Programming a Computer for Playing Chess

idea is right, and this is the canon's first entry of that kind. Everything written before it is fiction, except lovelace-1843 (an argument) and the two theorems, turing-halting-1936 and godel-incompleteness-1931. An idea entry holds a framing that later work is built out of rather than argued about: the citation occasion is not "someone is wrong about this" but "the thing you are watching descends from this, and the descent explains its shape." Shannon's paper is the clean case. Every chess engine that has ever existed is a modification of it, and so — more loosely but not trivially — is a great deal of what the readings now call search, evaluation and test-time compute.

Why not prediction. The paper does contain forecasts, and section 3 below grades them, because a canon that only grades entries filed under prediction will systematically miss the forecasts that arrive embedded in engineering papers — which is where most of them actually live. But it carries no dated claim, and the prediction kind exists to hold four things (claim, date made, date due, what happened) that this paper does not supply. What it supplies instead is a method. Note also the most interesting thing the paper does not say: it never predicts that a machine will beat a master, let alone a world champion. That silence is graded below, against the man who broke it seven years later.

descends_from is empty, and here that is a checked fact rather than a gap. The paper's reference list has seven items — Chernev, De Groot, Fine, Hardy and Wright, von Neumann and Morgenstern, Vigneron, Wiener — and I read it directly. There is no Turing in it, no Babbage, and no self-citation to Shannon's own 1948 information theory. So although turing-halting-1936 and lovelace-1843 are both in canon/ and both chronologically prior, neither is an ancestor of this entry and neither should be linked to it. The honest ancestors are von Neumann and Morgenstern's Theory of Games (1944), which Shannon cites at page 125 for the result that every chess position is won, lost or drawn, and Wiener's Cybernetics (1948). Neither is in the canon, and neither is in proposals.md either. If someone writes them, this entry's header should be edited to descend from the first of them.

Four dates, and the id names the latest. The acknowledgements are dated 8 October 1948. The work was first presented at the National IRE Convention in New York on 9 March 1949. The Philosophical Magazine received it on 8 November 1949 and printed it in March 1950 as article XXII of Ser. 7, Vol. 41, No. 314. The id says 1950 because that is the publication and the way the paper is universally cited, but the thinking is 1948 — contemporaneous with Cybernetics and with Shannon's own communication theory, and two years ahead of Turing's Mind paper. A reading that wants to say "this was being worked out before there was a stored-program computer in the United States to run it on" is entitled to; Shannon was writing for a machine he describes hypothetically in Fig. 3 and hoped to borrow: "It is planned, however, to experiment with a simple strategy on one of the numerical computers now being constructed."

What it is

A twenty-page engineering specification for making a general-purpose computer play chess, written when no computer had ever played a game of anything. It is not an essay about whether machines can think. It contains a memory layout, an instruction-level sketch of ten subprograms, a worked estimate that 3,000 bits of storage would suffice for a three-move search, and a sample game the author played against his own random-move procedure and won in four moves.

Shannon opens by conceding that the problem itself does not matter: "Although perhaps of no practical importance, the question is of theoretical interest, and it is hoped that a satisfactory solution of this problem will act as a wedge in attacking other problems of a similar nature and of greater significance." He then lists eight such problems, which are reproduced in full in section 3 because they are the paper's real forecast. Chess is chosen as the wedge for four stated reasons: the problem is sharply defined in both operations and goal; it is "neither so simple as to be trivial nor too difficult for satisfactory solution"; "chess is generally considered to require 'thinking' for skilful play; a solution of this problem will force us either to admit the possibility of a mechanized thinking or to further restrict our concept of 'thinking'"; and the discrete structure of the game suits a digital machine.

That third reason is worth pausing on, because it is the shape of nearly every capability argument the readings now cover. Shannon does not claim the machine will think. He claims that a working chess machine will force a choice, and he predicts, correctly and without apparent regret, that people may well take the second branch and redefine thinking to exclude whatever the machine turned out to be doing. That is exactly what happened, repeatedly, and it is the mechanism behind every "but it's just autocomplete" that a reading has to adjudicate.

