Field Notes

How we extended SDPO to learn from production data: single rollouts, off-policy, with a 5x improvement on APEX over zero-shot.

Our case for continual learning, and why it's the next era of software.

A concurrent, multi-LoRA training platform for continual learning; 2.81× faster than serial baselines, open-sourced in SkyRL with UC Berkeley Sky Lab and Anyscale.

Trajectory's plug-and-play platform drops any model into a continual-learning loop. We post-trained open-weight Nemotron 3 Super to near-frontier legal performance, fast.

A look at four pioneers (Harvey, Clay, Decagon, and Mercor) building products that learn from how they're used, and how Trajectory is partnering with them to make continual learning real in production.