The Palo Alto startup is raising around $50M to fix what its founders call AI's toddler-level visual understanding.
A team of researchers who helped build some of the most advanced AI systems at Google DeepMind and Apple have left to start their own company. Trajectory, a new Palo Alto-based startup, is focused on making AI systems that can actually see and interpret the physical world, something current models handle about as well as a three-year-old handles a juice box.
The company is targeting roughly $50 million in seed funding to build multimodal AI systems with dramatically better visual reasoning. Its core bet: the rapid iteration cycles that turbocharged the “vibe-coding” movement in software development can be applied to train AI products that learn continuously across all kinds of industries.
The team behind Trajectory #
Trajectory’s founding roster includes Andrew Dai, who spent more than 14 years at Google DeepMind and led data and pre-training efforts for the Gemini model family. Yinfei Yang, formerly the chief research scientist at Apple, rounds out the industry side. Seth Neel, drawn from Harvard’s AI research community, brings the academic perspective.
Their thesis is straightforward. Large language models have gotten remarkably good at processing and generating text. But when it comes to visual understanding, interpreting images, video, spatial relationships, and the messy physical world, performance remains shockingly primitive. In English: your AI assistant can write a legal brief but struggles to tell you whether a shelf is about to fall over.
Trajectory wants to close that gap by building feedback loops that let models iterate on visual data much faster than current training pipelines allow. The founders have described their approach as borrowing from the vibe-coding playbook, where developers use rapid, almost intuitive cycles of building and testing to ship products at speed. Applied to visual AI, the idea is that models should be learning continuously from new data rather than waiting for the next massive training run.
Why this matters beyond Silicon Valley #
The immediate applications Trajectory is targeting include robotics and physical systems. Think warehouse robots that need to navigate cluttered environments, autonomous vehicles parsing complex intersections, or manufacturing lines that need to spot defects in real time.
What investors should watch #
A roughly $50 million seed round is substantial but not outlandish by 2026 AI startup standards. What sets Trajectory apart is the specificity of its focus. Rather than building yet another general-purpose chatbot or coding assistant, the company is targeting a well-defined technical bottleneck, visual and multimodal reasoning, where progress has genuinely lagged.
Coverage of Trajectory’s funding efforts first surfaced in The Information in January 2026, with Bloomberg following up in April. The company has not indicated any involvement with blockchain technologies or digital assets.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our