Former Google and Apple Researchers Launch a Startup to Create Artificial Intelligence for AI

A team of AI researchers who previously worked at Google DeepMind, Apple, OpenAIand Meta Superintelligence Labs announced Wednesday they are launching a new startup called The waywhich aims to help companies continuously improve their AI products by training real-world users.

Trajectory aims to create an AI platform that can learn continuously, an ability that researchers have long held as a major barrier to AI progress. OpenAI, Google, and Anthropic have found successful studies of AI models, especially in fields such as literature, mathematics, and science. However, these systems stop being smart after graduation. Although there have been recent developments leading to continuous learningTech companies often struggle to create AI products that learn from their mistakes in real time. In December 2025 at NeurIPS, one of the biggest annual AI conferences, Turing award winner Richard Sutton said continuous learning is essential for the most intelligent builders.

Trajectory has raised $15 million in seed funding at a cost of $115 million, led by the venture capital firm Conviction, backed by Bessemer Venture Partners, Radical VC, and BoxGroup. Individual investors participated in the round, including the chief scientist of Google DeepMind, Jeff Dean, and the so-called “godmother of AI,” Stanford professor and CEO of World Labs Fei-Fei Li.

Trajectory’s CEO and cofounder Ronak Malde was previously an AI researcher at Windsurf, and later became one of the few employees who went to work at Google DeepMind when it hired top startup talent. A $2.4 billion deal last year. Among the founders of Trajectory is Arjun Karanam, a former AI researcher at Apple who works for the company. Vision Proand Michael Elabd, who has worked in the past The robotics division of Google DeepMind.

Malde tells WIRED that some advanced AI products, such as Cursor, are already doing continuous learning — using it. the reality of how people interact and their products to perform after training and sending frequent sample corrections. They say this is the main reason AI products have taken off very quicklyand it’s one of the reasons why big AI labs exist rushed to create a vibe coding application their own. With Trajectory, Malde and his team of 11 researchers and engineers hope to apply a similar approach to controlling AI-powered devices outside of the writing environment.

“Even the most powerful AI today is still static. The AI ​​model you used yesterday will make the same mistakes today,” says Malde. “A number of companies have started learning continuously. What we are doing is building a platform for every company to learn continuously.”

The problem with applying this principle to other domains is that validation is easily determined – the code runs or not – but other industries have looser definitions of success. Karanam says one of the Trajectory platform’s offerings is helping to optimize the AI ​​model for business needs.

Instead of starting with an off-the-shelf model from OpenAI or Anthropic, Trajectory has customers starting with an open source model that was trained after another AI product the company envisions. For Decagon, a client that builds AI customer assistants, Trajectory records when its AI is down — say, a customer trying to get a refund gets their question to a human — and uses that experience to train a new model on a weekly basis. Trajectory says that once these models are trained, they outperform frontier labs on narrow tasks that are critical to the company’s products.