Computer Chess, from a DBA's Chair
In mid 2003 I was a young college student — studying algorithms, C++, and AS/400 operations — with a 2.4 GHz Pentium 4, a stack of external drives, and a chess habit. I cranked the hash as high as the machine would allow, hand-edited my opening book so the engine always steered into the Ruy Lopez, hoarded endgame tablebases across every drive I could afford, and ran engine gauntlets at every time control just to see what my setup could do. This is that curiosity, more than twenty years later — the engineering side of chess, looked at the way I now look at a slow query.
Michael Paycer
The Yahoo Chess era, and a habit that never quite ended
If you were online for chess between roughly 2000 and 2008, you remember the Yahoo! Games chess rooms — a Java applet, unmoderated lobbies, and engines finally strong enough to run on any PC sitting in the corner. That was where a lot of us met computer chess for the first time: not as a research topic, but as a thing you ran, tuned, and argued about at 2 a.m. The engines were how I loved the game. The Ruy Lopez and the Caro-Kann were the openings I fell for; the machine was how I got to play them a thousand times over.
Two decades on, I spend my days tuning Microsoft SQL Server — diagnosing slow queries, reading execution plans, deciding what to cache and what to precompute. And the funny thing is how much of computer chess reads like that job. A transposition table is a plan cache. An opening book is a warmed result cache. A tablebase is a precomputed, indexed lookup. Alpha-beta pruning is skipping the partitions you've already proven can't hold the answer. So this cluster is a passion project with a professional accent: how chess engines are actually built, told by someone who never stopped tinkering. Every page leads with the human — the player, the tinkerer, the college student with the Pentium 4 — and then goes technical.
Everything an engine does is either search or evaluation
If you take one idea from this whole cluster, take this one — it structures every page that follows. An engine has exactly two jobs. Search is how far and how cleverly it looks ahead through the tree of possible moves. Evaluation is how well it judges a single position once it stops looking. Strength is the product of the two: a deep search wedded to a wise evaluation. Everything else — bitboards, hashing, neural networks, tablebases, opening books — is in service of doing one of those two things faster or better.
The whole cluster hangs off this split. When a page talks about NNUE, it's improving evaluation; when it talks about pruning, it's improving search.
The same instincts, applied off the clock
This isn't a stretch metaphor bolted on after the fact. The reason computer chess grabbed me — and the reason it still does — is that the engineering rhymes, almost line for line, with the database work I do for a living. You'll see this callout again on every page in the cluster; here's the whole map at once:
- Search vs. evaluation = diagnose before you prescribe. Look ahead, then judge.
- Alpha-beta pruning = skipping the partitions you've proven can't contain the answer.
- Transposition table = a plan / result cache, keyed by a hash of the position.
- Opening book (.ctg) = a warmed result cache for the moves you've seen a million times.
- Endgame tablebases = a precomputed, indexed lookup — the engine stops calculating and reads the answer.
- Fishtest & SPRT = statistically-gated regression testing. Nothing merges without proof.
- Bitboards = bit manipulation as a first-class citizen.
- Engine provenance = inherited-code governance, derivative works, and licensing discipline.
Any DBA who has watched a vendor benchmark fall apart under scrutiny already has the instinct that runs through the whole set: trust, but verify.
What each page covers
Nine deep dives plus a shared glossary, each written to stand alone. Start with the history if you want the story, or jump straight to the machinery. Pages go live as they ship — this is an actively-building cluster.
-
From Yahoo Chess to Stockfish 18
History · 2000 → 2026Twenty-five years of engines, and the first-hand Yahoo-era material you won't find on a wiki: the ytoics ICS bridge, the room backdoors, and the rating exploit that hit a 16-bit ceiling.
-
Search vs. Evaluation
How an engine thinksMinimax, alpha-beta, iterative deepening, move ordering, quiescence, and the horizon effect — plus classical handcrafted evaluation. The conceptual core of the whole cluster.
-
NNUE: Neural Evaluation
How neural nets took overThe AlphaZero shock, then the shogi import that let a neural evaluation ride a classical search. The incremental-update trick, HalfKA features, and Threat Inputs in Stockfish 18.
