Former Google and Apple Researchers Launch Trajectory for AI With Continuous Learning

Former researchers from Google and Apple founded the startup Trajectory to create AI systems that continuously learn from user data. Their main belief: fast iteration cycles can solve one of the key problems facing modern AI.

Why the Current Approach Doesn't Work

Most AI products suffer from a fundamental limitation: the lack of a fast feedback loop. The typical process looks like this: researchers train a model on historical data, the product team deploys it to production, and then the company waits months for new data to accumulate for retraining. During this time, the model degrades, users encounter errors, and data distribution shifts. Trajectory believes this is fundamentally wrong. If a system can learn in real time, it can adapt almost instantly to new scenarios and errors.

Inspiration from Fast Development

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

Advantages of this approach:

• Model updates in real time based on new data
• Error detection and correction in hours, not months
• Adaptation to each user's or client's specific needs
• Reduced risk of model degradation in production
• Reduced costs for retraining and redeployment

How This Could Work Technically

Trajectory is working on an architecture where the model doesn't just make a prediction, but simultaneously learns from the result of that prediction. This requires solving several non-trivial problems. The first is data validation. How do you distinguish useful signal from noise? If a user clicks a button by accident, that shouldn't train the model. The second is quality control. How do you prevent the model from learning on its own errors? The third is scalability. How does this architecture work when millions of users are generating data simultaneously? Companies that can solve these challenges will gain a huge advantage in iteration speed and product quality.

What This Means for the Industry

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