You can’t just give OpenAI a mulligan on vision, if you want to claim superhuman performance. Not saying there is zero progess here but on the question of inducing a game from experience, and beating humans on a level playing field, it’s a step back compared to DQN on Atari.https://twitter.com/Smerity/status/1031248288958177280 …
-
As a vision researcher I believe the perception of the API could be replicated with modern techniques (unlimited data after all), though not without effort. OpenAI focused on showing that complex coordination and long horizon play was possible, and did that.
I would be more impressed, for sure, if they managed to put all that together in an amount of compute that didn’t utterly dwarf what a human used.
-
-
There are caveats ofc., but OpenAI has been transparent about the compute & simplifications they used. Focusing on a subset of a problem at first is one of the basic principles of AI research, so it's hardly fair to single them out for an 'would have been more impressed' attitude
-
Surely I didn’t single them out. I wrote a whole arXiv on AlphaGo, in January...
- 1 more reply
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.