Trajectory wants businesses to train their own software models

By Daniel Okafor · Reporting from New York ·

Ronak Malde tells Sequoia Capital that customized models beat expensive frontier systems, showing why businesses must own their training data to build real margins.

Corporate software today runs on rented brains that cannot remember what happened. On the Sequoia Capital podcast, Ronak Malde said that modern artificial intelligence remains trapped in what he called pseudo-continual learning. Companies query giant frontier models the way drivers once consulted folded paper road maps from a motor club. If a road closes mid-trip, the traveler pulls over to call headquarters for an updated route. Current systems ingest billions of words during massive training cycles, yet they freeze in place once deployed.

At legal software firm Harvey, closed frontier models left valuable industry knowledge stranded. When Harvey used Trajectory’s platform to train a specialized model on internal workflows, legal agent accuracy jumped from 4% to 18%. The tailored model ran six times cheaper than the general frontier equivalent. For an enterprise spending on computational tokens, that difference determines whether automated tools deliver a margin or burn capital.

The MapQuest trap

Software teams currently operate like artisans who forge every tool by hand. Malde said that when he worked at Windsurf, building a frontier coding model demanded dedicated research staff and compute. Cursor ran through the exact same drill to produce its composer tool. Each company built bespoke infrastructure to log interactions, replay tasks, and balance loss functions.

The platform packages trace logging, task evaluation, and model deployment into a loop that feeds production data back into training. Malde’s team modified Group Relative Policy Optimization to learn from single trajectories rather than sixteen identical runs across simulated environments. Instead of scoring ten thousand traces, operators grade one hundred representative samples, while a secondary system generalizes the criteria across the rest. This lowers the barrier to entry, allowing corporate buyers to shape algorithms around their own records.

Can an agent learn without forgetting?

The strongest case against this approach comes from the researchers who built the modern discipline. Computer scientists Rich Sutton and David Silver have said that continual learning remains impossible on current transformer architectures. Neural networks suffer from catastrophic forgetting: when you tune a model on specialized data, it loses general competence. Giant frontier labs say that massive foundation models, supported by large-scale reasoning at test time, will inevitably outrun fragile, custom-tuned engines. Why hire engineers to supervise local models when frontier providers handle the underlying intelligence?

That objection misreads how commerce functions. General intelligence is expensive, whereas commercial operations require narrow precision. In real offices, an investment bank or a municipal court does not require an oracle that recites poetry. It requires software that formats balance sheets or checks statutory filings without error. As Malde said, only 1% of enterprise workflows currently use machine agents, precisely because general models fail on local nuances. History shows that technology begins as a centralized utility before decentralizing into specific tools. Mainframes gave way to personal computers because local operators demanded control over their daily tasks.

Specialization always defeats the oracle

Economies expand through the division of labor, as Adam Smith said centuries ago. When every firm relies on identical foundation models, no business builds durable competitive advantage. The value of enterprise technology lies in the accumulated knowledge of the people who do the work. Trajectory’s approach allows firms to route 90% of routine tasks to small, cheap, locally trained models, reserving expensive frontier systems for the rare 10% that require novel synthesis. This shifts corporate leverage away from foundational monopolies and back toward the enterprises that possess real operational data. The future of computational labor will not belong to one omniscient machine. It will belong to institutions that train their own tools on their own terms.

Sources

  1. A Model For Everyone: How Ronak From Trajectory Is Building Products For Continual Learning