Bringing Continual Learning to Every Company: Trajectory’s Ronak Malde

Trajectory bets continual learning will customise every company's model

In Trajectory's telling, the next phase of AI is continual learning, where a model improves from the work it does in production instead of periodic offline retraining. Its platform turns agent traces into a trained model, and its research aims to shrink that loop until the only input needed is a company's sense of what good looks like. Malde argues the architecture debate is already being settled, and the payoff is cheaper customised models for every enterprise.

Takeaways

• Continual learning updates models from live use. He contrasts it with pseudo continual learning, where agents are smart but retrained in offline batches, like the AAA paper maps his family collected. Changing a route meant calling AAA or waiting for a mass update.
• Trajectory sells one-click training for agent traces. Trajectory's platform handles the loop in one place. Agent traces go in and a trained model comes out. He borrows the philosophy from Apple's combination of the computer and the phone, and expects uploading models back to production to become standard.
• New architectures already support continual learning. Rich Sutton and David Silver argued on this show that today's architectures cannot do it. Malde points to DeepSeek V4.1 Flash, which puts fast and slow memory inside the weights and adds skip connections between layers 2 and 50.
• Harvey's legal model got six times cheaper. Harvey built a legal agent benchmark and trained with Trajectory, lifting accuracy from around 4% for an open-source model to around 18% while costing six times less than frontier models. Rogo's model came out several times more token-efficient and cheaper.
• Two research bets remove simulation from training. A group size-one variant of GRPO learns from a single real rollout per task rather than replaying the same task eight or 16 times, removing the need for a simulated environment. A meta-grader propagates reward from 100 representative traces to 10,000.
• Continual learning keeps humans in the loop. Recursive self-improvement would put a model in a black box and make it the best AI researcher so it can improve itself, which raises alignment questions. Continual learning instead improves at the tasks humans set.
• Malde wants enterprise AI adoption at 99%. Only about 1% of enterprises have adopted AI. Malde wants to take that to 99% by letting every company customise its harness and model, and argues from Adam Smith's The Wealth of Nations that abundance comes from specialisation.

The ingredients of continual learning

1. Trace production trajectories, often using tools such as LangChain and Braintrust.