A new AI startup, Trajectory, has burst onto the scene with a mission to revolutionize how AI products learn and evolve. Founded by a stellar team of researchers from Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs, Trajectory aims to build a platform that enables continuous improvement for AI products based on real-world user interactions.
This initiative tackles a long-standing challenge in AI: the static nature of models after initial training. While giants like OpenAI and Google have made incredible strides in developing increasingly capable AI for tasks like coding and math, these systems typically stop learning once their training phases are complete. Trajectory seeks to bridge this gap, allowing AI to learn and adapt continuously from its experiences.
Funding and Founders: A Powerhouse Team
Trajectory has successfully secured a $15 million seed round, boasting a $115 million post-money valuation. This impressive funding was led by Conviction, with key participation from Bessemer Venture Partners, Radical VC, and BoxGroup.
The venture also attracted individual investors of significant repute, including Google DeepMind’s chief scientist, Jeff Dean, and the acclaimed “godmother of AI,” Stanford professor and World Labs CEO Fei-Fei Li. The founding team itself is a collection of AI luminaries. CEO and co-founder Ronak Malde previously worked at Windsurf before joining Google DeepMind in a major acquisition. He is joined by Arjun Karanam, an ex-Apple AI researcher who contributed to the Vision Pro, and Michael Elabd, formerly of Google DeepMind’s robotics division.
The Vision: Continuous Learning for All AI
Ronak Malde highlights that some leading AI coding products, such as Cursor, already employ an early form of continuous learning. They leverage real user interaction data for post-training, consistently shipping model improvements. Malde believes this real-time learning is a key factor behind the rapid success of AI coding tools, prompting major labs to develop their own applications in this space.
Trajectory’s 11-person team of researchers and engineers intends to apply this proven technique to AI-powered tools beyond the coding realm. “Even the most powerful AI today is still static. The AI model that you used yesterday is going to make the same mistakes today,” Malde explains. “A couple companies are starting to get to that world of continual learning. What we are doing is building the platform for every single company to get to continual learning.”
How Trajectory’s Platform Works
A core challenge in extending continuous learning beyond coding — where success is easily verifiable (code either runs or it doesn’t) — is defining success in other industries. Arjun Karanam notes that Trajectory’s platform is designed to optimize an AI model for a business’s specific needs, tackling this very problem. Instead of starting with generic, off-the-shelf models, Trajectory guides customers to begin with an open-source model that has been carefully post-trained for their specific AI product.
For example, with their customer Decagon, which creates AI customer support agents, Trajectory logs instances where the AI falls short — such as a customer’s return query being bounced to a human. These instances are then used to post-train a new model as frequently as every week. Trajectory claims these specialized, post-trained models consistently outperform frontier lab models on the narrow, critical tasks that matter most for a company’s product.
Empowering Businesses with Self-Improving AI
Many corporate executives are eager to integrate AI into diverse tasks, but current implementations often require hiring “forward-deployed engineers” or consultants to build and maintain AI products. Michael Elabd emphasizes that Trajectory’s goal is to create a self-improving product, thereby reducing the need for companies to retain in-house engineers for continuous AI troubleshooting.
Trajectory is already collaborating with customers across various fields, including enterprise sales startup Clay and legal AI startup Harvey. While currently focused on AI-native companies, Trajectory plans to expand its platform to Fortune 500 enterprises in the future. Though critics might argue that weekly updates don’t constitute “true” continual learning in the strictest sense, Elabd views this as just the beginning. He envisions an industry shift towards AI that learns from experience, with Trajectory ultimately aiming to enable daily, hourly, or even per-interaction model updates. “Every day may not be enough. It could be every hour, it could be every interaction,” says Elabd. “Maybe every company doesn’t need just one AI, you could train an AI to learn for every person at every company.”
Source: Wired – AI
