← All matches

Kellin Pelrine versus KataGo

Go · 2023 · the human won the match

Who played whom

The person: Kellin Pelrine
The machine: KataGo and the third-party build JBXKata005; separately, Leela Zero
Behind the machine: FAR AI, MIT and UC Berkeley for the attacking team; KataGo itself is David Wu's engine
Where: the KGS Go Server, ordinary online conditions
Game: Go, 2023
Result: the human won the match

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.

What is proved and what is not

These two lists are the reason the page exists. A match is a result about a particular series of games under particular conditions, and almost everything usually drawn from one does not follow from it.

Proved

  • A human, using no machine during the games themselves, beat engines at settings normally out of human reach: 14 of 15 games against JBXKata005, then the strongest bot on the KGS server, plus games against KataGo and Leela Zero running 100,000 visits per move.
  • Handicaps did not break it. Against the same JBXKata005 the human gave three, five and nine stones and won in every case.
  • The flaw turned out to be humanly legible. The adversarial algorithm found not a memorised move sequence but a strategy that can be explained, learned and executed by hand — and after publication other people repeated it against other engines.
  • Superhuman strength and robustness are different properties. The vulnerability had been in KataGo since 2019, thousands of amateurs and professionals had picked the program apart in that time, and nobody found it until an algorithm was set on it.

NOT proved

  • That a human out-played a machine at Go. He did not find the strategy: an adversarial policy trained against KataGo did — some 1.01 million self-play games and about 2,223 V100 GPU-days of computing, roughly 14% of KataGo's own training budget. The human learned the technique from that policy's game records. Machine help did not vanish from this story; it moved upstream of the games.
  • That there was no machine during the games. There genuinely was none: the authors write 'without any algorithmic assistance', and play took place under ordinary KGS conditions. But 'a human beat the machine' without saying where the strategy came from misleads, and 'a machine helped the human' without saying it was absent from the games misleads exactly as much.
  • That this was a match. It was not: no regulations, no time control, no independent referee, no distance announced in advance. The games were played on a public server, and the side that counted and published the result was the attacking side — Kellin Pelrine is a co-author of the paper and his team's resident Go expert.
  • That 'KataGo lost 14 of 15'. The fifteen-game series was played against JBXKata005, a third-party KataGo build unaffiliated with the authors. Of the games against KataGo itself and against Leela Zero at 100,000 visits the paper says they were won but gives no score, and names no network version for them either.
  • That humans are once again stronger at Go. The same adversarial policy that beats KataGo loses to amateurs, and KataGo at ordinary settings still beats the strongest people alive. What was found is not a general weakness but one specific blind spot: a misreading of groups with a cyclic shape.
  • That the hole has been closed. Since mid-December 2022 KataGo's training has included positions from this attack, and the 2022 version of the adversarial policy did all but stop winning against the updated network. But the authors show that the underlying vulnerability survives even in networks trained specifically to defend against it.
Sources: T. T. Wang et al., 2023; FAR AI, 2023; Financial Times (reported by Richard Waters), 2023

The hardware

Computing units
no figure was published

Nobody published the machine side's hardware: the bots lived on the KGS server, which belonged to somebody else and was outside the authors' control. What is on record is not nodes but search settings, and those matter more: KataGo and Leela Zero played at 100,000 visits per move — full search, not a bare policy network. This is the load-bearing detail. Without search, the adversarial policy beats KataGo in over 99% of games; with search the job is markedly harder, which is precisely why the human's games count for something. The opponent in the fifteen-game series was JBXKata005, an independent KataGo build unaffiliated with the authors and the strongest bot then playable on KGS; its settings are not given in the paper. The KataGo network used for the main machine experiments is named exactly — b40c256-s11840935168-d2898845681 — but no network number is given for the games against the human. This is where the imprecision that travelled through the press comes from: the Financial Times of 19 February 2023 ran the series as 'man beats machine at Go', when the fifteen games were played against a third-party build.

Sources: T. T. Wang et al., 2023; Financial Times (reported by Richard Waters), 2023

Energy

Every line here states the boundary it was measured inside, and the boundary is printed above the number rather than under it: eight hundred watts of silicon and eight hundred watts at the wall are not one figure at two precisions, they are two different figures. Training and playing are kept apart for the same reason and are never added: one was spent once, before the match; the other over the hours people sat at the table.

