François CholletVerified account

We need to start using @spelunkyworld as AI benchmark. AI that figures out what to do with Eggplant wins. ;)

Thanks! This has got to be the clearest explanation of RL that exists.... 2 tweets!

It also helps if you just use API access and sidestep all the other tasks humans need to do while playing these games. It's almost as if most video games are laughably primitive, and the "difficult" ones just add more distractions for the player.

This seems to be the real strength of relational networks: They just seem to generalise a whole lot better in RL. Meanwhile CNNs seem to work as a very good particular-situation-detector - there's a reason RL CNN architectures are very different to other CNNs!

Physics simulation or real world data need ti be a new benchmark