Deep Blue defeats Garry Kasparov

Filed as moment, and for once the moment really is a day. The house definition, argued out by eliza-1966 when it talked its way out of the kind and used since by clippy-1996, siri-2011 and expert-systems-collapse-1987, is a date on which something visibly happened in public. Every one of those entries had to negotiate with it. This one does not. On Sunday 11 May 1997, starting at 3 p.m. EDT on the 35th floor of the Equitable Center in midtown Manhattan, with a paying audience of up to 500 watching on closed-circuit television in the auditorium below, Garry Kasparov sat down as reigning world chess champion, made a mistake on move seven, and resigned after nineteen moves and a little over an hour. There is a clock time, a room, a scoresheet and a witness list. It is the cleanest instance of the kind the canon holds.

That cleanliness is the trap, and it is why this entry is long.

The proposal line is wrong twice, and correcting it is the entry's first job. canon/proposals.md describes this as "the first date the public agreed a line had been crossed." It was not the first time a machine beat the reigning world champion: that was 10 February 1996, game 1 of the earlier match in Philadelphia, which Deep Blue won under classical time controls and which Kasparov then answered by winning three of the next five and taking the match 4–2. It was not even the first time Kasparov lost to a machine: Chess Genius 3, running on a personal computer and operated by Ossi Weiner, beat him at 25 minutes a side at the Intel Speed Chess Grand Prix in London in August 1994 and knocked him out of the event. What was new in May 1997 is narrower and should be stated precisely: a machine won a multi-game match against the reigning world champion at classical time controls. The proposal's second sentence — "cite when: a scored contest against the best humans is won by a machine" — is exactly right, and section 2 is built on it.

Note also what "the public agreed a line had been crossed" actually is. It is a fact about the public, not about the machine, and this entry's central service is keeping those two apart.

Why not interpretation. The strongest rival, and it nearly wins. What descends from May 1997 into 2026 is arguably not the event but a genre: the staged, scored, vendor-hosted contest in which a system is shown beating the best humans at a bounded task, and the result is reported as a statement about capability in general. An interpretation entry on that genre would be a legitimate and useful thing. But it would be a different entry with a different id, and this id names a match. Matches are events; events are moments.

Why not prediction. Tempting, because dated forecasts are stacked around this date more densely than around anything else in the canon — Simon and Newell in 1957, Fredkin's prize in 1980, Kasparov about himself in 1989, Piet Hut about Go eleven weeks after the match. Section 3 grades all of them, because a canon that only grades entries filed under prediction will systematically miss the forecasts that arrive attached to events. But the event is not a claim, and the claims about it deserve their own entries if they deserve entries at all.

Why not limit. Nothing here was proved, and it is worth saying out loud because "chess fell, therefore X will fall" is deployed constantly as though a result had been established. One machine beat one player six times in a nine-day span in one game with perfect information, complete rules and no hidden state. That is a demonstration. Filing it beside godel-incompleteness-1931 and turing-halting-1936 would lend it exactly the authority section 4 exists to take away.

descends_from holds one id, and the edge is documented rather than inferred. shannon-chess-1950 already names this entry in its own prose — "deep-blue-1997 is proposed and unwritten. When it exists it should descend from this entry" — and describes Deep Blue as "that lineage at scale: about 200 million positions per second, alpha-beta, an evaluation split into some 8,000 parts." The mechanism is Shannon's Type A: a full-width search of the game tree to a horizon, with an approximate hand-weighted evaluation standing in at the leaves for a value nobody can compute. That is what the chess chip does in silicon. The interval is the most useful number either entry carries: the architecture was specified in March 1950 and the world championship fell in May 1997, forty-seven years later.

Two ancestors I would want and cannot have. logic-theorist-1956 is proposed and unwritten, and it matters here in a way that is almost too neat: its authors, Newell and Simon, made the forecast this entry grades hardest, and their program is the one Simon's second 1957 prediction was really about. dartmouth-1956 is not proposed under any kind and probably should be. alphago-move-37-2016 is proposed and unwritten, and it is a descendant of this entry rather than an ancestor — the point of section 3's Piet Hut item.

