← All games

Nine matches against the machine

Backgammon, draughts, chess, Othello, Jeopardy!, Go and poker: one passport each

What this page is

Nine times a machine was put in front of a person who was among the best in the world at something, in front of an audience, with the result written down afterwards. Eight of the nine the machine won. This page lists them; each one has a page of its own behind it, and that page is where the figures live.

The passports are all cut the same way, so a backgammon exhibition in 1979 and a Go experiment in 2023 can be read against each other line by line: who sat down, what stood on the other side, what it drew, what it cost, how long it was given to think, and — the part the pages exist for — what the result proves and what it does not.

Every figure carries two things beside it and neither is a footnote: how good the figure is (measured, specified, reconstructed, or good only for an order of magnitude) and the boundary it was measured inside. Watts of silicon and watts at the wall are two different quantities, not one quantity at two precisions. Where nobody ever published a number, the page says so and explains what was published instead. That happens often.

BKG 9.8 vs Luigi Villabackgammon · 1979
Luigi Villa — BKG 9.8 · Carnegie-Mellon University
On 15 July 1979 Hans Berliner's program won a seven-point match 7-1 against the freshly crowned backgammon world champion. It was the first time a machine beat a reigning world champion at a board game, and Berliner himself put the result down to the dice.
the machine won the match · the match passport →
Chinook vs Marion Tinsleydraughts · 1994
Marion Tinsley — Chinook · University of Alberta
In August 1994 Jonathan Schaeffer's program drew all six games it played against Tinsley, after which Tinsley withdrew on health grounds and handed over the title. Two years earlier in London the same Tinsley had beaten Chinook 4-2 with 33 draws.
the machine won the match · the match passport →
Deep Blue versus Garry Kasparovchess · 1997
Garry Kasparov — Deep Blue · IBM
The rematch, 3–11 May 1997: six games, 3½–2½ to the machine. A year earlier Kasparov had beaten the previous version 4–2.
the machine won the match · the match passport →
Logistello vs Takeshi MurakamiOthello · 1997
Takeshi Murakami — Logistello · NEC Research Institute
From 4 to 7 August 1997 Michael Buro's program won all six games against the reigning Othello world champion: 6-0 in games, 264-120 in discs. It ran on an ordinary desktop machine with a single Pentium Pro under Linux.
the machine won the match · the match passport →
Watson on Jeopardy!Jeopardy! · 2011
Ken Jennings and Brad Rutter — Watson (DeepQA) · IBM Research
A two-day game, 14–16 February 2011, against the show's two most successful champions. Watson won; IBM gave the million-dollar prize away.
the machine won the match · the match passport →
AlphaGo versus Lee SedolGo · 2016
Lee Sedol — AlphaGo (the version DeepMind later named AlphaGo Lee) · Google DeepMind
Five games, 9–15 March 2016, won 4–1 by the program. Lee Sedol took game four, which remains the only game a human has won from AlphaGo in a formal match.
the machine won the match · the match passport →
Libratus versus four professionalspoker · 2017
Jason Les, Dong Kim, Daniel McAulay, Jimmy Chou — Libratus · Carnegie Mellon University; computing at the Pittsburgh Supercomputing Center
Twenty days, 11–30 January 2017, and 120,000 hands of heads-up no-limit hold'em. The machine finished 147 milli-big-blinds per hand ahead at 99.98% statistical significance, and beat each of the four men separately as well.
the machine won the match · the match passport →
Pluribus at a six-handed tablepoker · 2019
Thirteen professionals in the five-humans-and-one-AI format; Chris Ferguson and Darren Elias in the one-human-five-copies format — Pluribus · Carnegie Mellon University and Facebook AI Research
Two experiments of 10,000 hands each, six-player no-limit hold'em. Against five professionals the machine won 48 milli-big-blinds per hand; against one human and five copies of itself, 32. The blueprint strategy took 12,400 core hours and, by the authors' own arithmetic, about $144 of cloud rental.
the machine won the match · the match passport →
Kellin Pelrine versus KataGoGo · 2023
Kellin Pelrine — KataGo and the third-party build JBXKata005; separately, Leela Zero · FAR AI, MIT and UC Berkeley for the attacking team; KataGo itself is David Wu's engine
In February 2023 a human won 14 of 15 games against the strongest bot then available on KGS, and beat KataGo and Leela Zero running 100,000 visits per move. The strategy he used was not his own discovery: an adversarial policy trained against KataGo found it.
the human won the match · the match passport →

What each win cost

In every one of these matches one side could be counted and the other could not. BKG 9.8 was a share of a time-shared PDP-10 and one author. Chinook in 1994 was an eighteen-processor Silicon Graphics server and a set of endgame databases. Deep Blue was four hundred and eighty chess chips fabricated for this one task, thirty nodes to drive them, five engineers and a grandmaster on the payroll. Watson was ninety servers in the room next to the studio. By 2019 the whole machine side of a poker result was two processors.

The bars put the machine side on a scale of powers of ten. What is counted is whatever the source actually published — nodes for Watson, processors for Pluribus, chips for AlphaGo — and the unit is printed beside every bar, because normalising them would mean inventing a core count for a 1979 PDP-10 and then forgetting it had been invented. The human bar is the same in every row, because it was the same in every row.

The human, any of the nineone brain, 20 WR
Deep Blue, chess, 1997510 nodes and purpose-built chipsM
KataGo, Go, 2023no configuration was ever published
Red is the human side, blue the machine. Sources and methods are on each match's own page; the tag on every bar says how good the count is. Two rows are worth reading twice. AlphaGo in Seoul is 48 first-generation tensor processors, which is DeepMind's own figure, published eighteen months after the match in the AlphaGo Zero paper; the widely quoted 1,920 processors and 280 graphics cards is the largest line of a scalability table in the 2016 Nature paper and was never the Seoul machine. KataGo in 2023 has no bar at all: the bots stood on a public Go server that belonged to somebody else, and nobody published what they ran on.

The point of the column is not that the human was better after all. Nothing in these matches was measured in cost. They compared playing strength, playing strength is what they compared, and on that measure the results are exactly what the scoresheets say. A win with a thousand processors is a win.

It is worth writing the cost down because it is usually left out, and because leaving it out quietly turns nine results about games into one claim about people. The two columns are not the same kind of quantity. One says how well the moves were chosen. The other says what it took to choose them — and only one of the two has been falling for forty years.

The human column

On any chart the human side needs at least two lines: the brain (20 W) and the person at the board (105–126 W). A single line against the machine's watts is not a comparison, it is a choice of answer. And neither line is truly commensurable with the machine's: the machine's number excludes its training, and the human's excludes the twenty years it took to learn the game.

the brain at rest · one organ
20 WR
The adult brain at rest. A physiological floor, not the price of a single move.
basal metabolism · the whole body at rest
60–90 WR
The whole body at rest — basal metabolic rate. An adult lying down, fasted, thermally comfortable.
a person at the board · the whole body, seated
105–126 WR
A person sitting at a board. If watts are going to be set against watts, this is the honest figure to use.
what thinking adds · the whole brain, task-evoked
no figure was published
The increment from hard mental work. There is no reliable figure, and that is itself the finding: the increment is small enough to be hard to measure at all.

These figures are the same on every passport, because the human side of the table was the same in all nine matches. None of them is a stopwatch reading of a person playing: they are the physiological floor and the physiological cost of sitting somewhere for a few hours. What they leave out is the twenty years it took to learn the game — the same thing the machine figures leave out when they count the match and not the training.

Where to go from here
To cite this page: “Nine matches against the machine”, You vs Computer, https://youvscomputer.com/games/matches.html