On Wednesday, a group of researchers who built machine learning systems at Google DeepMind, Apple, OpenAI, and Meta announced they're launching Trajectory, a San Francisco–based startup that wants to solve one of artificial intelligence's most stubborn problems: the fact that today's AI models are static, frozen in time the moment their training ends.
This matters because it points to a fundamental limitation holding back AI progress. The most powerful AI systems today—GPT-4, Claude, Gemini—learn from vast amounts of data during training, then stop improving. They make the same mistakes tomorrow that they made yesterday. While OpenAI, Google, and Anthropic have found remarkable success training increasingly capable models for coding, math, and science, those capabilities plateau the moment deployment begins. In December 2025, Turing award winner Richard Sutton argued at NeurIPS, one of AI's largest research conferences, that continual learning is essential for building superintelligent agents. Trajectory is betting it can build the infrastructure to make that real.
The startup has already raised a $15 million seed round at a $115 million post-money valuation, led by venture firm Conviction with backing from Bessemer Venture Partners, Radical VC, and BoxGroup. Individual investors include Google DeepMind's chief scientist Jeff Dean and Stanford professor Fei-Fei Li, the "godmother of AI" and CEO of World Labs.
Trajectory's CEO and cofounder Ronak Malde previously worked at Windsurf, a coding startup that Google DeepMind acquired for $2.4 billion last year, then hired its top talent including Malde. His cofounders are Arjun Karanam, a former Apple researcher who worked on the Vision Pro, and Michael Elabd, who previously led robotics research at Google DeepMind. Together with 11 researchers and engineers, they're targeting a specific opportunity: AI coding products like Cursor have already cracked the continual learning problem by using real-world user interactions to retrain and regularly ship improvements. Malde argues this is precisely why coding products have exploded in adoption, and why major labs have rushed to build their own coding tools.
The trick is applying this to other domains. Code is objectively verifiable—it either runs or it doesn't. Customer support, legal analysis, and sales don't have such clear success metrics. Trajectory's platform starts companies with open-source models customized for their specific use case, then continuously logs failures and retrains. For Decagon, which builds AI customer support agents, the system tracks when queries get bounced to humans and uses those incidents to post-train a new model as often as every week. Trajectory claims these narrowly optimized models outperform frontier lab models on the tasks that matter most for each business.
This approach addresses a growing pain for enterprises: most companies adopting AI need to hire expensive "forward deployed engineers"—consultants and technical staff embedded inside organizations to continuously troubleshoot their AI stack. OpenAI, Anthropic, and Palantir have all built teams to fill that gap. Trajectory's bet is that a platform solving its own problems eliminates that need.
The startup already counts customers like Clay, an enterprise sales platform, and Harvey, a legal AI startup. While currently focused on AI-native companies, Trajectory plans to eventually market to Fortune 500 firms. Elabd argues the entire AI industry is moving toward this paradigm—learning from experience rather than staying frozen in place.