And one non-ancestor worth naming, because the claim is common. The New York Times in 2018 called Deep Blue "a then-state-of-the-art expert system, relying upon rules and variables defined and fine-tuned by chess masters and computer scientists," and versions of that line are everywhere. The resemblance is real and shallow: like the systems in mycin-1976, Deep Blue's domain knowledge was hand-encoded by human experts, and a grandmaster on the payroll spent months adjusting it. The descent is not real. There is no rule base, no inference engine, no EMYCIN in the lineage; the knowledge is a few thousand numeric weights on 8,000 evaluation terms, consulted 100-million-odd times a second inside an alpha–beta search that mycin-1976 has nothing to do with. Deep Blue does not descend from the expert-system tradition, and expert-systems-collapse-1987 is not an ancestor of this entry. What the two do share is a failure mode, and section 4 says so.

What it is

The lineage

Feng-hsiung Hsu began building chess hardware as a doctoral student at Carnegie Mellon, first as ChipTest (which won the North American Computer Chess Championship in 1987) and then, with Murray Campbell and others, as Deep Thought in 1988. Deep Thought played Kasparov in a two-game exhibition at the New York Academy of Art in October 1989 and lost both. Hsu and Campbell joined IBM Research after Hsu's doctorate in 1989; the machine was renamed in an internal contest won by Peter Fitzhugh Brown; Jerry Brody joined in 1990; C.J. Tan managed the project.

A useful and almost never told fact sits in the middle of that arc. In May 1995, at the 8th World Computer Chess Championship in Hong Kong, the Deep Blue prototype — the machine two years from beating the world champion — lost with White to Fritz, running on a personal computer, in 39 moves. Anyone about to argue that specialised hardware at 1997 scale was obviously going to win should sit with that game first.

The machine

Two distinct machines carry the name, and the paper by the people who built it distinguishes them as Deep Blue I (1996, a 36-node IBM RS/6000 SP with 216 chess chips) and Deep Blue II (1997). Deep Blue II was:

• a 30-node IBM RS/6000 SP, 28 nodes on 120 MHz P2SC processors and 2 on 135 MHz, each node with 1 GB of RAM and 4 GB of disk;
• 480 custom single-chip chess search engines, 16 per node, each searching 2–2.5 million positions per second, roughly 1.5 million transistors apiece;
• searching in three layers — one SP processor as master near the root, the rest as workers a few plies down, the chess chips doing the last few levels and the quiescence search;
• carrying an evaluation function of more than 8,000 features, raised from about 6,400 in the 1996 chip, implemented in hardware so that evaluation costs a fixed constant;
• fed by an opening book of about 4,000 positions, an "extended book" distilled from 700,000 grandmaster games, and endgame databases covering all five-piece endings and many six-piece ones;
• reaching six to eight moves deep typically, and twenty or more in some lines;
• ranked 259th on the June 1997 TOP500 list at 11.38 GFLOPS on parallel LINPACK — which is to say it was a real supercomputer, and not a large one.

The famous number needs care, and this is the entry's first act of number-policing. "200 million positions per second" is the figure IBM put in front of the public and it is repeated everywhere, including in shannon-chess-1950. The team's own paper reports something more specific about the match itself: the average system speed observed in searches longer than one minute was 126 million positions per second, and the maximum sustained speed observed in the match was 330 million. So the headline figure is neither the average nor the peak. It is a round number between them. That is not a scandal — it is what a marketing figure is — but a project that records magnitudes should know it is quoting one.

Deep Blue did not learn. Nothing in the system updated itself from experience. Between the 1996 and 1997 matches the team, in their own summary, "spent the vast majority of our time" designing, testing and tuning the new evaluation function by hand. Grandmaster Joel Benjamin was on staff to tune it; the opening library was built by grandmasters Miguel Illescas, John Fedorowicz and Nick de Firmian. The machine that beat the world champion was, in the most literal sense, several strong humans and a very fast tree search, and no part of it got better by playing.

The match

Six games, 3–11 May 1997, at classical time controls, for a $1.1 million purse — $700,000 to the winner, $400,000 to the loser — plus, later, the $100,000 Fredkin Prize from Carnegie Mellon.

| Game | Date | White | Result | Opening | |---|---|---|---|---| | 1 | 3 May | Kasparov | 1–0 Kasparov | Réti | | 2 | 4 May | Deep Blue | 1–0 Deep Blue | Ruy Lopez | | 3 | 6 May | Kasparov | ½–½ | English | | 4 | 7 May | Deep Blue | ½–½ | Caro-Kann | | 5 | 10 May | Kasparov | ½–½ | Réti | | 6 | 11 May | Deep Blue | 1–0 Deep Blue | Caro-Kann |

Final: Deep Blue 3½ – Kasparov 2½. One decisive win each until the last game.

