Exactly thirty years ago, in February 1996, the reigning world chess champion lost a game to a computer for the first time. Garry Kasparov, the thirteenth world champion and the undisputed world No. 1, resigned to an IBM machine in Philadelphia — but won the match, 4–2. A year later, in New York, IBM brought an upgraded Deep Blue and took the rematch. It went down in history as the moment when “computers caught up with humans.” Nicholas Carlini, an AI security researcher at Anthropic, asked a more arrogant question: had they, though? More precisely — did it really take a multimillion-dollar supercomputer, or would an ordinary desktop PC, the kind sitting in the window of every computer shop in 1997, have been enough?
For the anniversary of the Pentium Pro processor (it turned 31 — it was released on November 1, 1995), Carlini bought one second-hand, “vintage,” as the listing put it, installed the modern chess engine Stockfish on it, and got a program rated around 3300 Elo at classical time controls — higher than any human has ever played in history. His conclusion, which I’ll unpack with numbers below: yes, Deep Blue was unnecessary. A Pentium tower under an engineer’s desk would have done.
To appreciate the audacity of that conclusion, recall what Deep Blue was as a machine. In its final version — thirty 200 MHz PowerPC 604e processors (the 1996 version had 120 MHz 604s) and 480 custom chips that did only one thing: generate chess moves and evaluate positions. This was not “a computer running a program”; it was an architecture with chess soldered into the hardware. Speed — more than 200 million positions per second. Ranked 259th in the world among supercomputers at the time it was built.
A 200 MHz Pentium Pro, in hardware terms, was an application-level dwarf: an Intel VS440FX motherboard, 64 megabytes of SIMM memory on a 66 MHz bus. At launch, the processor cost $1325, four 16 MB sticks another roughly $2000; with the case, monitor, and graphics card, a full configuration came to 4–5 thousand dollars — a serious workstation, certainly, but not a supercomputer. And just a year later, by the 1997 rematch, such a machine would have cost around a thousand: the $400 Pentium MMX had arrived, and nobody in their right mind was buying the top-end Pro.
The difference was not just megahertz. Deep Blue’s chips were like the blitters in nineties graphics cards: specialized logic, devastatingly fast at one operation and helpless at everything else. Want to change the program? Resolder the board. So Carlini’s experiment conceptually comes down to a question nobody was seriously asking in 1996: what if Deep Blue’s entire scope could fit inside an ordinary program?
Kasparov’s lost first game of 1996 was historic: a reigning world champion resigned to a machine under tournament conditions. Then came an anecdote straight out of “it wasn’t the machine that broke the man, it was uncertainty.” In the first game of the 1997 rematch, Deep Blue played 44.f6 — a move so unexpected that Kasparov refused to believe a machine had calculated it: he was convinced there was a human behind the box, inserting an “inhumanly human” continuation. Later, the explanation turned out to be as prosaic as possible: the engine’s search had stalled at that moment, and the program returned a fallback fail-safe move — precisely the one it would have played without any intrigue at all. In the second game, Kasparov resigned in a position that every analysis says was still a draw, and openly accused IBM: a grandmaster consultant, he claimed, was pulling the levers behind the machine. He lost the match and later admitted that he had lost it psychologically.
It is the best anecdote in the whole story because it flips the familiar narrative on its head. The machine had no psychology — it beat the human being with his strongest resource: the ability to defeat himself. Deep Blue was merely a silent box, but a box that moves “like it’s alive” breaks the living faster than the living can break the box.
It is easy to mock IBM for the expense, but one has to admit: in 1997, the alternative truly was not obvious — for a reason that is easy to forget. By then, half a century of attempts to make a chess program had already happened. Alan Turing and his economist friend David Champernowne wrote the Turochamp algorithm on a sheet of paper as early as 1948: a machine capable of running it simply did not exist, and the program’s first “games” were played on paper, with a human calculating the moves by hand. Claude Shannon, in 1950, divided future programs into two types — “brute force,” with exhaustive search, and “selective” search with heuristics — and, as it turned out, humanity spent the next twenty years choosing between those two branches.
