Hacker News Comments on
AlphaGo and the Future of Computer Games: A Conversation at the University of Alberta, Version 1.0
Paul Lu
·
Youtube
·
53
HN points
·
1
HN comments
- This course is unranked · view top recommended courses
Hacker News Stories and Comments
All the comments and stories posted to Hacker News that reference this video.> no one right now is even really tryingAnd we're supposed to judge by the author's description of "silence" and "nervousness" that befell an expert panel. I can assure you that most AI researchers are trying, and are just not in the business of writing long-form articles to the public asking for donation.
> See eg. the recent paper by Grace et al.
A self-selected group of NIPS/ICML authors don't constitute experts. NIPS/ICML authors are the core of the community. The experts would be the top 1% of the community, i.e. either the authors with the most citations or most papers or just generally regarded highly by peers.
edit 1: Go players are not the experts I'm talking about. I'm talking about AI experts, and no not amateur AI hobbyists who know how to do Pseudo Monte Carlo. I mean, such as, people doing RL research. Watch, for instance, this: https://www.youtube.com/watch?v=UMm0XaCFTJQ
⬐ apsec112"And we're supposed to judge by the author's description of "silence" and "nervousness" that befell an expert panel."I make this judgment based on, among many other things, the tiny budgets given to people like Tetlock to study predicting events even a few years out; the fact that Kurzweil's very simple methods, basically "just draw a line through the curve", are still considered big news among many financial and political elites; that nobody had bothered to spend $100K on a good survey methodology for AI prediction, before the paper I linked came out earlier this year; that a friend of mine, who is supposed to run a (small budget) government program on forecasting, has to ask me where to get datasets on past tech progress because nobody has ever bothered to compile them into a standardized form, and so on.
"I can assure you that most AI researchers are trying"
What serious forecasting attempts, with specific dates attached to specific events, have been done in this vein?
"The experts would be the top 1% of the community"
IIRC, NIPS has around 5,000 people, so the top 1% would be like 50 people, and most of them won't respond to a survey. That's not a reasonable sample size.
(edit: this article doesn't ask for donations to anything; the links at the bottom are all to various papers and research materials, so getting money is obviously not the main goal)
(edit 2: the video linked is from after AlphaGo came out. I'm sure many people, after AlphaGo happened, claim that it was easily predicted. Again, hindsight is 20/20.)
⬐ TwoBitMost computer games these days have terrible AI. I'm starting to think that it's intentional. Otherwise I would fare much worse in the game.⬐ ktRolster⬐ naringasThe first time I ever encountered a game that had superior AI to humans was Uniracers, back in the 90s. Great game, but the manual stated that they had to make the AI worse before they shipped because no one could beat it.At the time it seemed novel, and even rather hilarious. Now it seems somewhat standard (aimbots can win every time, for example).
⬐ taneqThe purpose of video game AI is not to beat you, but to lose to you in a fun way.⬐ eruIt's much more fun for a human to win with a small force against seemingly overwhelming number of enemy mooks.If computer games had awesome AI trying to beat you, you'd need the overwhelming superiority of numbers just to stay alive.
That applies to the computer competing with you, though. There are games, like eg RPGs, with more space for cooperative AI.
I wanna play civilization vs a crowd trained (and probably intentionaly crippled) AI of this kind.⬐ skybrianThe sound seems pretty poor so I'm not going to listen to this. Anyone want to summarize?⬐ kami8845⬐ deepnetNobody? Come on, somebody help the poor guy out.AlphaGo's 1st incarnation included an input scalar that the Net was taught represented the level of Human Play to Emulate (handily included in the dataset).Thus AlphaGo plays as a human (before MCTS), it's choices model a challenging oppenent.
If the dataset included the name of the player in the Dataset then AlphaGo could play in the style of that player.
This offers a human experience, and for teaching one can show what different experts at differing levels would choose.
⬐ taneqThis (along with the Neural Algorithm for Artistic Style paper) opens some interesting questions about simulating specific humans and the relationship of such simulations to mind uploading. Obviously it's a huge distance away from a proper upload, but with enough input data maybe this could form the basis of Alastair Reynolds' "beta simulations".