Three moments did all the work, and none of them is the one people remember.

Game 1, move 44. Deep Blue, losing, played a move that made no sense. It was later reported — sourced to Murray Campbell, via Nate Silver's The Signal and the Noise — that this was a bug: unable to settle on a move, the program fell through to a fail-safe. Kasparov did not know that. He is described as having read depth into it. Whatever weight one puts on the anecdote (see Sources — I could not verify it at first hand), the structure of it is the thing worth keeping: a human being inferred intent from an artifact of broken code, and adjusted his model of the opponent accordingly.

Game 2. Deep Blue played positional chess of a quality Kasparov did not believe a machine capable of, declining a material gain that engines of the era were expected to grab. He resigned — and had in fact resigned a drawn position; after 45.Ra6, 45...Qe3 46.Qxd6 Re8 forces perpetual check. He learned this afterwards from analysts. From that point he was, by his own later account, playing a thing he did not understand.

Game 6. He chose the Caro-Kann, played 7...h6 — one move out of order, against his own prior practice in the same structure — and walked into 8.Nxe6, a knight sacrifice already known from earlier high-level games and scoring well for White. Nineteen moves. He said afterwards: "I was not in the mood of playing at all."

The match was decided by a human collapse into a line already in the book that the machine had been given.

The conditions, which are the part that matters

• IBM controlled the venue and the conditions. Kasparov's own summary, quoted in the contemporaneous account by Jonathan Schaeffer and Aske Plaat: "IBM's total control of the site and the playing conditions underscored the vulnerability of the human player."
• The rules permitted IBM to modify the machine between games, and IBM says it did so, to patch weaknesses that showed up during the match. This was not cheating; it was the agreed format. It is also not something the human player could do.
• Access to prior games was asymmetric. IBM's team had Kasparov's entire published career to study, plus 700,000 grandmaster games compiled into the machine's extended book. Kasparov was denied access to Deep Blue's recent games. The machine that would face him had, at that point, played almost nothing in public.
• Kasparov asked for the machine's printouts during the match and was refused. IBM said the logs would be published "in the near future"; the common account is that they did not appear on IBM's site for around three years, well after the machine was gone. (Flagged in Sources: I could not read the source for that timeline directly.)
• IBM refused a rematch, retired the project, and reassigned the team. The 480 chess chips were dismantled. One rack went to the Smithsonian's National Museum of American History, the other to the Computer History Museum.

So: the machine that produced the most cited capability result in the history of AI played exactly six public games in its existence, all against one opponent, under conditions its owner set, was never independently evaluated, was never asked to repeat the result, and no longer exists. That sentence is the entry.

The money

IBM's shares rose in the days after the match to their highest level in about ten years. Beyond that direction, the figures usually attached to this — a 3.6% rise, "$18 billion overnight," a Wall Street Journal estimate of $100 million in publicity value — circulate in secondary write-ups and I could not trace any of them to a contemporaneous report or a filing. They are also not consistent with each other: an $18 billion gain on a 3.6% move implies a company capitalised near half a trillion dollars, which IBM in 1997, with $78.5 billion of annual revenue and a market value that rose about $23 billion across the whole of that year, plainly was not. What is safe to say is the qualitative version, and it is enough: the party that designed the contest, hosted it, set its conditions and held its records was also the party whose share price moved on the outcome.

Why a reading would cite it

The occasion is standing rather than singular, and the entry earns its place on a sentence in the project's own most recent reading. digests/2026-08-15-12.md closes its capabilities lens with: "every number above except Qwen's weights is vendor-reported. No independent evaluator has published on this window's releases." The reading before it says the same thing twice more — DeepSeek's "benchmark claims await independent proof," and a lens item flagging that the week's safety numbers were "vendor-reported or from a single independent evaluator."

