Building towards Superintelligence
Defining quality standards and curating the training data. The work that decides what the model learns to value.
I work on training data and evaluation at Meta. I write and build around world models, embodied intelligence, and the line between a useful tool and a mind.
I am Jesse He. I have been a commercial pilot, CPA, data engineer, and data scientist. Now I work on the training data and evaluations behind language and video models at Meta.
My work keeps leading me back to the same question: what would make a machine intelligent, not just useful?
I do not think prediction alone gets us there. Intelligence needs stakes. Something to fear. Something to long for. A world where action has consequences.
I write Decision Height to test that belief against the craft. Data. Evals. World models. Public experiments. I also share what I learned from taking the long way into AI.
Sometimes I help people and teams with hard, specific problems.
For teams: training data, evaluation, dataset diligence, and turning a vague concern into a test.
For people: making an unconventional move into serious AI work.
I keep this limited. Send me the decision you are trying to make and what you have already tried.
Defining quality standards and curating the training data. The work that decides what the model learns to value.
AI strategy for people who build real things. Written from inside frontier labs, not from the sidelines.
A few mentorship slots a quarter. Mostly folks breaking into AI at senior levels or considering a pivot.
An interactive cost-of-ownership model for the general-aviation fleet: set how you fly, compare any two aircraft, and find the planes nobody beats on both cost and speed.
Open Plane Math ↗Post-training · RL datamix · Data quality
PPL · CPL · Instrument Rated
Language models · AI systems
Robotics · OpenClaw · Papers · Businesses
Pilot before college. CPA at 23. Uber during its most chaotic years. Berkeley MIDS while shipping Llama and MetaAI to billions. Now: Meta Superintelligence Lab.
Read the flight plan →