Descriptions:
Latent Space hosts Ronak Malde, co-founder of the newly launched Trajectory.ai, for a wide-ranging conversation about his path from Stanford AI research through Windsurf (formerly Kodium) to founding a startup built around continuously learning AI systems. Malde shares a firsthand account of Windsurf’s surprise Google/DeepMind acquisition — including the moment the team assembled expecting Sam Altman and found Demis Hassabis on the call — and discusses his work training SWE-1, Windsurf’s proprietary coding foundation model, from massive agent activity data.
Trajectory.ai’s core thesis is that every future product should be a “living system”: one that improves continuously from its own real-world usage rather than relying on static model snapshots. Malde explains how the company has built on NVIDIA’s Neimotron 3 Super as a base model — a notable choice over the dominant Chinese open-weight alternatives — and why structured agent activity logs (tool calls, workflow steps, task outcomes) provide a uniquely rich training signal that most companies are not yet exploiting.
The episode also covers the rapidly closing gap between Western and Chinese open-weight models, why sectors like legal, finance, and healthcare will increasingly require self-hosted open-weight solutions for data sovereignty, and how Trajectory.ai is positioning itself to power knowledge-work companies that need their AI to get smarter with use rather than freeze at deployment time.
📺 Source: Latent Space · Published June 21, 2026
🏷️ Format: Interview
