A New Company Aims to Help AI Systems Get Better On Their Own

A group of former Google DeepMind, Apple, OpenAI, and Meta researchers announced Wednesday the launch of a new startup called Trajectory. The company aims to help other businesses make their AI tools smarter by having those tools learn directly from how real people use them. The company is led by CEO Ronak Malde, who previously worked as an AI researcher at Windsurf.

The startup has backing from some well-known figures in AI research, including Jeff Dean from Google DeepMind and Fei-Fei Li, a Stanford professor who also runs a company called World Labs. Their support shows that other AI experts see real value in what Trajectory is trying to solve.

The Problem: AI Systems That Stop Improving

Here's a common problem in AI today: companies build and train an AI system, put it to work, and then find it doesn't improve much after that. The traditional way works like this: a company takes old data, trains a system on it, launches the system, waits months or longer, then retrains it with new data. That gap means the AI sits largely frozen in between, even as the real world changes and users interact with it.

Trajectory wants to change this. The company is building tools that let AI systems learn and improve continuously as people use them, rather than waiting for big retraining cycles. Instead of snapshots, the system adapts as it goes.

This sounds straightforward but is surprisingly hard to pull off. When an AI system keeps learning on new feedback, there's a real risk it will accidentally forget things it once knew well. There are also real concerns about data privacy, preventing bad actors from tricking the system into learning the wrong things, and simply not letting the system get worse instead of better. These are genuine engineering puzzles, which is why most AI systems in use today still rely on the old batch retraining method.

Why Now?

The timing makes sense. AI companies have spent the last couple of years rushing to build and launch AI tools. Now that some of that initial work is done, they're asking: how do we keep these systems working well over months and years? How do we make them better without shutting them down to retrain?

This mirrors something we've seen before in technology. When cloud computing first arrived, companies just rented basic compute power. Later, the industry figured out how to build smarter systems that could automatically adjust to usage patterns and needs. The AI industry seems to be at a similar turning point now.

The list of investors backing Trajectory signals confidence in the team and the market opportunity. Jeff Dean has been central to building many of Google's core AI systems over decades. Fei-Fei Li spent years leading computer vision research at Stanford and has deep roots in the AI community. Their involvement carries real weight.

There's a growing crowd of companies offering different kinds of AI infrastructure tools. Companies like Weights & Biases, MLflow, Modal, and Anyscale are all building pieces of the AI tooling puzzle. Trajectory will need to show it does something genuinely different and useful, especially in the specific area of systems that learn on their own, to succeed in this crowded space.

What This Could Mean

The opportunities for continuous learning are broad. Recommendation systems on streaming services could get better by watching what users actually watch. Fraud detection systems could adapt to new scams in real time. Customer service chatbots could improve based on conversations. Any AI system that learns from feedback stands to benefit.

But adoption won't be automatic. Companies that already have working AI systems may be hesitant to swap them for something more complex and unfamiliar. The extra infrastructure and ongoing maintenance work could be hard to justify without proof it's worth doing. Regulated industries like finance and healthcare might move slowly until they see evidence these systems are safe and auditable.

The bigger picture here is about how we think of AI over time. Right now, most AI systems are somewhat static—they get built, they get launched, and they stay mostly the same until someone rebuilds them. If Trajectory and similar companies succeed, AI could become more like a living tool that adapts and improves as you use it, rather than a fixed tool that slowly falls behind the world. That shift could make AI genuinely more useful, though it also brings new challenges that teams will need to manage.

Whether Trajectory succeeds will depend on whether it can take a genuinely complex technical problem and make it simple and practical for ordinary businesses to use. If it can do that, it could help unlock a new generation of AI systems that don't just sit still after launch, but actually get smarter over time.