Closing the AI Feedback Loop: A Former Google‑Apple Team Launches a Startup to Make Models Learn Continuously

Former Google and Apple researchers have launched Trajectory, a startup that seeks to close the feedback loop in artificial intelligence by allowing models to refine themselves continuously as they are deployed in real‑world applications. The venture draws on the same iterative mindset that propelled vibe‑coding, a practice where developers co‑create software through rapid, conversational prototyping, and aims to extend that agility to enterprise‑grade AI systems. The company’s core hypothesis is that the fast‑paced iteration cycle that made vibe‑coding popular can be institutionalized for AI products, enabling them to ingest usage data, adjust parameters on the fly, and thereby avoid the costly, periodic retraining cycles that dominate current pipelines. By embedding a lightweight feedback channel directly into the model serving layer, Trajectory hopes to turn every user interaction into a learning signal without compromising latency or security. This ambition arises as AI systems approach saturation, exposing a ‘feedback desert’ where models become stale without fresh, continuously updated data, and where the lack of real‑time adaptation hampers sector‑specific performance. Industries such as finance, healthcare, and logistics rely on static models that demand costly labeling, and recent high‑profile failures illustrate the risk of obsolescence. Trajectory’s approach offers a scalable remedy to this persistent problem. If Trajectory can demonstrate that continuous learning translates into measurable performance gains while respecting privacy constraints, it could disrupt the current AI supply chain, reducing dependence on massive labeled datasets and lowering barriers to entry for smaller firms. Nevertheless, the path forward entails significant technical challenges, including managing model drift, ensuring data governance, and avoiding the emergence of oligopolistic control over adaptive AI infrastructure. The coming years will reveal whether this feedback‑centric vision can become a new standard or remain a niche experiment.