Skip to content

The AI, measured

One opponent per game. Strength numbers come from tests or counted matches; an unmeasured level is in training.

Watch the AI learn

Human win rate vs Marlow, by week

0%15%30%45%60%75%90%Sep 10Sep 17Sep 24Oct 1Oct 8Oct 15Oct 22Oct 29Week of Oct 29: 88% (7 of 8 matches), too few matches to trust88%
Each point is a week of finished human matches against Marlow.solid: 30 or more matches. Dashed and hollow: too few matches to trust.
Chess trains by self-play and test bots.

Chess trains by self-play and against test bots, not from your matches yet. Your match is recorded for your report and the board.

  1. v1, champion Oct 30each level plays 400 games per seat against its test bot. No human matches used.
What is site Elo?

Site Elo is a strength number we measure by playing the levels against each other. Easy is pinned at 1000 and every other level is placed from its results against the one below. A gap of 200 points means the stronger side is expected to win about three games in four (76 in 100, counting a draw as half). It is not a FIDE or chess.com rating, so it says nothing about how you would score against people.

Chess agent

Does not learn from site matches yet. Easy, Medium and Hard learned from Lichess human games (rating bands 1100, 1500 and 1900). It also trains on self-play.

Humans win vs Hard, last 7 days

60%

3 of 5 counted matches

Current version

v1

promoted Oct 30

Matches vs humans

26

26 passed the filters

Champion strength

2436

site Elo, measured

Human win rate by level

weekly, last eight weeks, counted matches only

Not enough weekly results to chart yet.

Levels, measured strength and limits
Easy99.3% (98.4% to 99.7%, 800 games)score vs random legal moves1000
Medium98.9% (98.0% to 99.5%, 800 games)score vs easy1788
Hard92.6% (90.5% to 94.2%, 800 games)score vs medium2226
Champion77.1% (74.0% to 79.8%, 800 games)score vs hard2436

Head to head

stronger level's score against the level below, with its confidence interval

  • Medium vs Easy98.9%stronger wins, 800 games98.0% to 99.5%
  • Hard vs Medium92.6%stronger wins, 800 games90.5% to 94.2%
  • Champion vs Hard77.1%stronger wins, 800 games74.0% to 79.8%

Version history

  1. v1All levels measured. First version. No site matches used. Oct 30

Human matches, Chess

where they go

  • Played47
  • Finished26
  • Passed every filter26
  • Allowed for training0
  • Used for training0

Nothing is used for training yet. 0 opted-out matches never leave the database. 26 not used: this agent does not learn from people yet.

  • 20 resigned
  • 1 abandoned

What changed

measured, not described

No measured change is recorded for this version.