Then came the era of modest machine victories. In 1967, Richard Greenblatt’s MacHack beat a human in an official tournament game for the first time. In 1968, grandmaster David Levy bet 1250 pounds on himself, wagering that no program would beat him within ten years — and only won that bet in 1978, defeating Chess 4.7 in a match in Toronto. The Soviet branch did not lag behind either: in 1966–1967, the Institute of Theoretical and Experimental Physics entered its program against Stanford’s American Kotok-McCarthy — the first intercontinental computer match, played by telegraph — and won it. And in 1974, the Kaissa program from the Institute of Control Sciences became the first official world champion among chess programs; Botvinnik himself also consulted for its creators. Its creators had Mikhail Botvinnik himself as a consultant, a man who spent his entire life wrestling with a program that would play “like a master”: with plans, attacks, the “trajectories” of human thought. His Pioner remained a closed Soviet project, while the world went with brute force.
And here comes the least obvious turn, the reason this story is worth telling. After 1997, computer chess entered, if you can call it that, an era of “soft stagnation.” Hardware grew exponentially: clock speeds rose twentyfold over ten years, then cores joined the party. But the Elo rating of the best engines crept upward reluctantly. Developers refined the same principle — alpha-beta search with a hand-crafted evaluation function (a giant set of weights along the lines of “isolated pawn — minus 12”), written by hand and polished over decades.
The final chord of that era could easily have restored the status quo for another decade. By the end of the 2010s, Stockfish played at around 3400 Elo on the CCRL lists — unreachable for any human, comfortably above Deep Blue. Yes, over twenty years — enormous growth. But here is the question Carlini asked himself, one that had no honest answer before his experiment: how much of that growth was “new thinking,” and how much was “new hardware”? The best engine available on 1996 desktop hardware came in at roughly 2300 Elo; Deep Blue delivered 2700; modern Stockfish — around 3700. It seemed obvious: even if the architecture was different, all the growth from 2300 to 3400 rested on raw speed. Or perhaps it did not.
The answer came from the side, from a game the average European chess player does not play — shogi. In 2018, Japanese researcher Yu Nasu described a trick called NNUE (efficiently updatable neural network): a small neural net that does not recalculate a position’s evaluation from scratch after every move, but carefully inches toward the new evaluation by updating a couple of dozen numbers. For shogi — with its outrageous branching factor — it proved a lifeline. Motohiro Isozaki, the author of one of the strongest shogi engines, predicted that porting the idea to chess would add around one hundred Elo points to Stockfish — and the prediction seemed so implausible that almost nobody believed it.
Then, in 2019, a programmer known as nodchip (Hisayori Noda), “just to give back to the community,” bolted NNUE onto Stockfish 10 as a proof of concept. Within mere months of its integration into the main branch (August 2020, Stockfish 12), it became clear: the network, despite nearly halving search speed, still made the engine stronger. It was the largest algorithmic shift in computer chess in more than twenty years — and it came from a game nobody in the industry had been thinking about.
And now we have everything we need to answer the original question quantitatively.
What does “installing a modern engine on a 1995 machine” mean technically? Stockfish 17 is designed for 64-bit processors, vector instructions, multicore systems, and a modern C++ library. The Pentium Pro is a 32-bit chip without SIMD, with one core and 64 megabytes of memory. Carlini’s team (to be honest: it was LLMs — Fable 5.1, Opus 5.5, and GPT-6; the author wrote the article itself) rebuilt the bitboards as pairs of 32-bit integers, cut multithreading, endgame tablebases, and the 71-megabyte neural net, replacing it with a tiny one. The first such “naive” port looked modest: 2496 Elo at a pace of one second per move.
Then came alchemy: tuning search parameters with SPSA (stochastic optimization over thousands of self-play games), replacing the evaluation network with a faster one, “correction history” in the search. The most intriguing part is the final step in the results table: the last chunk of plus 35 Elo came with no code changes at all — only a compiler swap and bytecode alignment to a 32-byte boundary, so the search loops would fit perfectly into the processor’s cache lines. The result: 2846 at one second per move, 3109 at ten, 3180 at thirty. Given that in a classical game the engine gets around six minutes per move, Carlini extrapolated the machine’s strength to 3300 Elo.
