ML Theory and Practice

From research to real-world deployment


When Bayesian Uncertainty Becomes Memory - A Path to Continual Learning

By inverting uncertainty into density, a model can recreate its own past and learn continuously without explicit replay buffers.

9 min read   ·   2025

On Anthropomorphizing Token Traces

My thoughts on the paper 'Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!'

10 min read   ·   2025