Own experiments, real data
The technical foundation behind the services: hands-on studies with real data, fixed seeds and reproducible setups — including what didn't work. Each post reads in 2–3 minutes on the surface; the full depth unfolds per finding.
RSS feed — new posts, no tracking.
Less bias, better transfer?
A frozen, self-supervised foundation model — trained only on everyday photos, never on medical images — detects intracranial hemorrhage on head CT at ~0.94 AUC. What drives the transfer isn't what I bet on.
Can a video-AI recipe learn language?
LeCun's newest self-supervised recipe trains video models at up to 20× less compute. I ported it to text and ran the controls. Result: a model that never reads more than 30% of any text learns better text representations than models that read everything — and the new invariance loss is not what makes that work.