Some Data from LeelaPieceOdds (www.lesswrong.com)

I downloaded and graphed a large set of games against LeelaPieceOdds — a Leela chess build trained to play with piece-removal handicaps (options like N/B/R/Q combinations such as QN or BBNN). The dataset is split by time control (Bullet/Blitz/Rapid) and human Lichess rating; Leela always plays fast so time controls mainly affect the human. Results show Leela frequently overcomes substantial material disadvantages and often feels like playing “while tired”: humans blunder a lot and wins are rarer than intuition suggests. The author notes uneven data (a huge spike from one 1400–1499 player), low sample sizes in some buckets, and anecdotally winning slightly more than the aggregate suggests. Recommendation: pick odds where your rating wins <1/3 to get the most striking demo. Technically, the story highlights two clear levers to improve LeelaPieceOdds: more search/compute (deep search favors the machine) and per-player adaptation (modeling a specific opponent’s typical mistakes). Both would raise performance differently: search improves raw tactical/simulation strength, while personalization exploits systematic human errors. The author frames odds chess as a small analogue for AI “boxing” experiments — useful for probing control trade-offs (limited serial reasoning, information, practice) but not decisive evidence about escape risks; the experience nuances the intuition that structural advantages always beat intelligence differences.