Recursive self-improvement writing successor weights directly rather than searching for them. AdderBoard asks for the smallest transformer that adds two ten-digit numbers at 99% or better on a held-out test set; when the newsletter picked it up, the hand-coded table was led by a 121-parameter single-layer Qwen3-style decoder whose weights were set analytically by Codex, using tied embeddings, RoPE digit routing and a carry computed through the final norm (leaderboard). It is the constructive counterpart to AlphaEvolve’s search for activation functions and a step past Codex being instrumental in creating its own successor. The record did not last: 36 parameters a week later.
Development
A 121-parameter adder is hand-coded rather than trained
The author's framing: recursive self-improvement can now straight-shot the weights of a perfect successor model, writing them analytically instead of finding them by gradient descent.
Claim
The AdderBoard competition to find the smallest transformer that perfectly adds ten-digit numbers is led by a 121-parameter model whose weights were hand-coded by Codex rather than trained.
Reported metric: 121 parameters
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to February 25, 2026 (theinnermostloop.substack.com)
- AdderBoard: Smallest transformer that can add two 10-digit numbers (github.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.