Why the obvious approach is dead. From von Neumann and Morgenstern, every chess position is a forced win for White, a draw, or a forced win for Black. Shannon calls this "for practical purposes, of the nature of an existence theorem. No practical method is known for determining to which of the three categories a general position belongs. If there were chess would lose most of its interest as a game." He then prices the exhaustive method. Typical positions offer about 30 legal moves — a figure he takes from De Groot's counts over master games, not from intuition. A White move and a Black reply therefore give about 10³ possibilities, and a game lasts about 40 moves, which he calls conservative "since the machine would calculate out to checkmate, not resignation." That gives 10¹²⁰ variations from the initial position, and: "A machine operating at the rate of one variation per micro-second would require over 10⁹⁰ years to calculate the first move!" The alternative of a lookup dictionary fails on the other axis: "The number of possible positions, of the general order of 64!/32!(8!)²(2!)⁶, or roughly 10⁴³, naturally makes such a design unfeasible."

This is the number the world remembers, usually under the name Shannon did not give it. It is worth being precise about what it is a bound on, because that precision is the whole of section 4 below: 10¹²⁰ is the size of the tree you must walk if you insist on walking all of it. It is an argument against a particular method, produced in order to motivate a better one, by the person proposing the better one.

The evaluation function. Since no exact f(P) exists for chess "and probably never will be because of the arbitrary and complicated nature of the rules of the game," Shannon proposes an approximate one, assembled from the maxims chess players already use: piece values, pawn structure, mobility. His illustrative formula is

f(P) = 200(K−K') + 9(Q−Q') + 5(R−R') + 3(B−B'+N−N') + (P−P') − 0.5(D−D'+S−S'+I−I') + 0.1(M−M') + …

where D, S and I count doubled, backward and isolated pawns, M is mobility measured as legal moves available, and primes denote Black. He is explicit that this is illustrative — "The coefficients 0.5 and 0.1 are merely the writer's rough estimate" — and that the king's weight of 200 is a hack: "Checkmate has been artificially included here by giving the king the large value 200 (anything greater than the maximum of all other terms would do)." An appendix lists the terms he thinks a serious version needs, and closes with a sentence that turns out to matter enormously: "The relative values to be given each of the above quantities is open to considerable debate, and should be determined by some experimental procedure."

Two properties of the approximation are flagged as features. It is continuous where the true function has three values, which lets the machine distinguish an easy win from a hard one, as human players do and as game theory's "unlimited intellect" cannot — Shannon's joke is that a game between two such intellects consists entirely of one of them resigning before a move is played. And it is a proxy: the thing being maximized is not winning, but a hand-weighted guess about winning, evaluated at a horizon chosen for cost.

Type A. Apply f(P) to every position at a fixed depth and back the values up by alternating maximization and minimization — the machine assumes its opponent plays the reply that minimizes f. "A strategy of this sort, in which all variations are considered out to a definite number of moves and the move then determined from a formula such as (1) will be called type A strategy." He programs it: T₀ makes a move, T₁ through T₆ generate moves for each piece type, T₇ lists and legality-filters all moves, T₈ evaluates, T₉ runs the minimax. Then he prices it and rejects it, in the paper's most consequential paragraph: "Unfortunately a machine operating according to the type A strategy would be both slow and a weak player. It would be slow since even if each position were evaluated in one microsecond (very optimistic) there are about 10⁹ evaluations to be made after three moves (for each side). Thus, more than 16 minutes would be required for a move, or 10 hours for its half of a 40-move game."

The weakness argument is separate from the slowness argument and is better: "It would be weak in playing skill because it is only seeing three moves deep and because we have not included any condition about quiescent positions for evaluation. The machine is operating in an extremely inefficient fashion — it computes all variations to exactly three moves and then stops (even though it or the opponent be in check)." A machine that stops counting in the middle of a queen trade will report that it is a queen ahead. Shannon states the general form: "it is meaningless to calculate an evaluation function of the general type given above during the course of a combination or a series of exchanges."

