Dota 2

Bob McGrew traces OpenAI’s shift from early robotics and game projects toward large language models, arguing that scale and generalization connected those efforts. He describes how reasoning and test-time compute may help models overcome pretraining data limits and make agents reliable enough to take actions. He also discusses how AI startups should begin with the strongest available model, then optimize for cost after finding value with users. The talk closes on AI’s effects on work, the need for useful software tailored to people’s workflows, and the prospect of robotics and scientific progress.

Frank Lantz argues that games create meaning through the gap between a simulation and real life, letting players become absorbed in an activity and then reflect on what it means. He describes game design as shaping spaces for thinking and doing, and sees games such as Universal Paperclips and Shadow of the Colossus as examples of that double movement. He discusses how communities can develop games and play styles through modding, why deep competitive communities matter, and how games can become sports over time. He also explains how NYU's Game Center teaches both coding and artistic judgment, and why creative work and entertainment are hard to predict.

Miles Brundage and Tim Hwang discuss how AI could be misused, how emerging capabilities may change the scale and reach of those risks, and why predictions about future applications require humility. They examine public perceptions of AI, interpretability, fairness, accountability, transparency, and robustness, including how these concerns arise in health care and across different regulatory settings. They describe the need for policy work to connect technical research with government, industry, and civil society, and discuss shared goals between near-term AI governance and long-term AI safety. The talk closes with debate about open publication and predictions for advances in game-playing systems and meta-learning.

The speakers explain how OpenAI built a Dota 2 bot through large-scale reinforcement learning, careful engineering, and repeated experiments. They describe choosing Dota for its Linux support and bot API, building a Python and Gym-compatible environment, and improving performance through self-play, replay data, and expert feedback. The project’s progress depended heavily on infrastructure, reliable code, useful metrics, and fast iteration rather than exotic models alone. The bot’s encounters with professional players revealed both unexpected strategies and exploitable gaps, showing how games can serve as scalable test beds for AI research.