Bringing Continual Learning to Every Company: Trajectory’s Ronak Malde
Ronak Malde argues that continual learning, where agents improve from real-world traces, is the next learning paradigm, and that Trajectory makes it a simple product so every company can train its own models.
Watch on YouTube ↗Charlie's summary
- ▶ 0:59
Today's agents are in 'pseudo continual learning': AAA paper maps now, MapQuest next, Google Maps as the goal
Ronak uses a road-trip analogy. Today's agents are like AAA maps: you consult them offline and any change needs a mass batch update. Letting everyone train and customize models is like MapQuest: still offline, but no trip to AAA. Real continual learning is Google Maps, where every route driven updates the whole system in aggregate. That is the world he wants to reach.
- ▶ 3:02
Architectural change is already arriving, but the missing layer is connecting models to the real world
Asked about Sutton and Silver's doubt that today's architectures can do it, Ronak points to models like DeepSeek, which add fast and slow memory in the weights and skip connections between distant layers. He wants to build on that research and connect it to real use. Companies know what good looks like, and the hardest part is turning that into model improvements.
- ▶ 4:35
Training a frontier model took Windsurf six to nine months and hundreds of millions of dollars; Trajectory packages it
At Windsurf, Ronak says, training a model needed hundreds of millions in R&D and compute, a large research team, and six to nine months, plus consulting-style handholding. His view is that a good product appears when every company is doing the same thing repeatedly. Package it with a few knobs for 'what good looks like' and hide the rest.
- ▶ 7:10
Harvey trained its own legal model, six times cheaper, rising from about 4% to about 18% accuracy
Harvey had built on closed frontier models and couldn't use its 30 years of domain data. With Trajectory it trained a model on its legal agent benchmark. Ronak says accuracy went from around 4% for an open-source model to around 18%, at six times lower cost than frontier models. That lets them spend many more tokens per task. He admits there is still much room to grow.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.