How an AI Startup Plans to Help Tech Companies Improve Their Models in Real Time

A group of AI researchers from Google DeepMind, Apple, OpenAI, and Meta announced this week the launch of Trajectory, a startup aimed at helping companies continuously improve their AI systems by learning from real users. The company is led by CEO and cofounder Ronak Malde, who previously worked as an AI researcher at Windsurf.

The startup has backing from respected figures in the AI world, including Jeff Dean, a chief scientist at Google DeepMind, and Fei-Fei Li, a Stanford professor and CEO of World Labs. Their support suggests that the industry sees value in what Trajectory is trying to tackle.

The Problem With Today's AI Models

Most AI models today follow the same pattern: they are trained once on a fixed set of data, deployed to users, and then periodically retrained with new information. Think of it like a car that gets maintenance on a schedule — it works fine, but it cannot adapt smoothly between service visits.

This approach has a real drawback. As the world changes and user behavior shifts, the model's performance can slowly decline. If a recommendation system was trained on last year's user preferences, it may not work as well six months later. Retraining takes time, resources, and planning, which means there is always a gap between when a problem appears and when a fix is deployed.

Trajectory's answer is to enable what researchers call continual or online learning — the ability for models to adapt in real time based on what users are actually doing. Instead of waiting for a scheduled retraining cycle, the model learns and improves continuously, even as it is serving users.

This sounds straightforward but is technically challenging. When a model learns from new information, there is a risk it will forget what it learned earlier — a problem computer scientists call catastrophic forgetting. The system also needs to handle data quality, protect user privacy, and guard against bad inputs that could make the model perform worse. These are not simple engineering problems, which is why most AI systems today still rely on the traditional scheduled retraining approach.

Why This Matters Now

The timing of Trajectory's launch fits a broader shift happening in the AI industry. After the initial wave of excitement around generative AI — chatbots, image generators, and the like — companies are now focused on making AI systems more practical and reliable. They want systems that actually improve over time, not ones that stay frozen until the next retraining cycle.

We have seen this pattern before. When cloud computing first arrived, the industry moved from simple rental servers to sophisticated platforms that automatically scaled resources based on demand. Now, AI infrastructure is going through a similar evolution. Trajectory is entering a crowded field. Companies like Weights & Biases, MLflow, Modal, and Anyscale are all working on tools to help companies manage and improve their AI systems. Trajectory will need to show that it offers something meaningfully different, especially in the specific area of continuous learning, in order to build a lasting business.

What the Platform Would Need to Do

Trajectory has not publicly shared detailed technical specifications, but the problem it is solving tells us what the platform likely needs to handle. It would need to accept streaming data from user interactions, build in safeguards to prevent the model from getting worse, and monitor performance changes. It also needs to update models quickly enough that users do not see delays in getting responses.

There are also practical questions about tracking and accountability. If a model is constantly changing, how do companies know which version made which decision? In regulated industries like finance or healthcare, this is critical for compliance. Companies need to audit what happened and why — you cannot do that easily if your model is learning and updating throughout the day.

Integrating with what companies already use is another challenge. Most organizations have existing systems for managing and training AI models. A new platform would need to fit in with these tools while remaining flexible enough to work with many different types of models.

What Comes Next

The potential market for continuous learning systems is large. Any industry using AI that benefits from staying up to date could use this — recommendation systems that learn what you like, fraud detection that adapts to new scams, or chatbots that improve with each conversation.

But adoption will face real obstacles. Companies that already have working systems for retraining models may not want to switch to something more complex. Continuous learning systems require more infrastructure and operational care, and many organizations will need to see clear proof of benefit before making the change. Industries with strict regulations may require extensive testing before allowing systems that essentially modify themselves.

The bigger picture here is that AI is moving from static products to adaptive ones. Models that learn over time could eventually provide better and more personalized experiences. But they also introduce new operational risks that teams will need to understand and manage carefully.

What determines whether Trajectory succeeds or fails is whether they can make continuous learning simple enough for companies to actually use, while proving it solves real problems companies face. If they do, AI systems could become genuinely smarter over time, rather than slowly degrading until someone gets around to retraining them.