Kasparov versus the World — computer analysis

Chessgames.com game link

August 2026 run

Every candidate move tested from moves 6 to 39 was searched for two hours on a dual AMD EPYC 9654-class machine (192 cores, 384 threads, 503 GB RAM, complete 6-piece Syzygy tablebases in memory). The 73 main searches examined about 200 billion nodes each. They reached depths of 21 to 118 ply overall; the non-losing candidates reached depths of 60 to 118. Every main search was given the same time, and the transposition table was cleared before each one, so the results were obtained under directly comparable conditions.

The raw log is here: kvw-2026-runs.jsonl (JSON Lines, one record per search). It contains the 73 main searches, three completed follow-up searches in the 16…Ra5 17. Nb5 line, and one follow-up stopped after 31 minutes as being of minor importance. Every completed record gives the evaluation, depth, node count, tablebase hits and full principal variation.

The three decisive moments

The public-analysis feedback loop

The objective sequence at moves 37 and 38 has an important historical context. The World Team's analysis was public, and after the game Kasparov acknowledged that he had followed the discussion on MSN and used it to his advantage. At this particular moment, however, the information may have worked in both directions.

Before move 37 the bulletin board had examined 37…e5, but much of the analysis feared 38. Bc1. The World therefore chose 37…e6??, which the present run shows was the losing move. The World Team was also preparing for the expected 38. Rd1 and had concluded that Black's only hope was the desperate 38…Ke4. Long contemporary lines appeared to give Black a draw, but later analysis found that White wins by a full-board zugzwang.

Kasparov said that he had intended to play 38. Rd1, but that over dinner a defensive move for Black came to him which he could not refute. Because he had admitted consulting the World Team's analysis, it has long been supposed that its 38…Ke4 line influenced his decision. There is no direct evidence that this is what happened, so it should remain a strong circumstantial explanation rather than a fact. Ken Regan's contemporary and post-game account is here.

The resulting irony is striking: the World Team's mistaken public analysis first hurt the World by steering it away from drawing 37…e5, then may have rescued it one move later by frightening Kasparov away from winning 38. Rd1. The game was not simply Kasparov against a crowd; it was a feedback system in which each side could be affected by the crowd's evolving beliefs.

Analyst scorecard

The four expert analysts who advised the World Team in 1999 — Irina Krush, Elisabeth Paehtz, Florin Felecan and Etienne Bacrot — each posted a recommendation per move. Krush was the principal advisor; the World Team usually voted in her choice.

The muted columns are seven further contemporary voices, recovered from the Internet Archive and from my own 1999 files: the St Petersburg GM Chess School; correspondence GM Roberto Alvarez’s Ajedrez de Estilo; Barnet Chess Club; the Computer Chess Team (the volunteer engine-analysis group I belonged to); the Finnish IM Antti Pihlajasalo, posting as “IM2429”; Canadian GM Duncan Suttles; and IM Georgi Orlov’s chessplayer.com column. The four analysts’ per-move commentary is preserved in gameanalysis.txt. “(absent)” means no pre-vote recommendation survives; “(none)” that the source explicitly made none.

Green = drawing; red = losing; shaded cell = the move the World Team played. (=) means the engine returned exactly +0.00; (draw) in the endgame rows means proved drawn by tablebase. Moves 1–4 were outside the 2026 run and carry no evaluation. Moves 51 onward are 6- and 7-piece endgames, where the verdicts are tablebase-exact rather than engine estimates; 39…Ke4 is plainly lost and was not re-run.

How the analysts did

• Krush found a drawing move at every single point from move 6 to move 36, and again at 53 — then picked the loser at 37 and 54. The two times she went wrong are exactly the two times the game was decided.
• Paehtz was the only one of the four to find 37… e5, the sole drawing move at the game's critical instant. She also went wrong more often than anyone else: losing recommendations at 19, 25, 32, 36 and 53.
• Felecan drew at 6, 7, 10, 15, 18, 25, 26, 52, 53, 54; lost at 19, 21, 36, 37, 51.
• Bacrot drew at 6, 7, 10, 15, 16, 18, 19, 21, 25, 26, 54; lost at 36, 37, 52. Absent at 30–33 (winning the French Championship at Besançon) and at 51 and 53.

Two patterns stand out. At the two fatal moments Krush chose the losing move both times, while at move 36 — where three of the four analysts went wrong — only her Kd5 held. And among all eleven sources, only Paehtz, Suttles and Orlov named 37…e5.

Discovery was not the same as selection

The scorecard is not a simple demonstration of “wisdom of crowds.” Krush served as the hub of the analysis network, and at move 51 the voters departed from her recommendation for the first time in 40 moves. At move 36 her drawing Kd5 defeated the losing b2 by only 37.69% to 37.11% of the vote — a margin of 0.58 percentage points. One move later, the majority view among the four official analysts was wrong: Paehtz alone found drawing 37…e5.

At move 54 the direction reversed. Paehtz, Felecan and Bacrot all supplied drawing queen moves, while Krush supplied losing b4. Across the nine surviving sources with a recommendation, five named Qd3 or Qd5 and four named b4; nevertheless b4 was played. The decisive failure was therefore not always a failure to discover the right move. At both Black blunders the correct move was already present in the network, but the authority-and-voting system did not select it.

Paehtz illustrates the value of dissent. She made more losing recommendations than any of the other official analysts, yet she was the only one of the four to solve the game's most important defensive problem. A system which filtered contributors only by their average record might have discarded the person holding the essential answer.

Kasparov's 17. Nb5

Kasparov writes that had the World played 16… Ra5, he intended 17. Nb5. Stockfish prefers 17. Bd2 (answering it with the surprising 17…Rh5), so his line was tested separately, two hours per position:

All equal, with the engine giving over 99% draw in each. Kasparov's intended move is perfectly sound, and 16…Ra5 wins nothing against it. An older run had 18. b3 at −0.21 and 18. Ra3 at −0.33; that apparent edge for Black disappears with more depth. These positions are not simple, though — the line after 18. Ra3 runs 95 ply and passes through 21…Rxb5 and 26. Re8+, both sides sacrificing material to reach a drawn queen ending.

Verdict

The World Team stayed on a drawing path until 37… e6??; Kasparov's 38. h6?? handed the draw back; and 54… b4?? finally lost it. Every other tested move in the critical set preserved the same game-theoretic outcome as the available alternatives, according to the two-hour searches or, from the seven-piece ending, exactly according to tablebases.

The historical lesson is different from the bare engine verdict. The World Team generated the correct answer at both of its losing turns but failed to select it, while its public analysis may have induced Kasparov's error between them. The match demonstrated the power of distributed analysis, but also the separate problems of correlated belief, authority and vote aggregation.