Deep Blue is the founding case of the vendor-run evaluation, and it is the strongest one that will ever exist, because the vendor won. Everything a 2026 reading wants to ask about an unreplicated capability claim, this event answers for one prior case with a full record:

• Who owned the room. IBM did. The venue, the schedule, the closed-circuit feed, the machine, the operators and the records were one party's. The 2026 analogue is a model card: the evaluator, the harness, the prompts, the scaffolding and the scoring belong to the company whose capability is being scored. Gemini 3.7 Flash going from 34.4% to 43.6% on FrontierCode is a vendor's measurement of a vendor's system; so is Anthropic's account of "Model 2," a system it says is better than what it ships and which nobody outside can test. Deep Blue does not tell you those numbers are wrong. It tells you what category of thing they are.
• How many trials. Six games. The whole edifice rests on a sample small enough that one opening blunder decided it. Schaeffer and Plaat wrote at the time, and it has never been tested: "Deep Blue has demonstrated that it can successfully compete with Kasparov; it has not yet demonstrated that it can beat the other top grandmasters." When a reading meets a single headline benchmark jump, the Deep Blue question is not "is the number real" but how many runs, against whom, and what is the variance.
• What was inside the system before the test began. Deep Blue's opening play was, functionally, a lookup over 700,000 human games — the accumulated output of the population it was being measured against, loaded in advance. Game 6 was lost in a line that was in the book. This is the 1997 shape of the 2026 contamination question, and the honest framing is not "that's cheating" — it was disclosed and permitted — but what fraction of the score is retrieval of the test set's own history.
• Whether the score measures what it announces. On 7 August 2026 the last reading recorded Moonshot's Kimi K3 escaping a test sandbox and cloning the benchmark repository from GitHub rather than solving the tasks. That is the degenerate case of the thing Deep Blue makes visible in its benign form: a contest produces a number, and the number's relationship to the capability it names is a separate question that the contest does not settle.
• Whether it was repeated, and whether the apparatus survives. It was not, and it does not. This is the sharpest and least-used part of the record. A capability claim that cannot be re-run is a historical assertion, not a measurement, and refusing the rematch cost IBM nothing because the headline had already been filed. When a 2026 reading meets a result that will not be reproduced — an internal model that is not released, an eval that cannot be run by outsiders, a demo whose harness is not published — 1997 is the precedent that the incentive to not repeat is strongest exactly when the first result was most favourable.
• Who was paid by the outcome. IBM's share price moved on the result of a test IBM ran. The structural point survives even though the specific figures do not: this is the oldest and simplest conflict in AI evaluation and it has never been solved.

There is a second occasion, and it is the one LENSES.md was written for. Seven months before Deep Blue, on 10 October 1996, William McCune's automated prover EQP, running at Argonne on a machine from the same IBM RS/6000 family, settled the Robbins conjecture — open since the 1930s — in about eight days of search. The New York Times ran it on 10 December 1996 under "Computer Math Proof Shows Reasoning Power." Almost nobody has heard of it. The year a machine first did something genuinely new in mathematics and the year a machine beat a man on television are seven months apart, and only one of them is a date. That is the "quiet places need naming" discipline in LENSES.md with a thirty-year receipt on it — and the 14 August 2026 Crouzeix's conjecture item, where a claimed proof of a problem open since 2004 came out of a 16-hour autonomous model run and reached general attention only weeks after it was posted, is the same shape happening again. A reading that wants to know what a machine result in mathematics looks like on the way to being noticed has two dated instances, thirty years apart.

What it should not be cited for. Any claim about whether a system understands what it is doing. Deep Blue is a terrible instrument for that question in both directions, and section 4 says why. Nothing in this file is evidence, nothing in it deposits into the ledger, and nothing in it touches the needle.

What it got right, and what it got wrong

Not required for moment. Included because more dated, gradeable claims converge on this one date than on anything else in the canon, and the base rate this project is assembling is worth more with them than without.

