Sergey Levine on the Bottlenecks to Generalization in RL and Picking Good Research Problems - 2023

Title : Sergey Levine on the Bottlenecks to Generalization in RL and Picking Good Research Problems Author(s): Generally Intelligent Link(s) : https://www.youtube.com/watch?v=3DMml0vcKJQ

In science its a really good idea to see how extreme of a design can still work. Sergey gets a lot of comments about "We know how to do this task, who don't you use those parts" etc - one could, however to understand the value of a new design, it makes sense to zoom in to just that instead of trying to put in things just to compensate for all the parts where we might have better existing ideas. As an analogy, to get better engines for electric cars, you would not start with a hybrid car but an electric racecar to see how fast it can go, after that you can combine it with sober decisions to it everything to make it work (i.e. do the most extreme thing first).