Notes

Working notes

Long-form write-ups here; the systematic material lives in three note repositories on GitHub.

Write-ups

1

The scaling result I published, and why I retracted it

64× on 128 cores is 50% efficiency — and two findings I had not expected.

Read the note

Note repositories

3

Planned

Listed as intentions, not as work already done.

  • A per-question error analysis of the rule router's failures.
  • Off-policy targets inside an on-policy learner, worked through.
  • Multi-node runs to settle what the falling efficiency actually is.