I certainly do like games.
The AI Box game always struck me as interesting for several reasons:
If you had this situation in real life, it'd be useless. Presumably a computer would have patience sufficient to outlast all humans trying to get into its noggin; you'd never be able to release it safely.
Eliezer looked for people with total confidence they'd succeed in the tiny two-hour time frame; I think that was to increase his chance of winning. This isn't just Dunning-Krueger - if you think you can always suss out lies and bad information, you're a sucker waiting to happen.
The way the challenge is set up, I think the AI's at an unfair disadvantage (and the setup's designed that way for PR reasons) because the human communicator need not engage the AI in an effort to evaluate it for release. I'd think forcing some level of engagement other than repeated, "No"s would be fair.
Oddly, I'm highly confident I'd win the AI Box game as the human, partly because I think that my being persuaded to release the AI over the very long term isn't merely likely, but inevitable. I think this leads to more diligence. I'm also in a profession where people lie to me on a semi-regular basis; that helps.
I think I could win as the AI Box against many people. But that's because I'd be especially friendly AI. If the other person is obliged to engage and stay in character.... well, I've got a plan.
I'll play as the human on negotiable terms; my charity would likely be National Center for Science Education. The only term I'd be pretty inflexible about is that I'd want to make the transcript publicly available. (In the extremely unlikely event that Eliezer wants to take me on, I'd be willing to place an embargo date on the transcript - say 18 months out.) Private message me if you're particularly interested.
Starcraft is a bad game, though; it's only popular because the ridiculously primitive 1998-era interface means that actual physical speed is required to control your units correctly, which adds barriers to entry to competitive play and makes it more challenging to play and therefore more impressive for someone to be good at. It's pretty much the embodiment of fake difficulty in game design.
The relative physical speed is what counts. The best players would benefit from a modern interface at least as much as much as the worst.
Fake difficulty is a meaningful word only in singleplayer. Fake difficulty is giving computer controlled opponents more hit points or map hacks instead of better AI. In multiplayer, the difficulty is provided and dependent on the human opponent who is subject to the same rules as you, and the game is just a medium - a chess board, a tennis court.
Edit: And barriers to entry are actually lower for Starcraft relative to other games because it's so old and so popular - there is an entire encyclopedia devoted to it it full of advice and ready to use game plans.
Fake difficulty applies to multiplayer too. Anything that adds barriers to entry or needless clicks is fake difficulty. Games like Starcraft, where you sometimes end up fighting the interface instead of your opponent, have a lot of fake difficulty. If you're going by That Other Site's definition of fake difficulty, the #1 thing on the list is "Bad technical aspects make it difficult," which certainly seems to apply!
For example, in Starcraft you have to micro all your workers to different mineral patches at the start of the game in order to get the most efficient economy possible. This is fake difficulty because games with real interfaces allow you to select all and click once, then the workers automatically fan out. Starcraft requires at least 8 (in practice usually 10) clicks in order to accomplish what other games do in 2. Further, some of the Starcraft community actually wants this "feature" to be preserved for Starcraft 2, as it "adds skill." Fortunately, I don't think Blizzard is going to acquiesce.
Um, I suppose your evidence is true, but the game is great in spite of its 1998-era interface. The balance between the races is sublime.
I never really got into playing starcraft because of the primitive interface, i could never really enjoy playing it, but I am into watching korean matches with english commentarys on youtube.
I think that the primitive interface makes the game less enjoyable for me, but doesn't add 'fake difficulty'. I like that its a very difficult game to play well in terms of micro and macro, and then on top of that starcraft is also rich in strategy and 'tradition' (for some reason I like that starcraft is a very old game)
I'll second the recommendation for the game of Go; I would be interested in finding out who here plays, and at what level; and my two years' experience with the game has taught me that it held many deep lessons about my own thinking and improving it.
These days my Go is on hiatus except for one DGS game.
Not entirely sure what I think about the donation angle.
I'd play Advanced Go with you. I'm a frustrated 2k in regular Go - frustrated because I've been stuck at 2k for a couple years.
I'd love to play with either of you, though I'd need nine stones to have a chance. I'm somewhere around 10-15k and I usually play on IGS.
The "Advanced" version sounds interesting, how in practice would you suggest implementing that for Go ?
We could play with very slow time settings, maybe 2 or 3 hours per player. Or we could play on DGS. Players can use whatever resources they want, except for other human beings.
I'm on KGS and DGS as "OneTrue." I'm a 2k on KGS, just registered on DGS.
Want to play a game on DGS? If I win, I'd like you to donate $20 to SIAI.
I'd be interested in a DGS game with the additional condition that each player document which computer resources they are using, either prior to the game or as soon as they start using them. I intend to use Kogo's Joseki dictionary and the fuseki.info openings database to start with.
I'm using the same handle on KGS and DGS as on LW.
I'll accept the bet, and request a matching donation to KIPP should I win.