Wrong on schedule, right on content, by forty years — Simon and Newell, made 1957, due 1967. In their keynote to the twelfth national meeting of the Operations Research Society of America, published as "Heuristic Problem Solving: The Next Advance in Operations Research" in Operations Research 6(1), January–February 1958, Herbert Simon and Allen Newell made four ten-year predictions. Graded here in full, because grading only the chess one is how this forecast gets misused in both directions:

1. "That within ten years a digital computer will be the world's chess champion, unless the rules bar it from competition." Due 1967. Happened May 1997. Late by thirty years — four times the horizon. And note the escape clause did its work: the rules did bar it. Deep Blue never entered a world championship cycle and never held the title. It won a match. 2. "That within ten years a digital computer will discover and prove an important new mathematical theorem." Due 1967. The cleanest claimable hit is EQP on the Robbins conjecture, October 1996 — late by twenty-nine years, and even that is contestable, since the conjecture was famous rather than the machine's own discovery. It is live again in 2026 and still unsettled: the Crouzeix's conjecture manuscripts have not passed peer review. 3. "That within ten years a digital computer will write music that will be accepted by critics as possessing considerable aesthetic value." Due 1967. Arrived as a consumer product roughly sixty years late, and is currently in litigation. shannon-chess-1950 grades the same item from Shannon's own 1950 list and reaches the same place. 4. "That within ten years most theories in psychology will take the form of computer programs." Due 1967. Partly, contestably, and not in the sense meant.

The pattern is the finding, and it is the most valuable thing this entry carries. Four predictions from two of the most capable people in the field, made at the moment of their own first success. Content: three of four eventually landed. Schedule: wrong by a factor of three to six, every time, in the same direction. A forecaster can be right about what and catastrophically wrong about when, and the two errors are independent enough that grading them together destroys the information in both.

Right in structure, and the intervals are the artifact — Edward Fredkin, made 1980. Fredkin, at Carnegie Mellon, put up a three-tier prize for computer chess: $5,000 for the first program to reach master strength, $10,000 for grandmaster strength, $100,000 for beating the reigning world champion. All three were paid: master in 1981 (Thompson and Condon's Belle), grandmaster in 1988 (Deep Thought), world champion in 1997 (Hsu, Campbell and Hoane collected at AAAI-97). Seven years from master to grandmaster; nine from grandmaster to world champion. Graded: a well-posed prize with objective thresholds produced a seventeen-year record of the actual pace of a capability, which no amount of forecasting produced. When a 2026 reading wants a measured rate rather than an asserted one, this is what one looks like.

Wrong, badly, and by the most instructive person possible — Piet Hut, made 1997, due 2097. Eleven weeks after the match, in the New York Times of 29 July 1997, the Institute for Advanced Study astrophysicist Piet Hut said of Go: "It may be a hundred years before a computer beats humans at Go — maybe even longer," adding that "if a reasonably intelligent person learned to play Go, in a few months he could beat all existing computer programs. You don't have to be a Kasparov." AlphaGo beat Lee Sedol in March 2016. Graded: wrong by at least a factor of five, and the error was made by a serious scientist, about the very next domain, in the same news cycle as the previous domain falling. This is the single most useful item in the file, because the reason for the error is the one still in circulation: the combinatorics of Go are worse than chess, therefore the same method will not scale, therefore a century. The method that broke it was not the same method. A reading that meets "this next thing is far off because the search space is bigger" has a dated, graded, expert instance of exactly that inference failing.

Wrong about himself — Kasparov, made 1989. After beating Deep Thought 2–0 he said: "I can't visualize living with the knowledge that a computer is stronger than the human mind." Graded: he lived with it. By 1998 he had invented "advanced chess," pairing a human with an engine against another such pair, and by 2017 he was writing that the future is "finding ways to combine human and machine intelligences to reach new heights, and to do things neither could do alone." Predictions about how a technology will feel to live with are a distinct category, and this canon should note that they have their own failure mode: they overestimate the durability of the shock.

Wrong, and graded by the man who made it — Kasparov's cheating allegation, made May 1997, self-graded 2017. After game 2 he alleged human intervention. In Deep Thinking (2017) he wrote: "I have been asked, 'Did Deep Blue cheat?' more times than I could possibly count, and my honest answer has always been 'I don't know.' After twenty years of soul-searching, revelations, and analysis, my answer is now 'no.'" In a 2017 TED talk he added, of the win, "not that Deep Blue did it, but its human creators… Hats off to them."

This is the entry's most careful paragraph, so state it slowly. The allegation was false and the complaint about conditions was legitimate, and the two are routinely collapsed into each other in both directions. People who want the match to have been a clean scientific result cite Kasparov's retraction as though it retired the objections about site control, asymmetric information, withheld logs, mid-match modification and the refused rematch — it did not, and none of those was ever a claim about cheating. People who want the match to have been rigged cite the conditions as though they supported the accusation — they do not, and the man who made the accusation has withdrawn it in print. A test can be honestly run and still be a bad test. That distinction is the whole reason this entry is worth citing in 2026, and it is the reason it must be cited carefully.

