Research
Longer-form write-ups of investigations and experiments. Most run on my home lab; some are work-adjacent without representing my employer.
- Teaching an open model to say how sure it is A single forward pass, one LoRA epoch, and a size study from 0.8B to 27B. We turned open Qwen models into typed decision models that return calibrated probabilities for yes/no, pick-one and rating questions in one forward pass, with no generated text. One LoRA epoch makes every model from 2B up accurate and well calibrated on familiar tasks. Calibration on task types it never saw is where size matters, and where the work is still unfinished.
- A JEPA world model that learns Mario by playing itself Fifteen closed-loop iterations, a 3σ false positive, and what we're rebuilding before iter 16. We trained a self-supervised JEPA world model on NES Super Mario Bros and ran a closed-loop flywheel — collect, retrain the world model, train a policy in imagination, repeat — for fifteen iterations. The headline iter looked like a breakthrough; a five-seed re-run revealed it was a draw from a much wider distribution than we'd been measuring. This is the project so far, and the redesign that comes next.
- Sport-transfer SSL — what scaling up actually changed A follow-up: capacity, resolution, fine-tuning, and a real cross-sport win. After the first writeup, three follow-up questions kept nagging. With more compute we tested them all — and one downstream task we hadn't measured before flipped the story.
- Sport-transfer self-supervised learning — what we learned What we learned trying to teach a vision model about sports. A one-week investigation into whether self-supervised pretraining on in-domain video improves transfer to new sports tasks — and what those experiments taught us about how to actually measure SSL progress.