OK. I'm planning to use CGoban3 as my SGF editor, and Kogo's and MasterGo for openings.
I'm sending you an invitation on DGS now. If you don't like the settings, you can reject the invitation and send me a different one.
Looks like you got Black. Onegaishimasu !
I might use CGoban as well. If we both are, we might as well agree that it's OK to use its score estimator; though I wouldn't trust it much until the yose.
Interested onlookers, you can follow the game here. If you want to comment on the game, I would suggest a) using this comment thread and b) rot13ing your observations if they could influence play.
And White wins... on time. :-/
Thanks for the game, it was interesting.
I'm not at all sure at the point we had reached how to estimate who's in the lead (that can be one of the frustrating mysteries of Go). The CGoban score estimator says B+20-something (I think that overestimates the center) and the GnuGo estimator says W+15 (but doesn't give a "visual" explanation of its guess).
Would you be willing to write up comments on your moves and how you used the other resources, and make a post of them?
First: you might be interested in the "Malkovitch" games at GoDiscussions.
LW isn't the venue for a deeply commented game of Go, but it might be worthwhile for the players in such a game to post here with observations on where they felt the game highlighted this or that aspect of their thinking processes.
OK. I think having both players make such posts would be unnecessary clutter, so how about if we combine both into one post and the winner posts it? :-)
BTW, if you wanted to play, Blueberry, I'll offer you the same conditions.
2k on what scale ? I was ranked as 2k on KGS before I quit - I'm not sure how well that reflects my playing ability, I was playing blitz almost exclusively.
Are you on KGS ?
How well does Advanced Go really work for reasonably good players (where I'm defining "reasonably good" to include 2k players, for this purpose)? What little knowledge I have of go software is well out of date, but I had the impression that unlike in chess, where computers were better than humans at tactics long before they were better overall, go programs aren't terribly good at anything (on a full-sized board) unless one throws an outrageous amount of hardware at them.
What do you mean by "better at tactics long before they were better overall"? Getting the tactics right seems to be the point. Do you mean 'good at seeking immediate goals but having a relatively poor lookup mechanism for evaluating possible future board positions'? Equivalently 'Better at thinking n moves ahead but worse at guessing how good the n configurations will end up at n+5'.
Yes, that's the kind of thing I mean, though perhaps for larger values of 5. It is customary for chess players to distinguish between tactics (stuff you can work out by searching) and strategy (stuff you can't, where you play according to general principles / feel / high-level anticipation of what sort of thing will be happening several moves down the line).
Of course in the limit of outrageously effective searching strategy gets absorbed into tactics, but even in chess no player (human or computer) can look that far ahead. And in the not-exactly-limit of merely very effective searching, you can afford to be not quite so good at strategy if you can stomp your opponent tactically. This is generally how computers win.
The fact that computers and humans have distinctly different skill profiles is what makes "advanced chess" interesting: a hybrid with the strategic understanding of a good human player and the tactical monstrosity of a good computer player is very strong indeed.
(Having said which, I believe there's some evidence that even a not-all-that-good human player armed with multiple computers running different programs can be scarily effective too.)
(Having said which, I believe there's some evidence that even a not-all-that-good human player armed with multiple computers running different programs can be scarily effective too.)
If you're thinking about the same thing I am, the player was "not-all-that-good" at chess, but knew a lot about chess programs and their different relative weaknesses and strengths.
Hypothetically, I wonder if that approach could be constructively imitated by a computer. A meta-chess program, dividing it's computational resources between several subprograms, and combining their input to play better than the subprograms would if they had the full computational resources.
I think we are indeed thinking of the same instance. And yes, it would be interesting to try getting a computer to play that way.
Here's a nice exploitation of a similar idea: The Fastest and Shortest Algorithm for All Well-Defined Problems; see also the discussion at Hacker News, where in particular you might want to read the comment from me that explains roughly what's going on and the comment from Eliezer that explains one way in which Hutter's description of his algorithm claims more than it really delivers. None the less, it's a very neat idea.
Since you're above the level of any commercial Go program, why would Advanced Go be any different than regular Go?
So the actual computer Go program isn't the advantage then: the advantage comes from having paper and pen, or a sample board, to try out different sequences? This strikes me as a little different in spirit from Advanced Chess, where the computer actually is a really good player. It's more like Postal Chess.
I enjoy Go, but I'm an absolute beginner. If I could remember exactly how many games I've played, I'm pretty sure I could count them on one hand.
I've been meaning to try out Dave Peck's Go Which is said to have a nice interface and doesn't require you to sign up for an account. You start a game by entering both players' email addresses.
I have an email account at gmail.com under the user name bjaress if
anyone wants to play.
I used to play casually, somewhere in the 10-13k range (I didn't play often enough or formally enough to have a tighter estimate), but have hardly played any go for years. Bloody stupid of me; it's a marvellous game.