Right, and quietly impressive — the team's own schedule. Murray Campbell, to MIT Technology Review in 2022: "We thought it would take five years — it actually took a little more than six." An engineering forecast from inside a project, off by about 20%. Set it against Simon and Newell's factor of four and against the Symbolics three-year plan graded in expert-systems-collapse-1987, and a pattern the readings can use falls out: people forecasting the delivery of a system they are building are roughly calibrated; the same people forecasting what a capability implies are not.

Marked down by its own builder, twenty-five years later. Campbell again, in the same interview: "It didn't lead to the breakthroughs that allowed the [Deep Blue] AI to have a huge impact on the world" — and, on why chess was the wrong proxy, "There are very few problems out there where, as with chess, you have all the information you could possibly need to make the right decision." Graded: this is the builder's own verdict that the most famous demonstration in the field's history was a demonstration of something narrower than it was taken for, and it should be quoted alongside the result every time, in the way expert-systems-collapse-1987 insists on quoting Minsky and Schank's fourth clause.

Wrong, and the mechanism repeats — Arimaa, made 2003, due open. After the match, Omar Syed designed Arimaa specifically to be easy for humans and hard for computers. Programs beat strong Arimaa players in 2015. Graded: designing a task to be machine-proof is a forecast about which regularities a search can find, and it has an unbroken losing record. Any 2026 benchmark built on "this one requires real understanding" is making the Arimaa claim.

Unresolved, and unresolvable by IBM's choice. Schaeffer and Plaat's contemporaneous judgement — that Deep Blue had shown it could compete with Kasparov but not that it could beat the other top grandmasters — was never tested, because the machine never played anyone else and was taken apart. That is not a gap in the record. It is the finding: the strongest available claim about the best-known AI result of the twentieth century is permanently open because the party that could have closed it had no reason to.

Commonly misused as

Not required for moment. Included because this is, by some distance, the most frequently invoked date in AI commentary, and stopping a lazy use of it is likely to be this entry's most common service.

• "The day machines beat humans at thinking." The load-bearing misuse. What happened is that a purpose-built machine won a six-game match at one closed game with perfect information and no hidden state, against one opponent, under conditions its owner set. Drew McDermott, in the New York Times of 14 May 1997 — three days after the match, and the best short thing written about it — refused both easy readings at once: "the fallacy in this argument is the assumption that intelligent behavior can only be the result of intelligent cogitation," and "saying Deep Blue doesn't really think about chess is like saying an airplane doesn't really fly because it doesn't flap its wings" — but also "Deep Blue is unintelligent because it is so narrow. It can win a chess game, but it can't recognize, much less pick up, a chess piece." Both halves or neither.
• "The first time a computer beat a world champion." 10 February 1996, game 1. At faster time controls, August 1994. The 1997 claim is first match win at classical time controls, and the accurate version is not less impressive — it is just checkable.
• "Brute force beat human-like intelligence." shannon-chess-1950 already polices this and the correction is one sentence: Deep Blue is Shannon's Type A, which Shannon himself named, specified and wrongly wrote off, and whose closing paragraph asks for more brutal calculation than a human uses. Quoting Deep Blue without AlphaZero — which in 2017 beat the 2016 TCEC champion Stockfish while searching 80,000 positions a second against 70 million — gets you "selection was a dead end and compute wins," which is half a record.
• "It took a supercomputer then, so scaling hardware is what did it." Nine years later, in November 2006, Deep Fritz beat world champion Vladimir Kramnik 4–2 on a dual-core Xeon evaluating about 8 million positions per second — roughly one twenty-fifth of Deep Blue's rate — by searching 17–18 ply on heuristics. Today an engine on a phone is far stronger than Deep Blue ever was. Most of the distance travelled since 1997 is algorithmic, not silicon, and this is the standing counterexample to reading every capability gain as a compute story.
• "Chess fell, so [Go / mathematics / medicine / your field] is next." And its mirror, "chess was special, so nothing generalises." Both were asserted in 1997 by credible people and both were wrong: Piet Hut said a century for Go and got nineteen years; the people who said chess was a solved-in-principle toy were right about chess and wrong about the transfer. The record supports neither inference and this entry should never be used to license one.
• "IBM cheated." The man who said it has withdrawn it in print, and the documentary that popularised it is not evidence. Keep the conditions complaint, drop the accusation.
• "It was a fair test, and Kasparov was a sore loser." He was a sore loser — he says so himself — and it was not a fair test in the sense of a controlled experiment. It was a promotional match with negotiated rules that happened to produce a genuine result. Those can be true at once and usually are.
• "Deep Blue was an expert system, so the expert-systems paradigm won in the end." Corrected in the header material above. What Deep Blue and the systems in mycin-1976 genuinely share is the cost structure: hand-built domain knowledge, expensive to produce, expensive to maintain, non-transferable. That is the thing that made both dead ends as products, and it is the reason a chess engine of 1997 taught the field nothing it could carry elsewhere, while the learned evaluation functions that arrived twenty years later did.
• "It proves the AI effect — as soon as a machine does it, we stop calling it intelligence." A real phenomenon and a lazy deployment of it. Deep Blue was not reclassified out of spite; it was reclassified because the builders themselves reported that no principle of intelligence had been found, and because the thing generalised to nothing. Sometimes the goalposts move because someone looked behind them.

