We trained a single AI for the past 10 months, something we haven't seen before in reinforcement learning. OpenAI Five at The International was 1.5 months old. OpenAI Five at Finals was 10 months old. *Huge* difference in performance. And the curves still haven't leveled off.
19
99
21
513
I agree. The relevant number is GFlops (related with CPU/GPU/whatever power and number of processing units) and consumed energy. Time alone is meaningless.
1
3
Replying to @notjfmc @aminorex
(The x-axis is total amount of compute consumed, and comes with labels!)

Apr 16, 2019 · 5:16 AM UTC

3
7
Replying to @gdb @notjfmc @aminorex
Do you have a graph where the x-axis is the cumulative number of RL training sessions? Would be great as a benchmark for future RL tasks... and congrats to @OpenAI :)
Replying to @gdb @aminorex
Thanks! Twitter crops the image in the default view. I'd add it to the text together with what PF/s means (petaflops?) for people not familiar with ML papers. I guess that energy is not relevant in this experiment, just that performance is still growing?
Replying to @gdb @notjfmc @aminorex
Is this actual or theoretical pfs-days? That is, have you factored in limited GPU/CPU utilization in making these estimates?