I find this idea very appealing; although like you, not for the sake of the singularity, but for my own entertainment.
First, I have a confession to make: I don’t really care how much money gets donated to the Singularity Institute, nor am I trying to drum up money for some other cause. I mainly want you all playing games.
I approve. In a loosely related example I was recently accepted a 4:1 karma bet proposed on how Alicorn would respond a comment. I mention this because turning predictions into bets is an exercise in calibration. It is staking (the usual trivial amounts of) status more specifically on predicted outcomes rather than on eloquence of expression or, more importantly, aiming 'predictions' at whatever is expected to give the greatest political payoff rather than what is correct.
As a general policy, it's probably not wise to have significant bets riding on the subjective opinion of someone aware of the bet. I would have been sorely tempted to spin my evaluation of the comment if what I said determined where some money went and I had a preference between the various options.
Good point, I do consider that bet to be an anomalous example of the class 'disagreements that can be resolved to testable predictions that can be bet on rather that rhetorised about'.
I see mutual consent as an important element of games.
If you really believe the world is at stake and have a way to extract value from me without my consent then I don't philosophically objection to you playing that game. In the sense that I tend to approve of people doing what is rational for them even if I have to punish, shame or implement potentially terminal deterrent measures.
If you really believe the world is at stake and have a way to extract value from me without my consent then I don't philosophically objection to you playing that game.
In other words, you consent to game playing.
Sure, if 'I will kill you if you try anything but acknowledge that you are making the right move given available info' counts as consent.
I guess that bit about "mutual consent" was sort of a cryptic remark on my part.
What I was trying to say is that I generally feel everyone except the players should butt out unless there's a dispute. If I suggest that a particular game be played or offer "official" rules as a third party, I won't mind at all if the players agree to do it differently or plug a loophole. I think it's important for everyone involved to have that attitude.
Anyone play Age of Empires? I'd wager a buck that I'd win a game of AoE 2, to be played Saturday.
Now that is tempting. If I had an acceptable net connection (in terms of ping) I'd take you up on that challenge.
That could work.
What's a 'build order' again? Something to do with 'Fuedal Rushes' if my memory serves me.
I never really got into the online guides. I'm entirely self taught. Would 7 PM (PST) work for you?
Sorry, have been out of town.
I never really got into the online guides. I'm entirely self taught.
I've read the guides but never implemented them.
Would 7 PM (PST) work for you?
I'm in Melbourne. I believe that makes 7pm your time 2pm the next day for me, which is fine. I'll message you.
Some of you have been trying to raise money for the Singularity Institute, and I have an idea that may help.
The idea is to hold public competitions on LessWrong with money going to charity. Agree to a game and an amount of money, then have each player designate a charity. After the game, each player gives the agreed upon amount to the charity designated by the winner.1 It’s a bit like celebrity Jeopardy.2
Play the game here on LessWrong or post a record of it.3 That will spread awareness of the charities and encourage others to emulate you.
The game can be as simple as a wager or something more involved:
We’ve already played few games on LessWrong, more on those in a moment.
In most ways the AI problem is enormously more demanding than the personal art of rationality, but in some ways it is actually easier. In the martial art of mind, we need to acquire the realtime procedural skill of pulling the right levers at the right time on a large, pre-existing thinking machine whose innards are not end-user-modifiable.
First, I have a confession to make: I don’t really care how much money gets donated to the Singularity Institute, nor am I trying to drum up money for some other cause. I mainly want you all playing games.
Not just playing them, of course. Playing them here in front of the rest of LessWrong and analyzing the moves in terms of the “personal art of rationality.”
We need more approaches to improvement. Even in Eliezer_Yudkowsky’s Bayesian Conspiracy fictional series/manifesto, many other schools of thought (called, for dramatic effect, “conspiracies”) were present. As I recall, the “Competitive Conspiracy” was mentioned frequently, but there are other reasons for choosing to start with games.
Games are fun, of course. They also deal with the “personal art” of getting familiar with your own brain, which I think has been underrepresented on LW. I do believe there are certain important things we can’t learn properly just by reading, arguing, and doing math (valuable as those techniques are). Games are an easy, intuitive first step to filling that gap.
We’ve already had a few games here. Warrigal held an Aumann’s agreement game competition. I created Pract and played it with wedrifid. Neither one has caught on as a LessWrong pastime, but the comments revealed that people here know many interesting games.
And there’s the donation hook. Some of you believe the world is at stake, so that’s a nice motivator.
Of course it’s not always the case that you have exactly one winner. I suggest that if there’s no winner each player give to their own charity, and if there are multiple winners each player splits up their donation evenly among the charities of the winners. ↩
This is, of course, not a new idea. I’m just suggesting that we adopt it and make it a common practice on LessWrong. ↩
Unless it contains some remarkable insight, I wouldn’t make it a top-level post. The comment area here or in one of the open threads would be good. ↩
I see mutual consent as an important element of games. ↩