On Wednesday, a team of AI researchers with prior tenures at Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs launched a new startup called Trajectory. The company’s core mission is to help businesses steadily upgrade their AI products by leveraging ongoing training on real-world user interactions.

Trajectory is building a platform that enables AI to learn continuously — a capability long cited as one of the biggest remaining barriers to advancing mainstream AI development. While leading research labs including OpenAI, Google, and Anthropic have successfully built increasingly powerful AI models, particularly for specialized fields like coding, mathematics, and scientific research, these systems stop improving once their initial training phase wraps up.

Even with recent breakthroughs in continual learning research, most large technology firms have yet to overcome the challenge of building production AI products that learn from their mistakes in real time. In December 2025, Turing Award winner Richard Sutton emphasized at NeurIPS (one of the AI research community’s largest annual conferences) that continual learning is a non-negotiable requirement for building superintelligent agents.

The startup has closed a $15 million seed funding round, valuing it at $115 million post-money. The round was led by venture capital firm Conviction, with backing from Bessemer Venture Partners, Radical VC, and BoxGroup. High-profile individual investors also joined the round, including Jeff Dean, chief scientist at Google DeepMind, and Fei-Fei Li — often called the “godmother of AI,” who is a Stanford professor and CEO of World Labs.

Trajectory’s CEO and co-founder Ronak Malde previously worked as an AI researcher at coding startup Windsurf. Last year, Google DeepMind acquired Windsurf’s top engineering and research talent in a $2.4 billion deal, and Malde was one of the few employees who moved over to DeepMind as part of the transaction. The startup’s other co-founders are Arjun Karanam, a former Apple AI researcher who contributed to the Vision Pro project, and Michael Elabd, who previously held a research role in Google DeepMind’s robotics division.

Malde told WIRED that leading AI coding tools like Cursor already test an early iteration of continual learning: they use real-world data on how users interact with their products to run post-training updates and roll out steady model improvements. He argues this approach is a core reason AI coding products have grown so rapidly in popularity, and it’s also why major AI labs have raced to launch their own “vibe coding” tools in recent months. With a current team of 11 researchers and engineers, Malde and Trajectory aim to bring this same continuous improvement framework to AI-powered tools across industries far beyond coding.

“Even the most powerful AI available today is still static,” Malde said. “The AI model you used yesterday will make the exact same mistakes today. A small number of companies are starting to build out this world of continual learning. What we’re doing is building the infrastructure platform that lets every company access this capability.”

The biggest hurdle to expanding this approach beyond coding is that coding results are easily verifiable: a piece of code either works as intended or it doesn’t. Many other industries have far fuzzier definitions of what counts as a successful AI output. Karanam explains that a core offering of Trajectory’s platform is helping tailor and optimize AI models to each business’s unique specific requirements.

Instead of having customers build on top of off-the-shelf models from OpenAI or Anthropic, Trajectory has clients start with open-source base models that are post-trained specifically for their unique AI product use case. For one client, Decagon (which builds AI-powered customer support agents), Trajectory logs instances where the AI fails — for example, when a customer trying to process a return has their request incorrectly escalated to a human agent — and uses these real failure cases to post-train an updated model as frequently as once a week. Trajectory claims these custom, regularly updated models outperform leading frontier models from top labs on the narrow, business-critical tasks that matter most to their clients’ products.

Corporate leaders are eager to deploy AI for a wide range of business tasks, but today that often requires hiring teams of “forward deployed engineers” — specialized consultants and technical staff embedded within a company to build and maintain custom AI products. Firms including OpenAI, Anthropic, and Palantir have moved quickly to meet this demand. Elabd says Trajectory’s goal is to build a self-improving product that eliminates the need for in-house engineering teams to constantly troubleshoot and update a company’s AI stack. The startup already counts clients across a range of industries, including enterprise sales startup Clay and legal AI startup Harvey. While Trajectory currently focuses primarily on working with AI-native companies, it eventually plans to roll out its platform to Fortune 500 enterprises.

Some critics point out that Trajectory has not yet delivered true continual learning, at least not as the term is traditionally defined in academic research. For now, the startup’s models only update once per week, and remain fully static between these scheduled updates.

Elabd counters that Trajectory is still in its early stages. He argues the entire AI industry is shifting toward a new paradigm where AI learns continuously from real-world experience, mirroring the trend already playing out in the AI coding space. Trajectory’s long-term goal is to build a platform that can update a company’s AI models every day — or even more often.

“Every day might not be frequent enough long-term,” Elabd said. “It could end up being every hour, it could be after every single user interaction. It’s also possible that every company doesn’t need just one monolithic AI — you could train a custom AI that learns for every individual employee at every company.”

This piece is an edition of Maxwell Zeff’s Model Behavior newsletter. Read previous newsletters here.