Current AI models, whether created by OpenAI, Anthropic, or built on open-source frameworks, face a significant challenge: they cannot learn from their mistakes without additional training. When businesses implement these models in specialized sectors like law and finance, they often resort to fine-tuning general models. This typically necessitates either building internal research teams or enlisting costly consultants, referred to as forward-deployed engineers (FDEs).
Trajectory, a startup launched about a year ago by three Stanford alumni with backgrounds at Google DeepMind and Apple, is tackling this issue head-on. Their goal is to automate the process that enables AI models to learn from user feedback and improve continuously for particular tasks.
Recently, Trajectory announced its success in securing $45 million in a Series A funding round, achieving a valuation of $300 million. This round was spearheaded by Sequoia, with additional investments from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures.
Co-founder Ronak Malde likened foundational models to an exceptionally bright PhD student perpetually starting new roles. "But what the economy values is someone who has been in the field for 30 years," he stated in an interview with TechCrunch.
Trajectory plans to address this limitation through a methodology called continual learning. Organizations can begin with an open-source AI model and integrate the Trajectory platform into their systems, enabling the model to gather feedback on its performance. This data is then utilized to retrain the model, promoting its improvement over time in specialized applications.
This functionality allows professionals in law or medicine to enhance the AI models almost autonomously, which Malde believes could significantly expedite and reduce the costs of implementing specialized AI solutions.
Companies such as Harvey and Rogo have already adopted Trajectory's platform to enhance open-source models, helping them evolve to function more like seasoned lawyers or financial analysts capable of drawing insights from past tasks, such as analyzing legal contracts or utilizing specific methodologies for company evaluations. Other clients using Trajectory include Decagon, Mercor, and Nvidia, and the company is now making its post-training platform broadly accessible.
Users of Trajectory's system have the autonomy to define "what good looks like" for their organization, translating this vision into tangible improvements for their AI models, Malde explained.
While Trajectory's methodology of continual learning shares some concepts with recursive self-improvement (RSI), co-founder Michael Elabd emphasizes the importance of retaining human involvement in the training process. This precaution aims to minimize the risks associated with wholly autonomous self-improvement, which could inadvertently lead to negative consequences.
Trajectory is not alone in its quest to simplify post-training processes; it faces competition from various startups, including Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, as well as Applied Compute, which boasts a recent valuation of $3.25 billion, and Prime Intellect, which assists companies in creating and training their AI agents.
Malde asserts that the greatest strength of Trajectory lies in its team. He, along with Elabd, are among the few researchers whom Google DeepMind acquired from Windsurf, while co-founder Arjun Karanam has expertise in generative foundation models from his time at Apple. The startup has also successfully recruited talent from OpenAI, Stripe, and Netflix.
The significance of design is paramount for a company specializing in AI training, according to Karanam. The team has developed an intuitive interface aimed at making post-training user-friendly for clients who may not have technical backgrounds.
