Notes on AI systems, financial infrastructure, and production ML. Updated as new pieces are finished.
TGAT runs at 5–6% GPU utilisation. JODIE at 1.5–2.5%. DyRep under 2%. These are the published benchmarks. A March 2026 paper fixes the inference problem — and the speedup numbers are remarkable.
A May 2026 paper feeds Soros's theory of self-reinforcing market dynamics to GPT-5, Claude, and Gemini across two boom-bust cycles. The forecasting improves. The reason why is harder to establish than the paper suggests.
Most academic papers on LLMs in finance report directional accuracy and call it done. A 2026 IEEE review asks what a fund manager would actually want to know.
An assumption baked into every forecasting model we build — and a February 2026 paper that formalises why it's wrong.
Why compliance and risk are the most underinvested, highest-leverage AI problem in global markets — and what the stack should actually look like.