Uncertainty Is Not A Compass
I trained a probe to find where an LLM's reasoning forks. It learned something real, but knowing where a model is uncertain doesn't tell you where to search.
Matthew Bragonje
Engineer & Experimentalist
I like probing machine learning systems until I understand why they actually work — causal inference, production ML, and the occasional LLM hacked open to see what's inside. I write about those investigations.
Latest Posts
View All →-
-
Hacking better answers out of a 0.6B model
Exploration of a novel method of improving the performance of an 0.6B model without touching weights
-
What Does Target ROAS Actually Do?
I built a simulated ad market to explore how bidding and pricing interact, and why individually rational choices can leave every competitor worse off.
-
Trying to Solve Capybaradoku With Linear Algebra
An attempt to solve a hexagonal puzzle with linear algebra became an investigation into its constraints, reducing the search space from billions of billions of possibilities to just 244 placements.