Jeff Dean, Google’s Chief Scientist and engineer behind MapReduce, Bigtable, TensorFlow, the TPU, and Gemini, spoke with Y Combinator’s Diana Hu about AI’s trajectory. Reflecting on his May 2025 prediction that AI is at the level of a junior engineer, Dean said the models are now “actually pretty capable” and the prediction seems “pretty spot on.” He admitted he underestimated how fast models would improve at complex, agent-based tasks, noting these systems are also excelling outside coding. For a 2027 prediction, Dean foresees greater automation of ML systems themselves, where models run experiments, break problems into subproblems, and improve through fully automated experimentation loops, applicable to any field with a measurable objective. The conversation also covered the origin stories of the TPU and MapReduce, Dean’s napkin-math instincts, and his advice for founders seeking defensible problems in an era of capable general models.