Hacker News Reader

Comment by 0x8BADF00D | Hacker News Reader

> Rapid used 6144 CPU cores and 8 GPUs to train our policy, collecting about one hundred years of experience in 50 hours.

That seemed an order of magnitude higher than I expected. Is training usually this computationally expensive?

View on HN

Disclosure: I work on Google Cloud.

Is it the cores that throws you off? (Those are used for the simulation, not the training). Second, I believe those were preemptible cores, so that's $60/hr for the cores and then $20/hr for V100s (which is what I think they used). $80/hr isn't bad considering how much a (small!) team of researchers costs.

The notable thing about this is how little computation it needed. We have a long way to go, but this is a big step in the right direction.

haha I just read it and thought that's a magnitude less than I expected. Pretty often it is. A lot of papers from high profile institutions have a lot of computing power availiable.

First, it seems to be a lot and be really expensive, but think of it in man-hours. It quickly diminishes.

Not to mention that evolution has had millions of years of optimizing this stuff.

Not always but a lot of machine learning techniques now are effectively just brute forcing and highly expensive band wasteful to compute. Machine learning is getting close but I don't believe we will get there with our current methods

Well if you look at the plot in the "Learning progress" section, you'll see that they did require almost two orders of magnitude more training time due to the randomizations they were adding to the simulation. Without these the policy isn't very robust but also takes a lot less time to train.

Yeah but it doesn't matter; One model to rule them all.

Reply on news.ycombinator.com