4 ms·
The original AlphaGo showing (vs Fan Hui) had obvious problems. They weren't catastrophic, it was clearly playing at a professional level stronger than 1d, but
by roenxi 8y ago
The original AlphaGo showing (vs Fan Hui) had obvious problems. They weren't catastrophic, it was clearly playing at a professional level stronger than 1d, but it also wasn't obviously superhuman. Lee Sedol could have expected to win his matches, and not been arrogant about it.
The gap between that AlphaGo and the AlphaGo that beat Lee Sedol was probably something like 10 years of human improvement compressed into a couple of months. 500 Elo is huge.
Even if they still have further to go, comparing this to AlphaGo's progress is reasonable. I'm going to live dangerously and make an assumption - you probably understand Starcraft better than Go, so can see the flaws more easily. Any mistake in Go comes back to something the AI just didn't see.
- andreyk 8y agoStarCraft is also a FAR bigger challenge than Go. So much so that even saying AlphaStar is similar to AlphaGo is quite a stretch (see https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/ https://deepmind.com/blog/alphastar-mastering-real-time-stra... - the architecture and to some extent training regime are very different). I'd be surprised if they can learn the macro without imitation, unlike in Go ; but we'll see. Would sure be nice if they also released the paper...
- darkmighty 8y agoYea stepping back a little I agree it's a fundamental achievement. It's all about expectations -- I were expecting a Lee Sedol kind of event, and it was more of a Fan Hui (or a bit further) kind of demonstration. A milestone for sure considering how difficult the problem is with various previously unconquered subchallenges (uncertainty, partial observability, massive control space), different but in some ways more difficult than OpenAI's Dota 2 matches. > I'm going to live dangerously and make an assumption - you probably understand Starcraft better than Go, so can see the flaws more easily Correct :) Although I'm not great at either. This comes closer to opening doors to real uncharted territory in robotics and general agency; it trained in crazy low wall-clock time. But now we're this far it makes me question even more the cost, flexibility, adaptability, of the AI less forgivingly. How much did it actually cost to train? Would it be economical to train one of those for every single task? (e.g. in a factory or warehouse setting), Would it be vulnerable to those exploitable strategic weakness in a real world setting?, etc. It's definitely not over yet as far as SC2 goes though :)