Boundary: playing the games
no figure was published

Not published.

where this number comes from

The games ran on the public KGS server, on other people's machines. What hardware turned JBXKata005, KataGo and Leela Zero is stated neither in the paper nor on the project site. The only measurable quantity on the machine side is the 100,000 visits per move used by KataGo and Leela Zero, and that describes work, not watts. Filling the gap with the TDP of some plausible graphics card would be invention.

Sources: T. T. Wang et al., 2023; FAR AI, 2023
Boundary: training the model
8,400–13,800 WR · reconstructed
where this number comes from

This is KataGo's own training, not the games. Reconstructed from the configuration: the KataGo repository documents the 'third major run' as 14 days on 28 V100 accelerators, 24 days on 36, and 119 days on 46 — quoted in the adversarial-attack paper, which totals it at 6,730 V100 GPU-days. The watt range is those same three phases: 28 × 300 = 8.4 kW at the start and 46 × 300 = 13.8 kW at the end, using the V100's nameplate TDP in SXM2 form (the PCIe card is 250 W). Those 6,730 days come to 40–48 MWh, accelerator silicon only — nothing for the machines holding them and nothing for cooling. The network actually attacked was trained longer: the authors give a lower bound of 15,881 V100 GPU-days, so not less than 114 MWh by the same arithmetic — but that is a conversion unit rather than real V100s, since the later training was distributed across volunteers' assorted hardware.

Sources: T. T. Wang et al., 2023; NVIDIA, 2018
Boundary: training the model
no figure was published
where this number comes from

This is the training of the attacking policy — the price of finding the very strategy the human later won with. The authors give the hardware actually used: 61.8 A4000 GPU-days, 348.6 A6000, 299.7 A100 40GB and 222.9 A100 80GB, which they themselves convert to 2,223.2 V100 GPU-days — about 14% of the compute that went into training KataGo, and roughly 1.01 million self-play games. Applying those accelerators' nameplate TDPs (140 W for the A4000, 300 for the A6000, 250–400 for the A100 40GB, 300–500 for the A100 80GB, the spread coming from PCIe and SXM variants the paper does not distinguish) gives something on the order of 6–8 MWh of silicon. This is an order of magnitude, not a measurement: how many accelerators ran at once is not stated, so there is no watt figure here, only a total.

Sources: T. T. Wang et al., 2023; NVIDIA, 2018

What it cost

Prices are held in the money of their own year and inside a stated boundary: the match hardware, the whole project and rented computation are three different sums, and which one is on offer is written above each.

What is counted: no boundary established
no figure was published

There was no money around these games at all: no purse, no sponsor, no stake — ordinary online play on a public server. FAR AI never published what training the attacking policy cost, and converting 2,223 V100 GPU-days into cloud dollars is not something this record will do: no provider and no rates are named, and any such figure would be our invention rather than a found number.

Sources: T. T. Wang et al., 2023

Time

Time control

ordinary online play on the KGS server; the paper does not state a time control

The person, per move
no figure was published
The machine, per move
no figure was published
Positions per second
no figure was published

Neither the time control, nor the human's time per move, nor the engine's appears in the paper — all three lines are empty. The one published number about the machine's thinking is 100,000 visits per move for KataGo and Leela Zero; that is a search budget, not positions per second and not seconds, and it cannot be turned into a rate without knowing the hardware. For scale, elsewhere in the paper KataGo plays at anything from 1 visit — that is, no search at all — to ten million.

Sources: T. T. Wang et al., 2023

The human side

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 · reconstructed

The adult brain at rest. A physiological floor, not the price of a single move.

where this number comes from

The twenty watts is a conversion, not a meter reading, and two independent routes arrive at it. One: Raichle and Gusnard (PNAS, 2002) give the brain as roughly 2% of body weight and roughly 20% of the body's oxygen and calorie consumption; 20% of an adult's basal rate (see body_resting, 60–90 W) is 12–18 W, and 20% of a seated person's rate (100–125 W) is 20–25 W. The other route is more direct: the classic Kety–Schmidt measurements put the brain's oxygen consumption at 3.3–4.0 ml per 100 g per minute (Sokoloff et al., 1955). For a 1,400 g brain that is 46–56 ml of oxygen a minute, and at an energy equivalent of about 20.9 J per ml it comes to 16–20 W. Both routes land on 16–21 W, and 20 is the rounding up. This does not earn an M: oxygen was measured, watts were not.