But the essence of the experiment, the reason the whole project was undertaken, is the meeting of engines from different decades. Carlini upgraded in the other direction too: he took Crafty 11.13 — the strongest “people’s” engine of December 1996 — and added every idea accumulated since then: fail-soft, late move reductions, correction history, a small neural network trained on modern Stockfish (without it, self-play would have taken roughly four thousand processor-years of 1996-level cores — on a modern desktop, the teacher generated the data in a matter of days). The 1996-vintage Crafty, updated with the knowledge of 2026, rose from roughly 1999 to 2718 Elo at the same one second per move — without a single hardware upgrade.
And now — the showdown: modern Stockfish on a 1995 Pentium Pro against upgraded 1996 Crafty on a modern Ryzen 7. A 100-game match: 78 wins, 17 draws, 5 losses — in favor of the 1995 machine. Modern software on thirty-year-old hardware crushes thirty-year-old software on modern hardware.
All told, the gap between the best engine of 1996 and current Stockfish is 1335 Elo points. Of those, around 720 are pure algorithmic progress, and only the remaining 615 come from the twentyfold growth in processors. More than half of chess AI’s evolution lives not in silicon but in programmers’ heads: in trained networks, incremental evaluator updates, and clever pruning heuristics.
Let me also mention the most honest part of Carlini’s article — where his method limps. Elo numbers are relative: they work well within a single ratings ladder. Carlini measured engines against their neighbors in CCRL, while Kasparov lived on the FIDE ladder; there is no reliable exchange rate between those ladders — like currencies of countries that do not trade with each other. The figure of 3300 (plus or minus 87) is an extrapolation from fast time controls, and debates over how much an engine gains when given more time could fill separate articles. But with a gap of roughly 500 points between the machine and Kasparov, the uncertainty in the ladders is just background noise. Carlini estimates the score of a match against the 1997 champion as “roughly 9 out of 10 in favor of the Pentium Pro,” and invites any 2700+ super-grandmaster to come and try: the engine is still online today and accepting challenges.
If you take one sentence away from Carlini’s experiment, it should be his own closing line: “The first time we do anything, we do it inefficiently.” First — scale; then — intelligence. Deep Blue is magnificent scale: millions of dollars, half a thousand custom chips, a machine room. A thirty-year-old desktop Pentium is intelligence, refined by another twenty years of other people’s ideas — from shogi programmers to enthusiasts rebuilding bitboards as 32-bit pairs.
And hanging over this is the question of why these forty pages are worth reading even if you do not care about chess. In 1997, looking at Deep Blue, a perfectly reasonable person could have concluded: “computer chess has matured; from here on, it is only a matter of adding more silicon.” That was wrong, spectacularly so: the biggest advances lay ahead, and they did not come from hardware. The same suspicion hangs over large language models today: we compare them by cluster size and assume that whatever they cannot do, they cannot do “in principle.” Chess history gives a wink: very little is truly impossible in principle — there are only things not yet done. Progress in thinking more often arrives dressed as somebody else’s idea from a neighboring game, falling into your code, than as a new server. Carlini makes the analogy directly and urges us not to draw grand conclusions from a toy experiment — but the very fact that the whole project (the ported engine, a homemade OS on a floppy disk, an HTTP server on vintage hardware) was assembled by several language models in a matter of days works as a metaphor better than any conclusion could.
The Pentium hums below. And the man who in 1997 could not forgive himself for losing to a box, in some sense left this story not as a loser: without his drama, we would not have hunted so greedily for an answer — and perhaps nobody would have put that answer into numbers even now.
Sources: Nicholas Carlini, “Algorithmic improvements in chess” (nicholas.carlini.com, November 1, 2026); Wikipedia — Deep Blue versus Garry Kasparov, Deep Blue (chess computer), Computer chess, Nicholas Carlini, Kaissa; chessprogramming.org — Kasparov versus Deep Blue 1996, Stockfish NNUE, World Computer Chess Championship; stockfishchess.org — “Introducing NNUE Evaluation” (August 2020); github.com/official-stockfish — Stockfish 12 release.