A graded trajectory is a complete recording of an AI’s attempt at a task plus a verdict on whether it succeeded, and it’s the fundamental unit of modern AI training data.
A free grader is any mechanism — passing tests, physical reality, paying customers — that produces those verdicts automatically at zero marginal cost, and since feedback cost caps practice volume, free graders determine how fast a model can learn.
The data flywheel is the compounding loop this creates (better model → more users → more graded trajectories → better model), and it’s strategically decisive because while model designs can be copied in months, an accumulated corpus of graded real-world trajectories cannot be copied at all — which is why companies now compete less over algorithms and more over owning the places where work happens and the mechanisms that grade it.
A trajectory is just a recording of an attempt at a task, from start to finish. Not the final answer — the whole journey: every step the AI took, every click, every line of code, every wrong turn and correction. Think of it like the dashcam footage of a driving lesson rather than just the exam result.
A graded trajectory is that recording plus a score: did the attempt succeed or fail? The grade is what turns a pile of recordings into training data, because now the AI can learn “do more of what the successful attempts did, less of what the failures did.” This is the core of how modern AI models improve through reinforcement learning: no grade, no learning signal.
A grader is whatever produces that score. And here’s the key economic insight: graders range from very expensive (paying a human expert to review every attempt) to completely free — where the world itself tells you the answer automatically. Three big free graders exist: execution (run the code — it either passes the tests or it doesn’t), physics (the robot either picked up the cup or dropped it; the car either stayed in the lane or didn’t), and markets (the customer either accepted the work and paid, or rejected it). Nobody has to be hired to produce these verdicts; they’re a byproduct of the task happening in the real world.
Digital Optimus will be an attempt to give the Tesla-XAI stack a market grader it lacks, by deploying screen agents into real workflows where “did the ticket resolve, did the human accept the output” scores the trajectory. Combined with Cursor (execution grader) and the vehicle/robot fleet (physics grader), the design on paper assembles all three — which is why it’s strategically coherent even if the execution has wobbled.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
I’ve been looking (passively) for a framework for thinking about how AI will progress for years and I’ve only been able to figure out where we end up but not the path. This article is the first I’ve seen with a really good structure for what gets mastered by AI first. Really good article.
glad you like it and find it useful