rsi.watch

Benchmarks

Three groups now publish RSI measurements. This page says what each one measures. For their results, go to their pages; rsi.watch does not rank models.

RSI Index · Vals AI

Launched Aug 2026 · page updated 2026-10-01

  • Four tasks: Compression (enwik8), LM Training, Harness Engineering, and Post-Training.
  • Each task is scored on a log scale: 0 is the starting baseline, 0.5 a published, human or frontier-model reference, 1 the theoretical best. The index is the mean of the four.
  • It ranks models.

Their page

RSI Bench · Scale AI

Launch post 2026-08-07 · final 50 tasks due Nov 1

  • Tests whether AI models can handle the stages of frontier AI development themselves.
  • Publishes individual graded runs. When we checked, there was no aggregate model leaderboard.
  • Open call for tasks: $2,000 per accepted task plus co-authorship. The code is MIT-licensed.

Their page

DeepRSI · Hillclimb

Announced on X about 2026-10-12

  • Announced by co-founder Jun Park as a measure of recursive self-improvement over 48 hours on ML tasks.
  • We did not find a public leaderboard page. Task count and scoring method appear only in secondary coverage, so they are not listed here.

Their page

Checked 2026-10-14.