I'm building the World Transition Model for systems that change.

Large language models learned from the world recorded in words. Abel is my attempt to learn from the world recorded in transitions: prices and order flow, power grids and energy demand, supply chains, macroeconomies, populations, and sensor streams from anything whose state evolves over time.

Different surfaces, one deeper problem: given the current state of a system, its history, its context, and the forces acting on it, what changes next — and what changes if you intervene?

I do not believe real predictive power comes from forecasting alone. It comes from the whole stack: information, representation, compression, conditioning on the right covariates, simulation, and causal structure. Correlation can make a model look brilliant in a backtest. Causality is what lets it survive regime change, intervention, and the quiet death of easy patterns.

But causality is not the whole model. A world model must also represent state, compress high-dimensional signals, reason over regimes, simulate possible futures, and connect prediction to action. The goal is not merely to say what comes next. The goal is to understand how systems respond when people act, environments shift, and variables are changed.

I'm proving Abel first in financial markets because markets are the most unforgiving dynamic system: liquid, reflexive, adversarial, and scored in real money in real time. Markets are the crucible, not the destination. If Abel can model transitions there, the same foundation can extend to other systems that evolve, react, recover, and adapt.

Prediction is not the product. It is the middle layer.
Decisions are the endpoint.

The path to Abel started in mathematics. Olympiad math carried me into Fudan's mathematics department without sitting the national entrance exam; in 2011 I took first place in the Shanghai division of China's national undergraduate mathematical modeling contest. Then Maryland, for a PhD in computer science under Tim Oates and Yun Peng, one of the pioneers of artificial life and multi-agent systems. Mathematical intuition on one side, machine intelligence on the other — that pairing runs through everything I've done since.

At Maryland I worked on teaching neural networks to see time as structure rather than noise. In 2015 I proposed encoding time series as images — Gramian Angular Fields, Markov Transition Fields — and then used FCN and ResNet for time-series analysis and classification. Published under Zhiguang Wang, those two lines of work became the standard baselines of the field — nearly every method since has been measured against them — and are closing in on 10,000 citations.

Then a decade on the production side of frontier AI: principal researcher and research director roles across Microsoft Research, Meta FAIR and Reality Labs, and Snap, taking successive generations of chatbots and large-model systems from research into products serving hundreds of millions of people — with a string of US patents along the way. Later, US CTO of Amber Group, a US-listed company.

I write about prediction, time, intelligence as compression, causality, markets, and where AI is actually going — not where the consensus insists it is going.