thedatasense.comCircuit Tracing: Finding Medical Features in Gemma 3Language models can answer medical questions with surprising accuracy. But do they actually encode medical knowledge in identifiable, interpretable ways? Or is it all just statistical soup? Using Neuronpedia, we ran a simple experiment to find out. W...Feb 4·4 min read
thedatasense.comWhat Does Medical VLM Actually See? Experiments with MedGemma and Sparse AutoencodersWhen a medical Vision Language Model(VLM) looks at a chest X-ray and says "cardiomegaly present," what's actually happening inside the model? It's a black box. Billions of parameters. Dense activation vectors where every dimension encodes a tangled m...Jan 24·5 min read
thedatasense.comFluent But Wrong: LLM and HealthcareOff late a lot of my research time is studying why medical models systems fail. Not the obvious failures where the model outputs gibberish, but the subtle ones where the output looks clinically appropriate, follows proper documentation structure, use...Jan 20·9 min read
thedatasense.comOpening the Black Box: How to See What Your Vision Language Model is Actually Looking AtWhen a doctor examines a chest X-ray and says "I see signs of pneumonia in the lower right lung," you can ask them to point at exactly what they're seeing. They can circle the cloudy region, explain why it looks abnormal, and walk you through their r...Jan 16·19 min read
thedatasense.comData Generating ProcessData does not just appear. Something creates it. A coin flip. A measurement device. A biological process. A human decision. Understanding that something, the mechanism that generates observations, is the key to understanding uncertainty. This mechani...Jan 14·7 min read