GTC invited Noam Brown, a core member of OpenAI responsible for the inference model o1, to discuss his academic research experience and views on the future of inference models. Focusing on Poker AI: His academic research began in 2012, during his doctoral studies at Carnegie Mellon University, focusing on how to achieve AI beyond human level in the imperfect information game of poker. Multiplayer Poker AI: In 2019, Noam Brown and colleagues further launched Pluribus, which is no longer limited to two players, but extends to multiplayer (six player) Texas Hold'em battles. In terms of inference algorithms, the team has significantly reduced the need for pre training by improving the computation strategy during the game stage, resulting in Pluribus training costs as low as $150 in the cloud, while inference uses approximately 28 CPU cores to perform deep thinking in about 20 seconds per game. The "Diplomacy" project during Meta: Diplomacy is a seven player game that emphasizes natural language communication and humanized strategies such as alliances and betrayal behind the scenes. Its true complexity stems from players bargaining through language and strategically backstabbing or compromising at critical moments. At this stage, he and his colleagues developed Cicero, which achieved human level natural language multiplayer interaction for the first time. This means that artificial intelligence must simultaneously master high-dimensional strategies and flexibility in language expression in multilateral games. Cicero is therefore regarded as an important milestone towards multi-agent environments and natural language reasoning. OpenAI's o1: Liam Brown's pursuit is to enable AI to no longer "develop inference methods separately for each game or application", but to directly utilize the inference process and time, fully achieving super strong decision-making ability in different scenarios. Deep learning is often seen as "System 1 (fast, intuitive)" computation. His work focuses on how to enable AI to allocate computing power more flexibly during the inference phase (System 2), achieving results beyond simple intuitive decision-making through longer or deeper level thinking. Noam Brown's research journey has brought about a significant breakthrough in the field of "imperfect information games", spanning multiple dimensions such as human-machine battles, multiplayer games, and natural language negotiations. Nowadays, he is committed to building more universal inference algorithms at OpenAI, enabling AI to extend inference paradigms to more fields, just like Transformer does to deep learning. This series of efforts will not only affect academic research, but also accelerate the emergence of the next generation of intelligent systems that think like humans or superhumans in practical applications. Video playback: https://www. (nvidia.com)/gtc/session-catalog/? search=Advancing%20AI%20Reasoning&search=Advancing+AI+Reasoning&tab.catalogallsessionstab=16566177511100015Kus /session/1733260712641001zu5P Full text reference: https://dtnews. (substack.com)/p/openai-o1-noam-brown-ai