<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom" xml:lang="en">
  <title>Speculative Decoding</title>
  <subtitle>Working notes on markets and machines.</subtitle>
  <link href="https://marketneutral.github.io/feed.xml" rel="self" type="application/atom+xml"/>
  <link href="https://marketneutral.github.io/" rel="alternate" type="text/html"/>
  <id>https://marketneutral.github.io/</id>
  <author>
    <name>Jonathan Larkin</name>
  </author>
  <updated>2026-11-11T00:00:00.000Z</updated>
  <entry>
    <title>How LLM-based AI beats its human teachers</title>
    <link href="https://marketneutral.github.io/posts/exceed-training-data/" rel="alternate" type="text/html"/>
    <id>https://marketneutral.github.io/posts/exceed-training-data/</id>
    <published>2026-11-11T00:00:00.000Z</published>
    <updated>2026-11-11T00:00:00.000Z</updated>
    <summary>A machine that only watches people can beat every one of them. No expert in the data. No practice. Here is the smallest proof I could build.</summary>
    <category term="ESSAY"/>
    <author>
      <name>Jonathan Larkin</name>
    </author>
  </entry>
  <entry>
    <title>autoresearch: sparks of an AI scientist</title>
    <link href="https://marketneutral.github.io/posts/ai-scientist/" rel="alternate" type="text/html"/>
    <id>https://marketneutral.github.io/posts/ai-scientist/</id>
    <published>2026-06-01T00:00:00.000Z</published>
    <updated>2026-06-01T00:00:00.000Z</updated>
    <summary>Karpathy&#39;s autoresearch hands the &quot;what should I try next&quot; step of science to a language model. It works on any problem with a score, and it is the start of a new way to do research. Finance should be excited, and careful.</summary>
    <category term="ESSAY"/>
    <author>
      <name>Jonathan Larkin</name>
    </author>
  </entry>
  <entry>
    <title>qex: a small language for quant alphas</title>
    <link href="https://marketneutral.github.io/posts/qex/" rel="alternate" type="text/html"/>
    <id>https://marketneutral.github.io/posts/qex/</id>
    <published>2026-02-08T00:00:00.000Z</published>
    <updated>2026-02-08T00:00:00.000Z</updated>
    <summary>Quant alphas in one line. No lookahead by construction, every expression aligned and NaN-safe, shared work computed once across thousands of signals, and backtests built in. A closed grammar small enough for an agent to search.</summary>
    <category term="PROJECT"/>
    <author>
      <name>Jonathan Larkin</name>
    </author>
  </entry>
  <entry>
    <title>The evolution of the hedge fund</title>
    <link href="https://marketneutral.github.io/posts/hedge-fund-evolution/" rel="alternate" type="text/html"/>
    <id>https://marketneutral.github.io/posts/hedge-fund-evolution/</id>
    <published>2025-10-02T00:00:00.000Z</published>
    <updated>2025-10-02T00:00:00.000Z</updated>
    <summary>Notes from a guest lecture at MIT on two trends reshaping hedge funds — unbundling the workflow across a crowd, and pairing human judgment with machine forecasts.</summary>
    <category term="TALK"/>
    <author>
      <name>Jonathan Larkin</name>
    </author>
  </entry>
  <entry>
    <title>Endowment asset allocation as a reinforcement learning problem</title>
    <link href="https://marketneutral.github.io/posts/endowment-rl/" rel="alternate" type="text/html"/>
    <id>https://marketneutral.github.io/posts/endowment-rl/</id>
    <published>2023-12-12T00:00:00.000Z</published>
    <updated>2023-12-12T00:00:00.000Z</updated>
    <summary>Can a Q-learning agent learn to run an endowment portfolio of stocks, bonds, and illiquid private fund commitments, and beat a simple rules-based baseline? TL;DR: yes.</summary>
    <category term="PROJECT"/>
    <author>
      <name>Jonathan Larkin</name>
    </author>
  </entry>
</feed>
