Scaling up Continual Learning — Ronak Malde, Trajectory
Real-world inference data—not benchmarks—is the key signal for continual learning at scale.
“the real unlock moving forward will be continual learning”
Ronak Malde, founder of Trajectory (formerly Windsurf, which trained the model behind a $2B DeepMind acquisition), argues that AI's next frontier is leveraging the hundreds of trillions of inference tokens generated daily as real-world training signal, rather than costly synthetic benchmarks that are disconnecting from real use cases. He frames continual learning—updating models continuously from live deployment feedback—as the natural successor to pre-training scaling. The talk is notable for citing Karpathy, Schmidhuber, Hassabis, and Nadella as aligned voices, signaling broad industry convergence on this direction.