
Episode 27: Noam Brown, FAIR, on achieving human-level performance in poker and Diplomacy, and the power of spending compute at inference time
Generally Intelligent
McTs
It predicts whatever it's going to do based on the dialogue history. It only does look back one turn, though. So if you've been trying to wait two turns ago and didn't say anything about it, then you're going to clear with this current bot. The planning engine ends up being super useful because these kinds of like tricks are really effective if you don't do planning. And yeah, that's there's some things that you can kind of get away with. There's ways to fool the language models. Like if you want the bot to do something, you just like copy and paste that like 100 times and click that into the dialogue. Then like assign a much higher probability
00:00
Transcript
Play full episode
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.