0 points [-]

Unreason is accepting the claims of a paper at face value, appealing to its authority

Which particular claim that the paper makes I accepted at face value and which you think is false? Be specific.

I was aware of AlphaGo Zero before I posted -- check out my link

AlphaGo Zero and AlphaZero are different things -- check out my link.

In any case, are you making the claim that if a neural net were able to figure out the rules of the game by examining a few million games, you would accept that it's a universal knowledge creator?

Comment author: curi 08 December 2017 11:02:59AM *  0 points [-]

If they wanna convince anyone it isn't using domain-specific knowledge created by the programmers, why don't they demonstrate it in the straightforward way? Show results in 3 separate domains. But they can't.

If it really has nothing domain specific, why can't it work with ANY domain?

Comment author: Lumifer 08 December 2017 03:56:12PM *  0 points [-]

Show results in 3 separate domains.

  • Chess
  • Go
  • Shogi
Comment author: HungryHobo 08 December 2017 03:07:06PM *  0 points [-]

You're describing what's known as General game playing.

you program an AI which will play a set of games, you don't know what the rules of the games will be. Build an AI which can accept a set of rules for a game then teach itself to play.

This is in fact a field in AI.

also note recent news that AlphaGoZero has been converted to AlphaZero which can handle other games and rapidly taught itself how to play Chess,Shogi, and Go (beating it's ancestor AlphaGoZero) hinting that they're generalising it very successfully.

Comment author: curi 08 December 2017 09:46:34PM 1 point [-]

Here are some examples of domains other than game playing: architecture, chemistry, cancer research, website design, cryonics research, astrophysics, poetry, painting, political campaign running, dog toy design, knitting.

The fact that the self-play method works well for chess but not poetry is domain knowledge the programmers had, not something alphazero figured out for itself.

Comment author: HungryHobo 08 December 2017 10:04:11PM 0 points [-]

This again feels like one of those things that creeps the second anyone points you to examples.

If someone points to an AI that can generate scientific hypothesis, design novel experiments to attempt to falsify them and run those experiments in ways that could be applied to chemistry, cancer research and cryonics you'd just declare that those weren't different enough domains because they're all science and then demand that it also be able to control pianist robots and scuba dive and run a nail salon.

Nothing to see here everyone.

This is just yet another boring iteration of the forever shifting goalposts of AI .

Comment author: entirelyuseless 09 December 2017 02:28:25AM 0 points [-]

Nothing to see here; just another boring iteration of the absurd idea of "shifting goalposts."

There really is a difference between a general learning algorithm and specifically focused ones, and indeed, anything that can generate and test and run experiments will have the theoretical capability to control pianist robots and scuba dive and run a nail salon.

Comment author: Fallibilist 09 December 2017 03:22:43AM *  0 points [-]

If someone points to an AI that can generate scientific hypothesis, design novel experiments to attempt to falsify them and run those experiments in ways that could be applied to chemistry, cancer research and cryonics you'd just declare that those weren't different enough domains because they're all science and then demand that it also be able to control pianist robots and scuba dive and run a nail salon.

We have given you criteria by which you can judge an AI: whether it is a UKC or not. As I explained in the OP, if something can create knowledge in some disparate domains then you have a UKC. We will be happy to declare it as such. You are under the false idea that AI will arrive by degrees, that there is such a thing as a partial UKC, and that knowledge creators lie on a continuum with respect to their potential. AI will no more arrive by degrees than our universal computers did. Universal computation came about through Turing in one fell swoop, and very nearly by Babbage a century before.

You underestimate the difficulties facing AI. You do not appreciate how truly different people are to other animals and to things like Alpha Zero.

EDIT: That was meant to be in reply to HungryHobo.

Comment author: curi 08 December 2017 10:22:51PM 0 points [-]

AlphaZero clearly isn't general purpose. What are we even debating?

Comment author: Lumifer 09 December 2017 12:34:00AM 0 points [-]

This sentence from the OP:

Like the algorithms in a dog’s brain, AlphaGo is a remarkable algorithm, but it cannot create knowledge in even a subset of contexts.

A bit more generally, the claim that humans are UKCs and nothing else can create knowledge which is defined as a way to solve a problem.

Comment author: Subsumed 08 December 2017 12:03:43PM 0 points [-]

I feel the term "domain" is doing a lot of work in these replies. Define domain, what is the size limit of a domain? Might all of reality be a domain and thus a domain-specific algorithm be sufficient for anything of interest?

Comment author: Mitchell_Porter 07 December 2017 09:09:59PM 0 points [-]

Four hours of self-play and it's the strongest in the world. Soon the machines will be parenting us.