Descriptions:
In this Two Minute Papers interview, host Károly Zsolnai-Fehér sits down with Jeff Dean — Google’s Chief Scientist, co-creator of MapReduce and TensorFlow, and longtime head of Google Brain — for a wide-ranging technical conversation about where AI is headed and what it will take to get there.
Dean pushes back on the popular narrative that the industry is running out of training data, pointing to underutilized video data, multi-pass training techniques, and algorithmic improvements as viable paths forward. He explains how reinforcement learning rollouts work in practice — generating hundreds of candidate solutions, filtering by compilation success and unit tests, and distilling the survivors back into training data — illustrating why more compute directly translates to better synthetic data quality. The discussion also covers continual learning architectures and the safety-testing checkpoints required before updated model versions reach users.
Looking a decade out, Dean reflects on the leap from LSTMs and early sequence-to-sequence models to today’s systems, and argues there is no reason to expect that pace to slow. He highlights multi-agent workflows — including Google I/O demos of models autonomously writing operating systems — as an early signal of what a further million-fold compute increase could unlock, with particular excitement about AI-accelerated scientific discovery.
📺 Source: Two Minute Papers · Published June 01, 2026
🏷️ Format: Interview
