Trajectory's Ronak Malde on continual learning: Harvey trained 6x cheaper than frontier

Training Data (Sequoia) · rss · 2026-11-05

Sequoia's Training Data podcast hosts Trajectory co-founder Ronak Malde on continual learning as the next paradigm:

Today's agents are stuck in 'pseudo continual learning'—smart but updated in offline batches
Two research bets: a Group Size-1 GRPO variant learning from single real-world rollouts (no simulated env), and a meta-grader propagating 100 representative traces across 10,000
Case study: Harvey trained a model 6x cheaper than frontier, lifting an open-source baseline from 4% to 18% on its legal benchmark
Continual learning keeps humans in the loop, unlike recursive self-improvement
Prediction: product builders will become gardeners setting trellises for self-growing products

More from Venture