Statistics and Computer Science at Harvard. I spend my time figuring out where machine learning can be applied: physics, politics, and more.
Finding gravitational lenses in wide-field sky surveys, and getting the models to survive the jump from simulation to real telescope data. Google Summer of Code.
Founding engineer. I build the forecasting models and the pipelines underneath them.
Statistics, with a concurrent master's in computer science. Quantitative Traders, AI Safety Team, poker.
Long-context models throw away tokens to save memory, then need them again. We put them back between turns — 91% retrieval where the baseline gets 24.5%, at the same cost.
Can you tell how long a model should think just by watching what it says? We built the benchmark to test it honestly. Mostly you can't — and the usual evidence that you can turns out to be noise.
A forecasting agent has to choose what to read. Relevant isn't the same as useful, and it can often tell the difference before opening anything.
Reads a live table and solves for the right play in under a second.
Traces industrial PFAS discharge downstream to the fish you'd actually eat. 3,274 river segments, built in 16 hours.
Sorted 300,000 galaxies and pulled out the 8,500 strange ones — mergers, starbursts, junk.