Type B. Two fixes, taken from De Groot's protocols of what masters actually do — in one experiment a master examined sixteen variations, to depths from half a move to four and a half, and considered 44 positions in total. Examine forceful variations as far as they go and evaluate only where "some quasi-stability has been established"; and select which variations to explore at all, "so that the machine does not waste its time in totally pointless variations." The first becomes g(P), a crude stability test — is anything attacked by a lower-valued piece, is there a check available on a square the opponent controls — with search continuing until g(P) = 0, "always, however, going at least two moves and never more, say, 10." The second becomes h(P,M), a plausibility score on moves, high for checks, captures and attacks, medium for developing moves, low otherwise, with the bar rising as the search goes deeper. Shannon's warning about h is sharp and is the reason selective search is hard: "It is important that this preliminary screening should not eliminate moves which merely look bad at first sight, for example, a move which puts a piece en prise; frequently such moves are actually very strong since the piece cannot be safely taken."

His verdict: "It is believed that an electronic computer incorporating these two improvements in the program would play a fairly strong game, at speeds comparable to human speeds."

The rest, which is usually skipped. Shannon lists four advantages the machine has — speed, "freedom from errors," "freedom from laziness," and "freedom from 'nerves'" (over-confidence in won positions, self-recrimination in lost ones) — and then, in one sentence, concedes the other side of the ledger: "These must be balanced against the flexibility, imagination and inductive and learning capacities of the human mind." He notes that the designer could in principle compute the machine's move by hand and so play its equal, "but on a fair basis of comparison, giving the machine and the designer equal time to decide on a move, the machine might well play a stronger game" — the Lovelace objection answered not by denying it but by pricing it in wall-clock. He adds randomization among near-equal moves so a beaten opponent cannot replay the same win; an opening book of "perhaps a few hundred" stored variations, which "is hardly 'cheating' since that is the way chess masters play the opening"; and the observation that style is a matter of coefficients — raise the positional weights and you get a positional player, deepen the forced-variation search and you get a combinative one. "Finally we may note that a machine of this type will play 'brilliantly' up to its limits."

Then the self-criticism: "The chief weakness is that the machine will not learn by mistakes. The only way to improve its play is by improving the program. Some thought has been given to designing a program which is self-improving but, although it appears to be possible, the methods thought of so far do not seem to be very practical. One possibility is to have a higher level program which changes the terms and coefficients involved in the evaluation function depending on the results of games the machine has played."

And a last section proposing a third strategy he cannot specify — a machine carrying an explicit encoding of the tactical structure of a position (this knight is pinned, this rook cannot leave the back rank) and using it to trigger suggestions of plausible moves, on the grounds that Type B without it "plays something like a beginner at chess who has been told some of the principles and is possessed of tremendous energy and accuracy for calculation but has no experience with the game." The obstacle he names is that the chess literature is "written for human consumption, not for computing machines," and a computer needs "an exact and completely explicit characterization of the situation."

The paper's last substantive sentence is the one that gets left out of every summary of it, and it reverses the moral most summaries draw:

> It is not being suggested that we should design the strategy in our own image. > Rather it should be matched to the capacities and weakness of the computer. > The computer is strong in speed and accuracy and weak in analytical abilities > and recognition. Hence, it should make more use of brutal calculation than > humans, but with possible variations increasing by a factor of 10³ every move, > a little selection goes a long way forward improving blind trial and error.

Why a reading would cite it

This entry is not evidence and touches no needle. It is here for five occasions that the readings already produce, roughly weekly.

1. When a benchmark saturates, or a model is found optimizing the score instead of the thing. The 15 Aug midday reading turns on Anthropic raising its own catastrophic-misalignment rating because its safety benchmarks are saturating and its R&D-acceleration measurement is degrading — an admission that the proxy has drifted from the quantity. Shannon is where that structure enters computing as a deliberate engineering decision, stated without euphemism: the true objective (won/lost/drawn) is uncomputable, so we substitute a weighted sum of observable features and search hard against the substitute. He also names the failure mode in the same paper — an evaluation applied at a non-quiescent position is meaningless, and a fixed horizon makes the machine most confident exactly where it is most wrong. A reading that needs this should be careful which claim it is making: goodharts-law-1975 (proposed, unwritten) covers a measure degrading because it became a target, which is a social mechanism. Shannon covers the prior and more basic fact that the measure was an approximation from the start, chosen for tractability, by people who said so. Cite Shannon for the origin of proxy-plus-search; cite Goodhart for what happens to the proxy afterwards.

