Trajectory, a startup that wants to make AI models easier for specialists to improve, has raised $45 million in a Series A round at a $300 million valuation. Sequoia led the round, with participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures.
Why do AI models need post-training?
The 18-person company was founded about a year ago by three Stanford classmates who previously worked at Google DeepMind and Apple. Its platform is built around a basic gap in current technology. Whether a model comes from OpenAI, Anthropic or an open source project, it doesn’t learn from its mistakes without further training.
Businesses adapting these models for specialized fields such as law and finance typically fine-tune general models. That often means hiring internal research teams or relying on expensive consultants, known as forward-deployed engineers, or FDEs. Trajectory wants to automate that work so models can learn from feedback and keep getting better at specialized tasks.
How does Trajectory's platform improve AI models?
The process works like this. A business starts with an open source AI model and connects Trajectory to its applications. The platform collects feedback on how the model performs, then uses that information to retrain it. Over time, the model improves at the specific job it is being asked to do.
Co-founder Ronak Malde compares today’s foundation models to a very smart PhD student who is always starting a new job. “But what the economy values is someone who worked on the job for 30 years,” Malde told TechCrunch.
Harvey and Rogo are already using the platform to post-train open source models. The goal is to make those models perform more like experienced lawyers or financial analysts. They do that by learning from earlier tasks, such as reviewing a particular type of legal contract or valuing a company using a specific methodology. Trajectory’s other customers include Decagon, Mercor and Nvidia. The company is now making the post-training platform publicly available.
Malde said the people using Trajectory decide “what good looks like for their organization and can turn that into improvements for their model.” Lawyers, doctors and other specialists can improve the models nearly on their own, he said. That could make deploying specialized AI faster and cheaper.
The approach has some similarities to recursive self-improvement, the idea that AI systems can improve their own capabilities. Co-founder Michael Elabd argues that keeping humans involved in training makes Trajectory’s method safer. Autonomous self-improvement, he says, could lead to unintended or even dangerous outcomes.
Trajectory is not the only startup chasing this market. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, is behind the fine-tuning platform Tinker. Applied Compute, which was valued at $3.25 billion last month, is also in the space. Prime Intellect offers a platform that helps companies build and train their own AI agents.
Malde believes the startup’s biggest advantage is its team. He was among the handful of researchers Google DeepMind acqui-hired from Windsurf. Elabd also worked at DeepMind, and third co-founder Arjun Karanam previously worked on generative foundation models at Apple. The company has also recruited researchers and product designers from OpenAI, Stripe and Netflix.
Karanam says the team developed an interface that makes post-training more accessible to nontechnical customers. That design work, he suggests, is what will let people outside engineering teams shape how their AI models behave.
Dieser Artikel wurde mit Unterstützung von KI verfasst.
News Factory APP - agentische News für besseres SEO & AEO.