The problem with AI models today

Current AI models from OpenAI, Anthropic and open source alternatives cannot learn from their mistakes without further training, limiting their usefulness in specialized fields.

How businesses adapt models now

Companies fine-tune general models for specialized work in law and finance by hiring internal research teams or paying expensive consultants called forward-deployed engineers.

Trajectory's solution

The Stanford startup founded about a year ago by three classmates with Google DeepMind and Apple experience wants to automate that process using continual learning.

The funding round

Trajectory's 18-person team raised forty five million dollars in Series A at a three hundred million dollar valuation, led by Sequoia with Nvidia, Bessemer Venture Partners and Bain Capital Ventures.

The PhD student analogy

Co-founder Ronak Malde compared foundational models to a very smart PhD student always starting a new job, but the economy values someone who worked on the job for thirty years.

“But what the economy values is someone who worked on the job for 30 years.”

How the platform works

Businesses connect Trajectory to their applications, allowing it to collect feedback on model performance. Trajectory then retrains the model to improve at specialized tasks over time.

Early customers and use cases

Harvey and Rogo use Trajectory to post-train open source models to perform like experienced lawyers or financial analysts. Other customers include Decagon, Mercor and Nvidia.

Safety through human involvement

Co-founder Michael Elabd argues that keeping humans involved in training makes Trajectory safer than recursive self-improvement, which could lead to unintended or dangerous outcomes.