Hacker News Reader

Doesn't machine performance already exceed human performance in a number of areas? As was just demonstrated this week when AI beat the world's beat Go player?

View on HN

From the paper:

> Defeat the best Go players, training only on as many games as the best Go players have played. For reference, DeepMind’s AlphaGo has probably played a hundred million games of self-play, while Lee Sedol has probably played 50,000 games in his life.

Seems a bit unfair. Go player study games of other players too, they don't just rely on personal experience.

But it's still much less than 100 million games. If I assume that playing a Go game (or looking at a record and understanding the moves taken) takes one hour (which is a wild guess since I'm not familiar with the game), 100 million games take nearly 11408 years.

In addition to what someone else who replied to you said about number of games, I also was listening to something recently talking about how much energy had to go into building one device that could beat humans at one task. Today and for the near future it's not a sustainable model at-scale.

But once you train the AI, it retains the knowledge forever. And I would assume it is not difficult to clone one, if that is even necessary.

Reply on news.ycombinator.com