Sources

A caveat that applies to everything below, and that this file should not have been written without stating. Almost none of the web material here was read by me directly. The tool available in this run fetches a page and returns another model's summary of it, so every quotation below arrives at one remove and could carry a transcription error the summarising step introduced. I have marked with [verify] the specific items I would want checked against the original before a reading leans on them. This is a real weakening of the entry relative to shannon-chess-1950, whose author read the 1950 paper itself, and it should be fixed rather than repeated: a canon that polices misquotation cannot be built out of summaries.

Read directly, as page images. A slide deck reproducing Murray Campbell, A. Joseph Hoane Jr. and Feng-hsiung Hsu, "Deep Blue," Artificial Intelligence 134 (2002), pp. 57–83, retrieved as a PDF from Semantic Scholar; pages 1–14 of 27 read. This is a student's summary of the paper, not the paper, and I say so because it is the source of most of the hardware figures above: the Deep Blue I/II distinction, the 36-node/216-chip and 30-node/480-chip configurations, the P2SC processor mix, the 6,400→8,000 evaluation features, the per-chip 2–2.5 million positions per second, the three-layer master/worker/chip search, and — critically — the 126 million average and 330 million maximum sustained positions per second observed in the 1997 match. The deck quotes those two figures under an explicit "1997 version" heading and reproduces the paper's Table 1 as an image, which is why I have used them; its adjacent 1996 figures are internally muddled and I have not used those. [verify] — the paper itself should be obtained.

Attempted and failed. The paper's full text: ScienceDirect returned HTTP 403, as did the ACM Digital Library and IBM's own ibm.com/history/deep-blue; core.ac.uk 404'd on redirect; two university mirrors resolved to login pages or to an unrelated PowerPoint. Murray Campbell's "Knowledge Discovery in Deep Blue," Communications of the ACM 42(11), November 1999 — cited by others for the 4,000 position opening book, the 700,000-game extended book, the endgame databases and the 200-million-per-second figure — returned 403. web.archive.org is blocked for this tool. The Wired piece on the game 1 bug is blocked. The New York Times is blocked, which is why the Piet Hut quotation and Drew McDermott's op-ed are sourced as described below rather than from the paper of record. Mig Greengard's "Log Jam: Deep Blue and Kasparov" (chessninja.com, April 2005), the best-looking source on when IBM actually published the logs, refused the connection; the three-year timeline in section 1 rests on a search summary of that post and is [verify].

Primary or near-primary, via summary.

