NeurIPS 2024 Math-AI Workshop talk: inference-time compute

NeurIPS 2024 Math-AI workshop keynote. Brown frames 2019→2024 AI progress as primarily training-compute scaling, argues inference compute is the next axis with much further to push. Anchors the case in his poker-AI lineage: Libratus (2017, heads-up no-limit Hold'em) and Pluribus (2019, 6-player) both won by giving the bot "thinking time" at inference. Public source of the "20 seconds of thinking = 100,000x more pre-training data" framing he later took to TED AI.