Former Google and Apple Researchers Launch a Startup to Build AI’s Missing Feedback Loop
Trajectory, founded by ex-DeepMind, Apple, OpenAI, and Meta researchers, is building tools to retrain AI from real user interactions.
Intelligence analysis by GPT-5.4 Mini
Trajectory says it can help companies turn real-world AI usage into regular model improvements. Its pitch is that AI products should learn from mistakes after deployment, not stay static after training.
Trajectory is trying to make AI more like a student that keeps learning after class. Instead of getting trained once and then staying the same, it watches how people actually use it and learns from mistakes.
That matters because many AI tools are a bit like a toy robot that can only follow the same instructions every day. If it keeps messing up the same way, it never improves unless people go back and fix it.
The startup says it can help companies improve their AI tools every week, and maybe someday much faster. The big idea is simple: AI should get better from experience, not just from one big lesson at the beginning.
Analysis
What Trajectory is building
Trajectory is a new startup from former researchers at Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs. Its core idea is straightforward: AI products should not stay frozen after training. Instead, they should learn from real-world user interactions and get better over time.
Why continual learning matters
The article frames this as a major gap in today’s AI systems. Models from OpenAI, Google, and Anthropic can become very capable, especially in coding, math, and science, but they usually stop improving once training ends. The startup is trying to close that gap by building a platform for continual learning, which researchers like Richard Sutton have argued is important for more advanced AI agents.
How the product works
Trajectory’s approach starts with an open-source model that is post-trained for a specific customer use case rather than relying on a generic frontier model. The company then logs moments when the AI fails or falls short, such as a support case being handed off to a human, and uses those examples to retrain the model regularly. In one example, the company says it can update models as often as weekly.
Who is using it
The startup says it already has customers including Clay and Harvey, and it plans to expand beyond AI-native companies to the Fortune 500. It has raised a $15 million seed round at a $115 million post-money valuation, led by Conviction with support from Bessemer Venture Partners, Radical VC, and BoxGroup, plus investors including Jeff Dean and Fei-Fei Li.
The open question
The article notes a limitation: weekly retraining is not true continual learning in the strict sense, because the model still stays static between updates. Even so, Trajectory argues the direction is clear, and it wants to move toward daily or even per-interaction updates.
Key points
- Trajectory was founded by former researchers from Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs.
- The startup aims to make AI products learn from real user interactions after deployment.
- It raised a $15 million seed round at a $115 million post-money valuation.
- Trajectory says it already works with companies including Clay and Harvey.
- The company’s current updates are weekly, which critics may say is not yet true continual learning.