Sources: M. E. Raichle, D. A. Gusnard, 2002; L. Sokoloff et al., 1955
basal metabolism · the whole body at rest
60–90 WR · reconstructed

The whole body at rest — basal metabolic rate. An adult lying down, fasted, thermally comfortable.

where this number comes from

Converted from kilocalories per day. One kcal/day = 4,184 J / 86,400 s = 0.04843 W. The Mifflin–St Jeor equation (1990), now the working standard in place of Harris–Benedict: a man of 70 kg, 175 cm, 40 years gives 1,599 kcal/day, which is 77 W; a man of 80 kg, 180 cm, 30 years gives 1,780 kcal/day, 86 W; a woman of 60 kg, 162 cm, 40 years gives 1,252 kcal/day, 61 W. Hence 60–90 W, covering ordinary adults of either sex. This is lying down and fasted; a person sitting at a board spends more — see the next row.

Sources: M. D. Mifflin et al., 1990
a person at the board · the whole body, seated
105–126 WR · reconstructed

A person sitting at a board. If watts are going to be set against watts, this is the honest figure to use.

where this number comes from

From the metabolic heat production tables underlying the thermal comfort standards ISO 8996 and ASHRAE 55. The 'met' is defined as 58.2 W per square metre of body surface and describes a person seated at rest in thermal neutrality. The standard adult is 70 kg, 175 cm, with a DuBois surface area of 1.8 m². That gives: seated quiet, 1.0 met — 105 W; seated at office work, 1.2 met — 126 W. This is the body's total heat output, not the brain's share. A caveat: the ISO and ASHRAE tables descend from Passmore and Durnin's 1967 review and have not been revisited since. The spread among advisers — '70–100' against '100–120' — is not a disagreement but two different definitions: basal rate in the first case, a seated person in the second. '150–300 W' for someone sitting is simply wrong: on the same tables 150–210 W is standing and light physical work, and 300 W is walking.

Sources: International Organization for Standardization, 2021; ASHRAE, 2023
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.

where this number comes from

The most direct test was done seventy years ago and has not been overturned. Sokoloff, Mangold, Wechsler, Kennedy and Kety (Journal of Clinical Investigation, 1955) measured cerebral blood flow and oxygen consumption in 13 people at rest and while doing mental arithmetic. The EEG changed unmistakably during the arithmetic — a shift to low-voltage, high-frequency activity — and pulse and blood pressure rose, while the brain's oxygen consumption did not move: in their words, the control and arithmetic values were identical. The authors state the conclusion worth remembering the paper for: neither the degree of mental activity nor the EEG pattern correlates with total cerebral oxygen consumption. Modern imaging did not overturn this so much as sharpen it. Locally, blood flow in a working region rises a great deal; across the whole brain, the task-evoked increase is usually under 5% of the ongoing baseline (Raichle and Mintun, Annual Review of Neuroscience, 2006). Five per cent of 20 W is under a watt, and that is a ceiling rather than a reading. What is not here and will not be: the claim that a grandmaster burns 6,000 kilocalories a day. It comes from an ESPN piece of 2019 quoting Robert Sapolsky, and traces back to a tripled breathing rate multiplied by 2,000 kcal. No calorimeter appears anywhere in that chain. What does rise during a hard game is sympathetic arousal — heart rate, blood pressure, breathing, muscle tone. That raises the whole body's expenditure, not the brain's, and how much it raises it over a chess match has never been published.

Sources: L. Sokoloff et al., 1955; M. E. Raichle, M. A. Mintun, 2006; Aishwarya Kumar, 2019; Andrew Gelman, 2025

These four lines are the same on all nine pages, because the human side of the table was the same in all nine matches. Twenty watts is the brain and not the person: what sat at the board was a whole body, and setting an organ against a machine's wall socket is a way of choosing the answer in advance.