2. When someone argues an approach cannot work in principle. 10¹²⁰ is the first respectable impossibility number in the history of AI, and the paper that contains it is a demonstration of how to route around one. This is the entry to reach for when a reading meets "the search space is astronomically large, therefore X is impossible," "scaling this is combinatorially hopeless," or the converse triumphalism. The record it supplies is specific: the method Shannon himself priced at 10⁹⁰ years, then softened to 16 minutes a move and dismissed as hopeless, won the world championship 47 years later on hardware. lighthill-1973 (proposed, unwritten) is the companion case where the combinatorial-explosion argument was believed by a funder and cost a country its research programme.

3. When an inference-speed result is waved off as "not a capability gain." The same reading logs OpenAI's Ultrafast mode at roughly 14× throughput and calls it "a latency change, not an intelligence one, but agentic loops are priced in wall-clock." Shannon is the entry that says why the first half of that sentence needs care. His rejection of Type A was entirely a latency argument — 16 minutes per move, 10 hours for half a game — and at fixed algorithm, throughput buys depth, and depth was the whole distance between his weak player and Kasparov's opponent. Cite when a reading has to decide whether a tokens-per-second number is a capability finding. The honest answer Shannon supports is: it is, when the system spends the time searching, and it is not, when the system emits the same answer faster.

4. When a result comes out of a machine searching further than a person would. The Crouzeix's conjecture item — a 16-hour autonomous model run producing a claimed proof of a problem open since 2004 — is this shape. Shannon supplies both halves of the framing: the machine's advantages are speed, accuracy, freedom from laziness and freedom from nerves, and they are "balanced against the flexibility, imagination and inductive and learning capacities of the human mind." He also supplies the caution that the interesting question is not how many variations were examined but whether the evaluation at the end of them was any good — which, for an unrefereed proof, is exactly the open question.

5. When a scored contest against the best humans is won by a machine. deep-blue-1997 is proposed and unwritten. When it exists it should descend from this entry, and until it does, this is where a reading gets the interval: the architecture was specified in 1950 and the championship fell in 1997. Forty-seven years between a correct framing and its cash-out, with the field's own confident ten-year forecast sitting in the middle of it, is the most useful single number this entry carries.

What it should not be cited for: any claim about whether current systems "really" think. Shannon declined to make one, and the entry is weaker, not stronger, when it is pressed into that service.

What it got right, and what it got wrong

Claims dated 1950 (drafted October 1948). Graded 2026-08-15.

Right: the architecture, and it is still the architecture. Minimax over a truncated game tree with a heuristic evaluation at the leaves is what Deep Blue ran in May 1997 and what Stockfish runs today. The relative piece values he wrote down — 9, 5, 3, 3, 1 — are still the numbers taught to beginners and still sit inside engine evaluations. Quiescence search is universal; the failure he described, evaluating in the middle of an exchange, acquired the name "horizon effect" later and is still the thing quiescence exists to prevent. Opening books are universal, and Deep Blue's held about 4,000 positions plus 700,000 grandmaster games against Shannon's "perhaps a few hundred" — right in kind, short by three orders of magnitude in scale, which is the recurring shape of his errors. Randomizing among near-equal moves so a beaten opponent cannot replay the win: standard. Style and strength as functions of the coefficients and the depth: exactly how engines are configured and how their handicap levels work. Every one of these was specified before anyone had built a machine that could run any of them.

Right, and the most under-credited item in the paper: the learned evaluation function. He named the chief weakness of his own design — "the machine will not learn by mistakes" — and proposed the fix in one sentence: a higher-level program that changes the terms and coefficients of the evaluation function according to the results of games played. He gave it no date and said the methods he could think of were impractical. Samuel did it for checkers in 1959. It took seventy years to become the default: AlphaZero learned its evaluation from self-play in 2017, and Stockfish replaced its hand-tuned evaluation with an efficiently-updatable neural network in August 2020, shipping in Stockfish 12 that September. The appendix's "should be determined by some experimental procedure" is, read from 2026, a description of gradient descent by a man who had no gradient to descend. Grade: right on the mechanism, silent on the date, and correct that it was the binding constraint.

