Trajectory: Ex-Google and Apple Researchers Launch Startup for AI's Missing Feedback Loop

A new startup named Trajectory, founded by a distinguished group of AI researchers from Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs, has announced its launch. The company's ambitious goal is to develop a pioneering platform that enables AI products to continuously improve by learning from real-world user interactions. This initiative directly addresses a long-standing challenge in artificial intelligence: the inability of most current AI models to get smarter after their initial training phase is complete, a limitation researchers have identified as a significant barrier to further AI progress. While leading AI labs like OpenAI and Google have achieved remarkable success in training increasingly powerful models for specific domains such as coding, mathematics, and science, these systems typically become static once their training is finalized. Despite recent advancements in continual learning, the broader tech industry has struggled to implement AI products that can adapt and learn from their errors in real-time. This deficiency was underscored by Turing award winner Richard Sutton in December 2025, who emphasized the critical role of continual learning in the development of superintelligent agents. Trajectory has successfully secured a $15 million seed funding round, valuing the company at $115 million post-money. The round was led by Conviction, with notable participation from Bessemer Venture Partners, Radical VC, and BoxGroup. The startup also attracted individual investors, including Google DeepMind’s chief scientist, Jeff Dean, and the renowned "godmother of AI," Stanford professor and World Labs CEO Fei-Fei Li. The co-founding team comprises CEO Ronak Malde, formerly an AI researcher at Windsurf and later Google DeepMind; Arjun Karanam, an ex-Apple AI researcher who contributed to the Vision Pro; and Michael Elabd, who previously worked in Google DeepMind’s robotics division. The company's approach involves starting with open-source models, which are then "post-trained" using real-world data reflecting how users interact with specific AI products. Ronak Malde highlights successful early examples in AI coding products like Cursor, which leverage user data for continuous improvements. Trajectory aims to extend this powerful technique beyond coding, applying it to a wider array of AI-powered tools. For instance, a customer like Decagon, which builds AI customer support agents, uses Trajectory to log instances where its AI fails (e.g., bouncing a query to a human) and then uses this feedback to post-train a new model, sometimes as frequently as weekly. Trajectory claims these specialized, post-trained models outperform frontier labs’ general models on tasks critical to a company’s specific product needs. While critics might argue that Trajectory's current weekly updates don't represent "true" continual learning in the traditional sense, the company asserts it's merely the beginning. Michael Elabd, a co-founder, explains that the AI industry is evolving towards a paradigm where AI learns from experience, mirroring the rapid advancements seen in AI coding. Trajectory's ultimate vision is to develop a platform capable of updating a company’s AI model daily, hourly, or even with every single interaction, potentially enabling tailored AI learning for every individual within an organization. This continuous feedback loop promises a future where AI systems are not static tools but dynamic, ever-improving collaborators.