Former Google and Apple researchers have launched a startup to build AI’s missing feedback loop

A group of AI researchers who previously worked at Google DeepMind, Apple, OpenAIand Superintelligence laboratories They announced on Wednesday that they will be launching a new startup called pathwhich aims to help companies systematically improve their AI products by training on real-world user interactions.

Trajectory wants to build an AI platform that can learn continuously, a capability that researchers have long viewed as a major barrier to further progress in AI. OpenAI, Google, and Anthropic have had success training increasingly capable versions of AI models, especially in areas such as programming, mathematics, and science. However, these systems stop becoming smarter after completing their training. While there were some recent ones Breakthroughs in lifelong learningTechnology companies have generally struggled to make AI products that learn from their mistakes in real time. In December 2025 at NeurIPS, one of the largest annual AI research conferences, Turing Award winner Richard Sutton argued that Continuous learning is essential To build super-intelligent agents.

Trajectory has raised a $15 million seed round at a post-cash valuation of $115 million, led by venture capital firm Conviction, with participation from Bessemer Venture Partners, Radical VC, and BoxGroup. Individual investors, including Google DeepMind’s chief scientist Jeff Dean, as well as the so-called “Godmother of AI”, Stanford University professor and World Labs CEO Fei-Fei Li, also participated in the round.

Trajectory CEO and co-founder Raunak Malde was previously an AI researcher at Windsurf, and later became one of the few employees who went on to work at Google DeepMind when it recruited top talent for the programming startup in A deal worth $2.4 billion last year. Other co-founders of Trajectory include Arjun Karanam, a former AI researcher at Apple who worked on Vision ProAnd Michael Al-Abd, who previously worked at Robotics division of Google DeepMind.

Maldi tells WIRED that some leading AI coding products, like Cursor, are already doing an early version of continuous learning — using… Real data about how people interact With their products to make post-training improvements and ship models regularly. This is the main reason behind it, he says AI coding products have taken off very quicklyIt’s part of the reason we have major AI labs He rushed to develop biometric barcoding applications Their own. With Trajectory, Maldi and his team of 11 researchers and engineers hope to apply similar technology to improve AI-powered tools outside of programming.

“Even the most powerful AI systems today remain static,” Malde says. “The AI ​​model you used yesterday will make the same mistakes today.” “Two companies are starting to access the world of continuous learning. What we are doing is building a platform for every company to access continuous learning.”

The challenge of applying this logic to other fields is that cryptography can be easily verified – either the code runs or it doesn’t – but some industries have loose definitions of success. Part of what Trajectory does is help refine the AI ​​model to a company’s specific needs, Karanam says.

Instead of starting from an off-the-shelf model from OpenAI or Anthropic, Trajectory asks customers to start with an open source model that is later trained on the specific AI product the company has in mind. For Decagon, a client that builds AI customer support agents, Trajectory logs when its AI fails — for example, a customer trying to generate revenue whose query bounces back to a human — and uses those instances to subsequently train a new model as often as each week. Track claims that these post-trained models outperform frontier laboratory models on narrow tasks of interest to the company’s product.