Interview, Fireside Chat
Noam Brown: AI vs Humans in Poker and Games of Strategic Negotiation | Lex Fridman Podcast #344
- Predicts potential use of large language models for NPCs in Elder Scrolls 7 if released between 2024 and 2026.
- Anticipates AI systems capable of generating complex dramatic interactions and passive-aggressive arguments that could cause emotional distress, though with less deep psychological impact than real-world equivalents.
- Envisions AI aiding geopolitical decision-making and running diplomatic simulations to avoid negative-sum outcomes like war.
- Foresees development of AI tuned to specific ELO ratings or human styles, such as mimicking Magnus Carlsen, to assist human players in strategy and self-assessment.
- Warns that increasingly human-like AI will create serious challenges for cheat detection in human-versus-human competition, necessitating improved defenses.
- Raises the prospect of a "civil rights movement for robots" regarding systems displaying emotion or suffering.
- Suggests AI may eventually determine optimal life strategies, though defining the correct reward function to minimize unintended consequences remains a difficulty.
- Notes the possibility that six-player poker is a "potential game" where approximating Nash equilibrium through self-play is provably effective.
- Identifies data efficiency and the availability of human data as major bottlenecks for scaling AI to complex domains like Diplomacy.
- Believes leveraging huge datasets across domains to specialize them will augment training data for games like Diplomacy and Hanabi.
- Expects short-term deployment of reinforcement learning in well-defined domains like code generation and theorem-proving with "incredibly powerful" results.
- Hopes long-term scaling in real-world settings will be achieved by overcoming current data efficiency barriers.
- Speculates on future legislation potentially making consumer lies illegal, while acknowledging the trade-off with being polite or nice.
- Suggests the approach used in Diplomacy could be applied to cooperative card games like Hanabi and to create NPCs in video games or the metaverse by conditioning language on specific intents.
- Predicts it is possible for AI to achieve superhuman poker performance by inducing emotional stress in human players without maximizing direct monetary objectives.
- Believes it is possible to use AI to assist lawyers in negotiations, though real-world deployment is currently limited by the lack of well-defined action spaces and reward functions in fields like divorce law.
- States the performance ceiling for games using language models and self-play is likely much higher than current levels.
- Notes research may expand self-play intents to include long-term cooperation, gossip, or social topics, increasing complexity.