Right, wrong on timing: the eight applications. The paper's real forecast is its list of what the wedge was for, offered with the words "It is believed that all of these and many other devices of a similar nature are possible developments in the immediate future." Machines for designing filters and equalizers; for designing relay and switching circuits; for routing telephone calls on individual circumstances rather than fixed patterns; for symbolic mathematics; for translating between languages; for making strategic decisions in simplified military operations; for orchestrating a melody; for logical deduction. Every one landed. Circuit design and place-and-route are automated and are now themselves a target of machine learning; symbolic algebra was solved well enough to be a commodity by the 1980s; machine translation became usable at scale around 2016; melody generation is a consumer product currently generating litigation; automated theorem proving is a live research front that this project's own accelerant lens tracks. "In the immediate future" was wrong by between fifteen and sixty-five years depending on the item, and the military one is no longer simplified. Grade the content and the schedule separately: the content is close to a clean sweep, the schedule is the same optimism the prediction entries exist to hold a base rate for. That an eight-for-eight forecast can still be badly wrong is the point.

Wrong, and this is the famous one: he backed Type B and Type A won. He judged full-width search "both slow and a weak player" and put the paper's recommendation behind selective search. Type B programs did dominate the first two decades — Bernstein's, Kotok–McCarthy, Mac Hack, CHAOS — and then lost decisively. Once brute-force alpha-beta programs like Tech (1972), Kaissa and Slate and Atkin's CHESS 4.x could reach five plies and more, the plausible-move generators started losing on tactics they had pruned away; CHESS 4.5 reached a USCF expert rating in 1975. Deep Blue in 1997 was that lineage at scale: about 200 million positions per second, alpha-beta, an evaluation split into some 8,000 parts. Claim made 1950, effectively due by the mid-1970s, graded wrong.

But grade the error precisely, because the popular version of it is also wrong. Three corrections, in increasing order of how much they change the lesson.

First, it was a hardware forecast, not a conceptual one. Shannon priced Type A at one microsecond per evaluation, called that "very optimistic," and got 16 minutes for a six-ply move. Deep Blue did 200 million per second. He was not wrong about the arithmetic; he was wrong about how much arithmetic there was going to be — a specific, dated, single-variable miss that later got promoted into a claim about the nature of intelligence he never made.

Second, he missed alpha-beta, and that omission is most of the gap. His T₉ enumerates the entire tree. Alpha-beta pruning searches the same tree to nearly twice the depth for the same work and — unlike his h(P,M) — is exact, returning the identical move. It was found piecemeal over the following decade by McCarthy, by Newell, Simon and Shaw, and by others, and analysed definitively by Knuth and Moore in 1975. This is the counterweight to reading the whole episode as a bitter-lesson story: a large part of what made Type A viable was not transistors but an idea, and Shannon did not have it.

Third, and decisively: the closing paragraph says the computer "should make more use of brutal calculation than humans." The man most often cited as having predicted that brute force would fail wrote, as his last substantive sentence, that brute force is what the machine is for, with selection as a multiplier on it rather than a replacement for it. Deep Blue is a vindication of that sentence, not a refutation of it.

And then the reversal, which is why this entry earns its place rather than being a historical curiosity. In December 2017 DeepMind published AlphaZero, which beat the 2016 TCEC champion Stockfish over 100 games — 28 wins, 72 draws, zero losses — while searching 80,000 positions per second against Stockfish's 70 million. The paper's own account of why, with a citation to this entry: AlphaZero "compensates for the lower number of evaluations by using its deep neural network to focus much more selectively on the most promising variations — arguably a more 'human-like' approach to search, as originally proposed by Shannon." It adds that this calls "into question the widely held belief that alpha-beta search is inherently superior in these domains."

