Every post in this series so far has been about the model lane of the TPU Model Performance Auto-optimization project: an autoresearch loop that proposes a hypothesis, patches model code, benchmark...

For the TPU Model Performance Auto-optimization project introduced in the first post of this series we proposed to use an LLM wiki to store information about model optimizations, experiment results...

In the first post I walked through the TPU Model Performance Auto-optimization repo — autoresearch, LLM wiki, xprof MCP, and per-experiment branches — and showed that Claude Code with Opus 4.7 coul...

Andrej Karpathy’s autoresearch is a genuinely brilliant idea. I was able to adapt it to the model-performance-optimization domain, and it proved to work in its original form — see my earlier write-...

I’d like to make a bold statement: given a sufficiently capable LLM, the right profiling tools, and a knowledge base that includes the model’s + framework’s source, an autonomous agent can drive an...

Recently I had to debug a performance issue in TensorFlow V1.x input pipeline. TensorFlow V2.x has a nice tensorboard profiler plugin which is super handy for such situations. Unfortunately it does...

It so happened that October turned out to be a month of Haskell for me. Accidentally I found a new Haskell MOOC by University of Glasgow. I got curious mostly because University of Glasgow is known...

While playing with F# I found that I’m really missing intellisense support for FSI interactive. It is very useful for exploring APIs and language features. I know, “official” suggestion is to write...

Idea to create a dev blog crossed my mind long time ago, but I didn’t get to finally do it until now. Hope someone will find my thoughts and experience useful. Today I want to share my experience ...