Trajectory Launches to Solve AI's Stagnation Problem With Continuous Learning

The Static Model Bottleneck

Trajectory, a newly announced startup founded by researchers from Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs, has secured $15M in seed funding at a $115M post-money valuation to address a fundamental limitation in modern AI: models stop improving once training concludes. According to Wired AI, the round was led by venture firm Conviction, with participation from Bessemer Venture Partners, Radical VC, BoxGroup, and individual investors including Google DeepMind Chief Scientist Jeff Dean and Stanford professor Fei-Fei Li, CEO of World Labs.

The core premise is straightforward but historically difficult to execute. Deploying an AI model in production generates a stream of user interactions, errors, and corrections—a feedback signal that today goes largely unused. Instead of incorporating these insights back into the model through post-training, companies release static systems that make identical mistakes across sessions. This gap between potential and practice has become a recognized impediment to AI progress. At the December 2025 NeurIPS conference, Turing Award winner Richard Sutton argued that continual learning is essential for building superintelligent agents, lending theoretical weight to the problem Trajectory aims to solve.

Why Coding Succeeded Where Others Faltered

The existence of viable examples provides a roadmap. Wired AI notes that coding-focused AI products such as Cursor have already implemented early versions of continual learning, using real data about user interactions to perform post-training and regularly ship model improvements. CEO and cofounder Ronak Malde, previously an AI researcher at Windsurf before joining Google DeepMind as part of a $2.4B acquisition of the coding startup in 2024, attributes coding’s rapid adoption partly to this feedback-driven iteration cycle. The verifiability of code—it either executes or it doesn’t—creates a clear signal for retraining.

Other domains lack this certainty. Writing quality, customer service responses, and design decisions resist binary evaluation. This ambiguity has forced most companies outside coding to accept static models as a trade-off between complexity and deployment speed. Trajectory’s founding team—which includes Arjun Karanam, a former Apple researcher who worked on Vision Pro, and Michael Elabd, formerly in Google DeepMind’s robotics division—aims to abstract the platform layer so that teams can build continual learning systems without solving evaluation from first principles.

Why This Matters

The implications extend across two fronts. For practitioners, continuous improvement unlocks competitive advantages in markets where model refresh cycles currently take months or years. For research, solving continual learning at scale addresses what has historically been a laboratory problem—how to make systems learn from experience without catastrophic forgetting or distribution shift. If Trajectory’s approach generalizes beyond coding, it could accelerate the timeline to AI systems that adapt dynamically to their deployment context, rather than remaining frozen snapshots of training-time performance. The research challenge is real, but the commercial incentive—every AI vendor’s desire to out-iterate competitors—suggests the problem may finally have institutional weight behind it.