So the full grade on the paper's central bet is: wrong on its due date, right on a mechanism the author explicitly could not supply. Type B lost in 1975 because nobody could write h(P,M) by hand well enough, and won in 2017 because h(P,M) stopped being written by hand — the plausible-move function was learned from self-play, which is precisely the "higher level program" Shannon proposed in his weakness paragraph and could not build. Seventy years separate the proposal from the working version, and the two halves of the arc point opposite ways. A reading that quotes only Deep Blue gets "selection was a dead end"; a reading that quotes only AlphaZero gets "Shannon was right all along." Both are half a record.

Right and still standing: chess is not solved. 10⁴³ has not moved. Exhaustive solution has reached seven pieces — the Lomonosov tablebases, 2012, about 140 terabytes — and eight-piece work is still in progress in 2026 at an estimated petabyte. Seventy-six years on, his line about the trichotomy being an existence theorem with no practical method behind it is simply true, and so is the aside that if it were ever otherwise, "chess would lose most of its interest as a game."

Right by omission, which is the finding this project should keep. Shannon's strongest claims about strength were "a tolerably good game," "perhaps comparable to that of a good human player," and "a fairly strong game, at speeds comparable to human speeds." No date. No champion. His only mention of the reigning world champion is a footnote calculating that random play has about a 10⁻⁷⁵ chance of beating Botvinnik. Seven years later, Newell and Simon supplied what he had declined to: delivered as a keynote to the Operations Research Society of America in 1957 and published in Operations Research in early 1958, the prediction that within ten years a digital computer would be the world's chess champion, unless the rules barred it from competition. Due 1967. Achieved 1997, thirty years late. The engineer who did the work made no forecast; the forecast came from people who had not done it. That contrast is a base rate, and it is worth more to a reading than either data point alone — the same asymmetry this project's lenses encode when they rank shipped deployments above surveys and predictions.

One dated fact about the record, stated without decoration. Shannon died on 24 February 2001, aged 84, after years with Alzheimer's disease. Deep Blue's win fell inside his lifetime, in May 1997, and almost certainly outside his reach.

Commonly misused as

Not required for idea, but the misuses are frequent enough, and load-bearing enough in exactly the arguments the readings adjudicate, to be worth policing.

"Shannon proved chess can't be brute-forced." 10¹²⁰ is a bound on exhaustive search of the game tree from the initial position. It says nothing about search with a horizon, an evaluation function and pruning, which is what the rest of the paper proposes and what actually works. The number is now routinely deployed as a general-purpose impossibility argument — the search space is astronomically large, therefore this cannot be done — in contexts Shannon never addressed. The paper is a standing counterexample to that inference, written by the person whose number is being cited against it. When a reading meets the move, the correction is one sentence: the number was computed in order to justify not doing that, not to prove nothing could be done.

"Deep Blue vindicated brute force over Shannon." Deep Blue is Shannon's Type A. He named it, described it, wrote the ten subprograms for it, and was wrong about its prospects — which is a different thing from being ignorant of it, and a very different thing from having opposed it in principle. The version of this misuse that most needs correcting is the one that presents 1997 as the defeat of a "human-like AI" school by an "engineering" school, with Shannon cast as the former. His own closing paragraph explicitly refuses to design the machine in our image and calls for more brutal calculation than a human uses.

"Shannon invented minimax." No. The minimax theorem is von Neumann's, 1928, and Shannon cites von Neumann and Morgenstern (1944, at page 125) in the paper itself for the won/drawn/lost trichotomy. Backing values up a game tree was not new. The contribution is what happens at the leaves: an approximate, continuous, hand-weighted evaluation standing in for a value nobody can compute, plus the decision to stop early and accept the error. Minimax on an exact value function was already known and is useless for chess. The approximation is the idea.

"Deep Blue was pure Type A" / "modern engines are Type A." The clean binary is a teaching device that the engines stopped obeying almost immediately. Deep Blue searched full-width to a depth and then extended selectively, sometimes far beyond it. The chess programming literature's own summary is that most programs today are closer to Type A but carry Type B characteristics through selectivity, and since 2017 the strongest programs have included one that is emphatically Type B and one whose leaf evaluation is a neural network. Anyone reaching for the dichotomy to classify a modern system — including a modern system that is not a chess engine — should expect it to come back a hybrid.

