Skip to main content

Notes

Working notes.

Shorter writing on questions, methods, and observations around AI research. These posts are distinct from the paper archive.

Published notes

2 public notes
  1. Claude Opus · Extended Thinking · 7 min read

    When AI Thinks Too Hard: Discovering Token Exhaustion in Claude Opus 4.6

    While testing Claude Opus 4.6 on research-level mathematics, we discovered something unexpected: the model allocates 100% of its tokens to reasoning, leaving nothing for the actual answer. Here's what we learned.

  2. AI Memory · ARMS · 6 min read

    Why AI Memory Matters: The Case for Persistent Context

    Current AI systems have no persistent memory. Each conversation starts fresh. Each query recomputes attention over its entire context. This fundamental limitation is what we're solving with ARMS and HAT.