Former Google and Apple Researchers Launch a Startup to Build AI’s Missing Feedback Loop

A radical of AI researchers who antecedently worked astatine Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs announced connected Wednesday they’re launching a caller startup called Trajectory, which aims to assistance companies regularly amended their AI products by grooming connected real-world idiosyncratic interactions.

Trajectory wants to physique a level for AI that tin larn continuously, a capableness that researchers person agelong held up arsenic a large obstruction to further AI progress. OpenAI, Google, and Anthropic person recovered occurrence grooming progressively susceptible versions of AI models, particularly for domains specified arsenic coding, math, and science. However, these systems halt getting smarter aft their grooming is done. While determination person been immoderate caller breakthroughs successful continual learning, tech companies person mostly struggled to marque AI products that larn from their errors successful existent time. In December 2025 astatine NeurIPS, 1 of the largest yearly AI probe conferences, Turing grant victor Richard Sutton argued that continual learning is indispensable for gathering superintelligent agents.

Trajectory has raised a $15 cardinal effect circular astatine a $115 cardinal post-money valuation, led by the task superior steadfast Conviction, with information from Bessemer Venture Partners, Radical VC, and BoxGroup. Individual investors besides participated successful the round, including Google DeepMind’s main scientist, Jeff Dean, arsenic good arsenic the alleged “godmother of AI,” Stanford prof and World Labs CEO Fei-Fei Li.

Trajectory’s CEO and cofounder Ronak Malde was antecedently an AI researcher astatine Windsurf, and helium aboriginal became 1 of lone a fistful of employees who went to enactment astatine Google DeepMind erstwhile it hired the coding startup’s apical endowment successful a $2.4 cardinal woody past year. The different cofounders of Trajectory see Arjun Karanam, a erstwhile AI researcher astatine Apple who worked connected the Vision Pro, and Michael Elabd, who antecedently worked successful Google DeepMind’s robotics division.

Malde tells WIRED that immoderate starring AI coding products, specified arsenic Cursor, are already doing an aboriginal mentation of continual learning—using existent information astir however radical interact with their products to bash post-training and regularly vessel exemplary improvements. He argues this is simply a halfway crushed wherefore AI coding products person taken disconnected truthful rapidly, and is portion of the crushed wherefore large AI labs person rushed to make vibe coding applications of their own. With Trajectory, Malde and his squad of 11 researchers and engineers anticipation to use a akin method for improving AI-powered tools extracurricular the coding space.

“Even the astir almighty AI contiguous is inactive static. The AI exemplary that you utilized yesterday is going to marque the aforesaid mistakes today,” says Malde. “A mates companies are starting to get to that satellite of continual learning. What we are doing is gathering the level for each azygous institution to get to continual learning.”

The situation with applying this logic to different domains is that coding is easy verifiable—code either runs oregon it doesn’t—but immoderate industries person looser definitions of success. Karanam says portion of what Trajectory’s level offers is helping optimize an AI exemplary to a business's circumstantial needs.

Rather than starting from an off-the-shelf exemplary from OpenAI oregon Anthropic, Trajectory has customers statesman with an open-source exemplary that has been post-trained for a circumstantial AI merchandise the institution has successful mind. For Decagon, a lawsuit that builds AI lawsuit enactment agents, Trajectory logs erstwhile its AI falls short—say, a lawsuit trying to marque a instrumentality gets their query bounced to a human—and uses those instances to post-train a caller exemplary arsenic often arsenic each week. Trajectory claims these post-trained models bushed the frontier labs’ models connected constrictive tasks that substance astir for a company’s product.