How Well Does RL Scale? / hacker news