How it went

The two sides

Kellin Pelrine is an American amateur, one rank below the top amateur tier. KataGo is an open-source program of the AlphaZero family which, given enough search, plays above any human. Between them stood a research group at FAR AI: Tony Wang, Adam Gleave and co-authors, Pelrine among them. What the group trained was not a stronger player but an adversarial policy — a program whose only job was to find where KataGo is wrong. Doing so took about fourteen per cent of the computation that had gone into training KataGo itself.

T. Wang, A. Gleave, K. Pelrine et al., Adversarial Policies Beat Superhuman Go AIs, ICML 2023 (arXiv:2211.00241); Pelrine's rank as reported by the Financial Times, February 2023.
What was at stake

No title was at stake — there are no titles for matches like this any more. What was at stake was the word superhuman. Since 2016 the default assumption had been that the gap only widens: programs get stronger, people stay where they are, and the last even-terms human win recedes for good. The question FAR AI asked was a different one. Does stronger than any human on average mean stronger in every position? They set out to answer it not by raising a human's strength but by looking for a hole.

The framing is set out in the FAR AI paper, ICML 2023 (arXiv:2211.00241).
What decided it

The first hole they found was a dull one: the adversary talked KataGo into passing early and won on the scoring rules rather than on the board. That was patched. Then a second one turned up. The adversarial program did not learn to play better. It found that KataGo systematically misjudges large groups connected in a closed ring: the program treats such a group as near indestructible when it is not. The recipe is short. Let the program build the ring, distract it with moves in other corners of the board, then surround the ring from the outside and take it. Against KataGo running enough search to be superhuman, the adversarial policy won over 77 per cent of games. Pelrine learned the recipe and carried it out himself, with no computer help during play: fourteen wins in fifteen games. The same plan worked against Leela Zero.

T. Wang, A. Gleave, K. Pelrine et al., ICML 2023 (arXiv:2211.00241); the 14-of-15 score was reported by the Financial Times, February 2023.
How it looks now

A record that had stood for seven years fell, and not because a human got stronger. It fell because someone located a place where the model is blind and then described that place to a person in words. Later work by the same group found the hole hard to close: defences trained against known attacks lost to new ones. The practical conclusion is modest and unwelcome to both sides of the argument. Superhuman is a claim about the average opponent, not about every position on the board — and finding where the boundary runs is still something a person cannot do without a program. Pelrine won fourteen games in fifteen, and he is not the one who found the opening. Of the five matches on this page, this is the only one the human side won, and it is also the only one where the win was prepared by a machine.

