Search Deepnet One

Down the rabbit hole. On purpose.

Machines that show their working.

Teaching models how to think.

Results may vary. These didn't.EngramSmall neural models that reason one checked step at a time and say so when they cannot.32×Parity, beyond trained lengthTrained on 8 bits, exact up to 256.5,000×A5, beyond trained lengthTrained two deep, exact 10,000 deep.81.8%Sudoku-Extreme, 0 wrongThe rest abstained. None wrong.0Kept errors in 52 million episodesNot one wrong step kept.engram.one ↗
Sutton and Barto, compiled.RelearnA templated, header-only C++ library for reinforcement learning, with documentation, examples and tutorials.GitHub Pages ↗Iron condors by conversation.AssistatronConnect Claude, ChatGPT, Gemini or Grok to a dedicated iron condor engine and a regulated broker.assistatron.com ↗

Notes from the rabbit hole

  1. Algorithmic AI Trading

    Seventeen years of algotrading

    Twenty-six years of watching the market, seventeen of trying to beat it, the last few with AI. This is the record so far: what I built, what the market thought of it and what is running today.

    3 Nov 2026 · 15 min read

  2. Algorithmic AI Trading

    Predicting Rare Events from a Growing Text History

    A forward-looking prediction problem that is really a timeseries: point-in-time profiles, an ever-growing text space, and why the hard part was leakage and out-of-sample validation, not the model.

    24 Jun 2026 · 11 min read

  3. Algorithmic AI Trading

    Tiny BERT in Production: Millisecond Text Classification on a CPU

    Four independent classifiers converged on the same recipe: a 2-to-4 layer BERT quantized with OpenVINO, running in single-digit milliseconds on a CPU with no GPU. How, and why the data split mattered more than the model.

    18 May 2026 · 14 min read

  4. Algorithmic AI Trading

    When the Simpler Model Won

    What a failed financial VAE taught us about representation learning. Across four model families and six trained models, the simplest model reconstructed best, and every layer of added sophistication made it worse.

    30 Apr 2026 · 10 min read

All posts →

AlgotradingSeventeen years of algotradingTrying to make the market pay for my models, stubbornly. The numbers so far.17years of trying10,665,733news articles hoarded4.3msper article, end to endThe story →