Was excited to join the morning keynote for PyTorchCon. Lots of fun discussions: RL is one of the most promising frontiers because outperforms imitation learning as it learns objectives not demonstrations, discovering strategies beyond training data that supervised learning can't reach, like o1's emergent reasoning traces. But pre-training is not dead. Architectural breakthroughs like diffusion models for coding and DeltaNet improving associative memory performance are promising frontiers. My philosophy is backing bold founders who build at the frontiers, e.g. Voyage AI building vector embeddings years before enterprise AI search took off.