T. Wang et al., ICML 2023; on the durability of the weakness, FAR AI's 2024 work on defending Go models.

Sources

ANSI/ASHRAE Standard 55: Thermal Environmental Conditions for Human Occupancy (the table of metabolic rates: seated quiet 1.0 met, seated office work 1.2 met; the standard adult's DuBois surface area 1.8 m²)
ASHRAE · ANSI/ASHRAE 55-2023 · 2023 · book · primary
original · archived copy · checked 2026-08-12
Price Index for Private Fixed Investment: Information Processing Equipment: Computers and Peripheral Equipment (series B935RG3A086NBEA; the fall stops around 2011 and the index rises from 2020)
U.S. Bureau of Economic Analysis · FRED, Federal Reserve Bank of St. Louis · 2026 · web page · primary
original · archived copy · checked 2026-08-12
Producer Price Index by Industry: Electronic Computer Manufacturing (series PCU334111334111; the same shape: a plateau from the early 2010s, rising from 2020)
U.S. Bureau of Labor Statistics · FRED, Federal Reserve Bank of St. Louis · 2026 · web page · primary
original · archived copy · checked 2026-08-12
The grandmaster diet: How to lose weight while barely moving (where the “6,000 calories a day” figure comes from, credited to Robert Sapolsky; no calorimetry underneath it)
Aishwarya Kumar · ESPN, 13 September 2019 · 2019 · press · secondary
original · archived copy · checked 2026-08-12
Consumer Price Index for All Urban Consumers: All Items in U.S. City Average, annual, not seasonally adjusted (series CPIAUCNS; a snapshot of the annual averages, 1913–2025)
U.S. Bureau of Labor Statistics · FRED, Federal Reserve Bank of St. Louis · 2026 · web page · primary
original · archived copy · checked 2026-08-12
Man beats machine at Go in human victory over AI
Financial Times (reported by Richard Waters) · Financial Times, 19 February 2023 · 2023 · press · secondary
original · archived copy · checked 2026-08-12
About that bogus claim that “chess grandmasters” burn 6000 calories per day (where the number came from: respiration measured on 11 ordinary players, the highest reading passed off as the average, multiplied against 2,000 kcal)
Andrew Gelman · Statistical Modeling, Causal Inference, and Social Science, 30 June 2025 · 2025 · web page · secondary
original · archived copy · checked 2026-08-12
Adversarial Policies in Go — the game records, the human-played ones among them
FAR AI · goattack.far.ai · 2023 · web page · primary
original · archived copy · checked 2026-08-12
ISO 8996: Ergonomics of the thermal environment — Determination of metabolic rate (tables of metabolic heat production by activity; 1 met = 58.2 W/m²)
International Organization for Standardization · ISO 8996:2021 · 2021 · book · primary
original · archived copy · checked 2026-08-12
A new predictive equation for resting energy expenditure in healthy individuals (the Mifflin–St Jeor equation)
M. D. Mifflin, S. T. St Jeor, L. A. Hill, B. J. Scott, S. A. Daugherty, Y. O. Koh · American Journal of Clinical Nutrition 51(2), pp. 241–247 · 1990 · paper · primary
original · archived copy · checked 2026-08-12
Two Centuries of Productivity Growth in Computing (table 10: the official BEA hedonic index — the real price of computing falling 18.7% a year over 1969–2004; the performance-based measure — 50.7% a year over 1969–2005)
W. D. Nordhaus · The Journal of Economic History 67(1), pp. 128–159 · 2007 · paper · primary
original · archived copy · checked 2026-08-12
NVIDIA Tesla V100 GPU Accelerator — Datasheet: 300 W TDP in the SXM2 form factor, 250 W in the PCIe one
NVIDIA · NVIDIA product datasheet · 2018 · vendor document · primary
original · archived copy · checked 2026-08-12
Appraising the brain's energy budget (the brain is about 2% of body mass and takes about 20% of the oxygen and the calories the body uses; the rate is remarkably constant across widely varying mental and motor activity)
M. E. Raichle, D. A. Gusnard · Proceedings of the National Academy of Sciences 99(16), pp. 10237–10239 · 2002 · paper · primary
original · archived copy · checked 2026-08-12
Brain Work and Brain Imaging (the increase in the brain's energy use that a task brings about is small next to the baseline — usually under 5%)
M. E. Raichle, M. A. Mintun · Annual Review of Neuroscience 29, pp. 449–476 · 2006 · paper · primary
original · archived copy · checked 2026-08-12
The Effect of Mental Arithmetic on Cerebral Circulation and Metabolism (13 subjects; doing arithmetic in the head shifts the EEG and raises pulse and blood pressure, while cerebral blood flow and the brain's oxygen consumption do not move; control CMRO₂ 3.3–4.0 ml O₂ per 100 g per minute)
L. Sokoloff, R. Mangold, R. L. Wechsler, C. Kennedy, S. S. Kety · Journal of Clinical Investigation 34(7), pp. 1101–1108 · 1955 · paper · primary
original · archived copy · checked 2026-08-12
Adversarial Policies Beat Superhuman Go AIs (appendix G.2 — the attack carried over to a human: 14 games of 15 against JBXKata005, wins over KataGo and Leela Zero at 100,000 visits, handicaps of 3, 5 and 9 stones, “without any algorithmic assistance”; appendix D — 2,223.2 V100-days for the adversarial policy, 1.01 million self-play games, no fewer than 15,881 V100-days to train KataGo, 6,730 of them the “third major run”; section 5.4 — KataGo fine-tuned against the attack from mid-December 2022)
T. T. Wang, A. Gleave, T. Tseng, K. Pelrine, N. Belrose, J. Miller, M. D. Dennis, Y. Duan, V. Pogrebniak, S. Levine, S. Russell · ICML 2023; arXiv:2211.00241 (v1 — 1 November 2022, v3 — 18 February 2023, v4 — 13 July 2023) · 2023 · paper · primary
original · archived copy · checked 2026-08-12