Dangerous capabilities can suddenly appear from gradual progress in AI / hacker news