AdderBoard

AdderBoard, maintained by Dimitris Papailiopoulos on GitHub, grew out of a prompt that asked Claude Code and Codex for the smallest possible addition transformer and got back 6,080 and 1,644 parameters. It admits only models that add two ten-digit numbers at 99%-plus accuracy (the launch framing was perfect addition), keeps two tables, weights learned from data and weights set analytically, and in this corpus is the leaderboard where a 121-parameter model hand-coded by Codex led before the record fell to 36 parameters a week later. The newsletter reads that drop, beside the NanoGPT speedrun’s fall to 88.1 seconds two days earlier, as a gauge of how fast capability density is compressing; the two leaderboards (NanoGPT speedrun and AdderBoard) are the corpus’s paired rulers for that compression.