Former Google and Apple Researchers Launch Trajectory for Continuously Learning AI

Former Google and Apple researchers have founded Trajectory, a startup building AI systems that learn continuously from user data. Their core conviction: a fast iteration cycle can solve one of the biggest problems in modern AI.

Why the current approach falls short

Most AI products share one fundamental flaw: no fast feedback loop. The typical process looks like this: researchers train a model on historical data, the product team deploys it to production, and the company then waits months for enough new data to accumulate for retraining. Meanwhile the model degrades, users run into errors, and the data distribution shifts.

Trajectory considers this fundamentally wrong. If a system can learn in real time, it can adapt almost instantly to new scenarios and errors.

Inspiration from rapid software development

Just as a developer can make a change, push it to production and get feedback within hours, an AI system should be able to update its weights based on live user data. Instead of a cycle of "months of planning, training, deployment," the gap between spotting a problem and fixing it would be minutes.

Benefits of this approach:

• The model updates in real time based on new data
• Errors are found and fixed in hours, not months
• Adaptation to the specifics of each user or customer
• Lower risk of model degradation in production
• Lower costs for retraining and redeployment

What this could look like technically

Trajectory is working on an architecture in which the model does not merely make a prediction but also absorbs the outcome of that prediction at the same time. That requires solving several nontrivial problems.

First, data validation. How do you tell useful signal from noise? If a user clicks a button by accident, that should not train the model.

Second, quality control. How do you keep the model from starting to learn from its own mistakes?

Third, scalability. How will this architecture work when millions of users generate data at the same time?

Companies that manage to solve these problems will gain a huge advantage in iteration speed and product quality.

What this means for the industry

If Trajectory's vision proves feasible, it could redefine what it means to "deploy AI in production." Instead of a one-time investment in training a model, companies could ship AI products as living organisms that grow and adapt to new realities. That will require new tools, new best practices and a new development culture, but the potential gains are huge.