The first half of this tweet is about something genuinely cool. Great work, OpenAI.
The second half is some serious eye-roll-inducing mental gymnastics. Why can't we be proud about our successes without fudging numbers to overreach with the conclusions?
The robot didn't get to train *at all* with tied fingers — it had to adapt on the fly.
(Also, humans have a billion plus years of evolutionary practice to solve the cube with untied fingers; the robot only gets about 10,000 years of untied practice.)
3
49
"Billion plus years" is obviously silly as we are going beyond multicellular organisms.
But evolutionary prior in human brain counting for way more than 10,000 years of training of a randomly initialized ANN seems equally obviously true? That has to be what @gdb tries to convey.
1
The question was "How many billion years of training did the Deep RL agent need again?"
Just wanted to point out that humans have an irreducible evolutionary prior, whose effect is hard to internalize. Could certainly have been more precise!
Oct 15, 2019 · 6:53 PM UTC
3
4