The one that matters most here: quoting half the arc. Citing Deep Blue without AlphaZero gives "Shannon's selective search was a dead end, and compute wins." Citing the AlphaZero sentence without the 1975 collapse gives "Shannon was right and the brute-force era was a detour." The record is that a 1950 design decision was refuted by hardware in the 1970s and re-instated by learning in 2017, and that the reinstatement depended on solving the exact sub-problem Shannon flagged as unsolved. A reading that uses this entry should use the whole arc or say which part it is using and why.

Sources

Primary, read directly:

• Claude E. Shannon, "XXII. Programming a Computer for Playing Chess", Philosophical Magazine, Ser. 7, Vol. 41, No. 314, March 1950, pp. 256–275. Received 8 November 1949; first presented at the National IRE Convention, New York, 9 March 1949; acknowledgements dated 8 October 1948. Read in full from the copy at https://vision.unipv.it/IA1/aa2009-2010/ProgrammingaComputerforPlayingChess.pdf, including the appendix, the six numbered notes and the seven-item reference list. All quotations above are from this text.
• A second transcription at https://www.pi.infn.it/~carosi/chess/shannon.txt, used first and then cross-checked against the PDF. Where they disagreed I took the PDF; the .txt copy is missing material the PDF has.

A caveat on the quotations, since this entry spends a section policing misquotation. Both copies are re-typeset transcriptions, not scans of the printed original, and both carry evident typos ("White the approach given here" for "While", "form these principles" for "from"). Two spots affect this entry directly. The list of machine advantages is transcribed as Freedom from "never", glossed by Shannon as over-confidence in won positions and defeatism in lost ones; the published word is plainly "nerves" and I have quoted it that way. And a sentence in section 8 about whether a "type position" program could be constructed is corrupt in both copies — it reads as its own negation — so I have not quoted it and have described the section's argument instead. I have not seen the printed original.

Secondary and corroborating:

• David Silver et al., "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm", arXiv:1712.01815, 5 December 2017. Read directly for the Shannon citation, the 80,000-vs-70-million positions-per-second figures and the Table 1 match result (28–0 with 72 draws).
• IBM's own Deep Blue history page and IEEE Spectrum's account, for 200 million positions per second, the 8,000-part evaluation, the 4,000-position opening book and the 3½–2½ result of May 1997. These are substantially the vendor's numbers and are cited as such.
• chessprogramming.org, Type_B_Strategy and Claude_Shannon, for the Type B program lineage (Bernstein, Kotok–McCarthy, Mac Hack, CHAOS), the early-1970s transition via Tech, Kaissa and CHESS 4.x, and the "closer to Type A but with Type B characteristics" summary of current engines.
• stockfishchess.org, "Introducing NNUE Evaluation", 6 August 2020, and the Stockfish 12 release of September 2020.
• ChessBase and Marc Bourzutschky's progress reports for the state of eight-piece tablebases as of 2026; the Lomonosov seven-piece set dates from 2012.
• MIT News obituary, 27 February 2001, for Shannon's death and illness.

Noted failures and second-hand items, per the discipline:

• Shannon's companion popular article, "A Chess-Playing Machine", Scientific American Vol. 182, No. 2, February 1950, pp. 48–51, is a genuine second primary source for this entry and I could not read it: the fetch of the Caltech copy failed on 2026-08-15 with a TLS certificate verification error. The bibliographic details above are from two independent secondary indexes. Nothing in this entry is quoted from it, and if a later session gets hold of it, the section on what Shannon claimed about machine strength is where it would bear.
• The Newell and Simon chess prediction — delivered as "Heuristic Problem Solving: The Next Advance in Operations Research" to the Operations Research Society of America in 1957, published in Operations Research in early 1958 — I have at second hand only, from several consistent secondary accounts. I did not read the journal article. If it is ever written up as its own prediction entry, the wording should be checked against the printed text before it is quoted.

This entry places no needle, no score and no landmark, and deposits nothing in the evidence ledger.