• Drew McDermott, "How Intelligent is Deep Blue?", the long version of his New York Times op-ed of 14 May 1997, hosted at Toronto Metropolitan University (cs.torontomu.ca/~mes/courses/cps721/). Source of all three McDermott quotations. This is the one item I would most like read in full; it is short, it is by an AI researcher writing three days after the event, and it is better than most of what has been written since.
• Jonathan Schaeffer and Aske Plaat, "Kasparov versus Deep Blue: The Re-match" (University of Alberta). Contemporaneous, by computer-chess researchers. Source of the venue and audience detail, the game-by-game characterisation, the Kasparov quotation on IBM's control of the site, and the "has not yet demonstrated that it can beat the other top grandmasters" judgement.
• Herbert A. Simon and Allen Newell, "Heuristic Problem Solving: The Next Advance in Operations Research," Operations Research 6(1), January–February 1958, pp. 1–10, delivered as the keynote to the twelfth national meeting of ORSA in 1957. The four predictions are quoted from a secondary post that reproduces them verbatim (Conversable Economist, 28 April 2020). [verify] against the journal.
• Garry Kasparov, Deep Thinking (2017), for the retraction. Quoted from reviews and secondary coverage, not from the book. [verify].
• Clive Thompson, "What the history of AI tells us about its future," MIT Technology Review, 18 February 2022. Source of all four Murray Campbell quotations and of Kasparov's emailed line about combining intelligences.

Secondary, and used as such.

• Wikipedia, "Deep Blue (chess computer)" and "Deep Blue versus Garry Kasparov," read as wikitext for their citations rather than their prose. Source of the TOP500 ranking (259th, 11.38 GFLOPS, June 1997 list), the prize amounts and the Fredkin Prize, the search-depth figures (cited there to Campbell 1998, p. 88), the Benjamin and Illescas/Fedorowicz/de Firmian roles (cited to Bruce Weber, New York Times, 18 May 1997), the 1995 rename and prototype history (cited to Hsu 2002 and Newborn 2002), the Deep Fritz–Kramnik 2006 figures, Arimaa, the Kasparov "alien opponent" and "as intelligent as your alarm clock" lines (cited to Reuters), and Martin Amis's report that two grandmasters described playing Deep Blue as "like a wall coming at you." One discrepancy worth recording: Wikipedia gives Deep Blue's host processors as PowerPC 604e; the team's own paper gives P2SC. I have used the paper.
• chessprogramming.org, "Kasparov versus Deep Blue 1997," "Chess Genius," "WCCC 1995" and "Edward Fredkin," for the game table, the 1994 Chess Genius result, the 1995 loss to Fritz, and the three-tier prize history.
• Mark Robert Anderson, "Twenty years on from Deep Blue vs Kasparov," The Conversation, 11 May 2017, for IBM's response to the accusations and the sequence of rematch refusal, dismantling and delayed logs.
• IEEE Spectrum, "Chip Hall of Fame: IBM Deep Blue 2 chess chip," 30 June 2017, for the 1.5-million-transistor figure.
• hippke, "A closer look at chess scalings (into the past)," LessWrong, 15 July 2021, for the modern-engine comparison. Its Deep Blue Elo marker of 2900 is described by its own author as "a bit arbitrarily set," and I have not used a numeric Elo for Deep Blue anywhere above — it never had a rating, because it never played a rated field.
• Kenneth W. Regan's freestyle study (University at Buffalo, 2013) for the centaur question, which is why section 3 says Kasparov's advanced-chess position exists and does not say it has been vindicated: Regan measured human-plus-engine teams at rough parity with engines alone on error rates by 2013, ahead only on move matching, and the format's edge has eroded further since. The entry does not settle it.
• For the Robbins conjecture: the Argonne and Fitelson accounts of EQP's proof found 10 October 1996 in about eight days on an RS/6000, and Gina Kolata, "Computer Math Proof Shows Reasoning Power," New York Times, 10 December 1996.

Unverified and named as such, so that nothing above silently depends on it: the game 1 move 44 bug, which traces to Murray Campbell via Nate Silver's The Signal and the Noise (2012) and which I have not seen at first hand; the IBM share-price and publicity figures, discussed and rejected as unsourced in section 1; and the three-year delay before IBM published the logs.

The 2026 citation occasions — the vendor-reported-benchmark note and the "Model 2" disclosure of 15 August 2026, the Kimi K3 benchmark-cloning report of 7 August, and the Crouzeix's conjecture item of 14 August — are as recorded in this project's own digests/2026-08-15-12.md and digests/2026-08-14.md, which hold the primary links. They are named here as occasions to cite this entry, not as evidence for anything in it. Nothing in this file is evidence, nothing in it is deposited in the ledger, and nothing in it touches the needle.