Why does so much brainpower and computation solve irrelevant problems? Here's my view: It's embarrassing to admit we lack simulators for meaningful problems. And solving the mismatch to worthwhile problems is a different type of "hard problem" than today's practitioners enjoy.
Greg Brockman
 @gdb
15 Apr 2019
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.
3
2
2
13
Replying to @dan_s_becker
Dota is a convenient testbed for pushing the limits of general-purpose deep RL technology. Here's a physical robotics problem we solved using the learning system we wrote for Dota: blog.openai.com/learning-dex…

Apr 15, 2019 · 10:52 PM UTC

2
4
1
32
Replying to @gdb
As a pre-mortem: If RL hasn't mitigated important problems (hunger, disease, climate change, etc) by 2040, will it be primarily because 1) We're insufficiently good at dynamic optimization OR 2) We're good at this type of optimization, but don't know how to apply it I think 2
1
3
The DOTA effort has lead to obvious progress on #1. There's some value to that. But we think too little about applications. And I speculate it's because we draw from a community whose talents are closer to pure optimization.
1
Replying to @gdb @dan_s_becker
I agree partially with @OpenAI: we need research to achieve technological progress; but, how much further we want to persist on the gaming exploration path without tackling key questions of what reasoning is and how it can be implemented?
1