Blog
Notes and experiments on software, engineering, and machine learning.
- Introducing Agentflare
A cloud space for building with agents
- The Termination Circuit (how reasoning models stop thinking).
Reasoning models since the dawn of o1 and R1 have a tendency to overthink. Despite a lot of work on early-exit methods and steering, open-weight and smaller reasoning models still produce long chains of thought before they answer. I worked on discovering how much of that thinking is required before the model already knows the answer and what makes the model stop thinking.
- Linear Probes add little for Verifiable Reward Hacking
Linear probes achieve near-perfect AUC at detecting reward hacks but add almost nothing over simple output checks. An experiment with GRPO on Qwen2.5-0.5B shows the limits of mechanistic interpretability when hacks are easily verifiable from the output.
- Notes on 'A Mathematical Framework for Transformer Circuits'
My notes and understanding based on Anthropic's 'A Mathematical Framework for Transformer Circuits' paper.
- How much does a code model know about what the code does?
A progress report on a side project: I built a small pipeline to read the hidden states of code models, then probed four sizes of Qwen2.5-Coder. Scale makes them settle the grammar earlier in the network and barely changes how well they track what a program does.
- Hello World
My first blog post where I share what this blog will be about.