-
Solving the Endgame
Tablebases · retrograde analysisHow perfect endgame results are computed by working backwards from checkmate — and what it meant in 2004 to ration which 6-man tables were worth the disk across a stack of external drives.
-
The Clone Wars
Code provenance · the ICGA caseFruit → Rybka → Strelka → Ippolit → Houdini → Fat Fritz 2. A litigious lineage of borrowed code, a GPL lawsuit, and why open, auditable engines outlasted every brilliant closed one.
-
Scripting Stockfish with UCI
Automation · the practical crossoverDriving an engine from the command line and PowerShell, running engine-vs-engine gauntlets with cutechess-cli, and why Fishtest/SPRT is exactly how you'd validate an index change.
-
How Engines Were Built, 2000 → Modern
Architecture · five construction shiftsFrom the single-threaded classical monolith with hand-tuned weights to distributed, statistically-gated development and a trained neural evaluation. The construction story, not the how-it-works story.
-
Inside .ctg: Opening Books
Books · steering the engineWhat a .ctg book really is, how manufacturers and human authors built them, and how an ordinary player hand-edited the tree to force the engine into their lines. Plus shared Shredder learn files.
-
Poke the Bear: Unsettled Arguments
Disputes · the caveat-sinkThe claims that aren't settled — Deep Blue's unknowable strength, the AlphaZero–Stockfish match conditions, Rybka's contested originality, and why cross-era Elo is barely real.
-
Chess Engine Glossary
Shared reference · every termBitboards to NNUE, quiescence to Syzygy, UCI to SPRT — the reference that every page links back to. Organized by theme so related ideas sit together.
Where computer chess stands right now
The crown moved from paid to free. In the Yahoo era the strongest engine was something you bought — Fritz, Shredder, Junior. Today the strongest engine on earth is open source, and the commercial engines sell on experience and content instead of raw Elo.
Stockfish 18
January 2026 · ~3,653 CCRL · SFNNv10 net with Threat Inputs · open source, Fishtest/SPRT-developed · reigning TCEC champion. The strength leader.
Fritz 21
May 2026 · ~3,620 Elo (+40 over Fritz 20) · adds Chess960 · ChessBase's commercial engine. Excellent — but open source now edges it out.
Mega Database 2026
11.7M+ games, 1475–2025. The raw corpus that feeds both the human study material and the engine's playable book.
Fritz Powerbook 2026
1.7M high-class games distilled into ~25M positions — the same .ctg format I hand-edited in 2004, now at a scale no single player could curate.
Figures captured July 2026 from stockfishchess.org, en.chessbase.com, and the CCRL 40/15 list (computerchess.org.uk). Engine ratings are build-, list-, and date-specific — see the Disputes page for why a single number is never the whole story.
Computer chess — FAQ
How do chess engines actually work?
Everything an engine does is either search — how far ahead it looks through possible moves — or evaluation — how well it judges a single position. Classical engines paired an alpha-beta search with a handcrafted scoring function; modern engines keep the fast search but replace the scoring with a small neural network (NNUE). Tablebases perfect the endgame and opening books handle the first moves, but search and evaluation are the whole game.
What's the strongest chess engine in 2026?
Stockfish 18 (January 2026), at roughly 3,653 on the CCRL 40/15 list — free, open source, and the reigning TCEC champion. ChessBase's commercial Fritz 21 is about 3,620 Elo: superb, but open source now wins on raw strength.
Are chess engine ratings the same as human Elo?
No. Engine ratings come from engine-vs-engine lists (CCRL, SSDF, CEGT) on standardized hardware and don't map onto the human FIDE scale. They also shift with hardware, time control, and build — the same engine can differ 100+ points between lists — so an engine's "Elo" only means something with its build, list, and test date attached.
Why is a SQL Server DBA writing about chess engines?
Because the engineering rhymes. A transposition table is a plan cache, a tablebase is an indexed lookup, alpha-beta is partition elimination, and Fishtest is regression testing. Computer chess is where I first met those ideas — years before they became my day job.
The player and the tinkerer
This cluster is the engineering side of a game I've loved from the other direction for just as long. If you came for the openings — the Ruy Lopez, the Caro-Kann, the champions and their famous games — that's the human side, and it's all next door.