It was a tough competition! I didn’t contribute much time to it, but it was interesting to read about the results and François’ comments once it ended.

I have been following Deepmind’s research for some time now. They do great research and produce great blogs/papers too. Deep RL is really amazing technology and what they achieve with DQN, R2D2 and now Agent57 shows it is quite capable at these non-real world tasks.

Personally, I believe that these systems aren’t going to take us towards AGI though. Something more logic-based, with hierarchical learning, program synthesis etc. Chollet hints at it quite a bit, or people like Goetzel and Hofstadter try to dig into it a bit more (though more theoretically than practically).

But who knows, exciting times.