FRITZ 20: When a Chess Engine Learns to Teach

A WEDNESDAY NIGHT IN SHANGHAI

At 11:40 on a Wednesday night in Shanghai, I opened FRITZ 20 for the first time. The apartment was quiet enough that I could hear the laptop's cooling fan running at full speed. On screen, a chessboard appeared, black and white, immaculate, offering no greeting and making no sound. I pushed my e-pawn to e4. The machine answered faster than one breath of mine.

The first thing I noticed was not strength. It was silence.

In 2026, in a cramped computer room, I watched a chess engine answer my move on a monochrome monitor for the first time. I sat there for a long while, not because I was amazed by its calculation, but because of a strange feeling: sitting across from me was something that would never sigh. No hesitation. No regret. None of that moment a strong human player still has, when a finger hovers over the queen, knowing the move is wrong, and makes it anyway.

Thirty-five years later, FRITZ 20 is introduced with three short lines: your personal chess trainer, your toughest opponent, your strongest ally. Those three lines are worth dissecting, because they mark a turning point for an entire industry of the mind. The race for raw strength ended long ago. What is being sold now is the ability to teach, which is a harder problem and a far less glamorous one.

FROM HAMBURG 2026 TO BONN 2026

ChessBase was founded in Hamburg in 2026 by Frederic Friedel and Matthias Wüllenweber, with a modest idea: digitize games so players could search them instead of flipping through opening books. Fritz 1 followed in 2026, tied to the names of Dutch player and programmer Frans Morsch and Germany's Mathias Feist. Fritz's difference was never raw power. It was designed for humans: a clear interface, adjustable levels, and a commercial philosophy of selling a sparring partner rather than a calculator.

Then came the Man versus Machine years. In October 2026, in Bahrain, world champion Vladimir Kramnik drew Deep Fritz 4-4 in a match billed as Brains in Bahrain. Four years later, between November and December 2026, in Bonn, Deep Fritz beat Kramnik 4-2, the first time a reigning world champion lost an official multi-game match to a chess program.

I covered Bonn as a commentator, and what I remember is not a tactic but a dry line of text in the corner of the screen. The machine announced it could mate in one. Kramnik, already lost, played on until the mate arrived. Collective memory recalls that moment as a human collapse. I see something else there, and I will return to it.

The key point about 2026-2026 is this: after Bonn, the question of whether machines are stronger than humans was settled. And when a question is settled, a market must find another one.

RAW STRENGTH BECAME A CHEAP COMMODITY

Around 2026, the gap between the strongest engine and the strongest human was only a few dozen Elo points, at times effectively zero. Today, public engine rating lists such as CCRL and TCEC place leading Stockfish builds above 3,600 Elo under tournament conditions. Meanwhile the human record, the 2,882 Elo Magnus Carlsen reached in May 2026, has stood untouched for more than a decade.

That gap of roughly 800 points is not a technical detail. It is an economic event.

When calculation becomes free, selling a strong engine is like selling bottled water. Everyone has it. Nobody pays for it. Which is why FRITZ 20's marketing does not say our engine is strongest. It says trainer, opponent, ally. Three nouns about the relationship between a person and a machine, not three adjectives about the machine.

WHAT FRITZ 20 ACTUALLY SELLS

According to the publisher, FRITZ 20 targets two very different groups: players taking their first steps into serious training, and players already competing at tournament level. The shared promise is more efficient, smarter and more individualized training.

Broken into components, that means tactics drills, opening repertoire management, endgame material, sparring against adjustable virtual opponents, game analysis with error feedback, and a progress-tracking layer.

Almost all of these existed fifteen years ago. What is new is how they are linked, and how automated the intervention has become. The system does not merely offer exercises; it decides which exercise comes next, based on an individual error profile. That is sound pedagogy. It is also what any experienced coach has always done by hand.

A dry detail to keep our feet on the ground: a full installation runs several gigabytes, requires Windows 10 or later, and ships with ChessBase's enormous annotated game database. A full license costs roughly the same as a few lessons with an experienced international master. That comparison forces a question that has nothing to do with technology: does the buyer need a tool, or someone to look at their games and tell them they were a coward?

THE SCIENCE BEHIND THE WORD ADAPTIVE

1. No. 1885: Hermann Ebbinghaus publishes his memory experiments and draws the forgetting curve. Reviewing at the right moment flattens it. That is the basis of every spaced-repetition system.

2026: Henry Roediger and Jeffrey Karpicke publish on the testing effect. Retrieving knowledge produces far more durable memory than rereading it. In other words, harder practice, better retention, even when it feels worse in the moment.

2026: Anders Ericsson and colleagues describe deliberate practice, progress coming from training at the right edge of difficulty with immediate feedback.

Combine the three and you get the whole formula an adaptive training system must execute: break knowledge down, deliver exercises just before forgetting, force retrieval rather than review, and adjust difficulty continuously.

The value of a modern training program lies not in the engine inside it, but in the quality of its scheduling algorithm and its ability to keep a learner in the zone of difficulty that is hard enough to matter. The engine is the motor. The gearbox decides whether the car climbs the hill.

And here every publisher hits the same wall: a scheduling algorithm can optimize, but it cannot manufacture motivation. It can only exploit motivation that already exists.

THE BORING PARAGRAPH NOBODY WRITES

I once sat with a fifteen-year-old student for forty minutes repeating a single endgame: the Lucena position, the rook-and-pawn ending every player must know by heart. No beautiful move. No combination. King to one square, rook to another, repeat. By the thirtieth repetition he yawned. By the fiftieth he played correctly with his eyes on the window. By the two hundredth, in a later session, he played correctly with his eyes on the board.

That is the entire story of chess improvement. Nothing romantic in the middle.

A machine helps here in a measurable way: it never gets bored, never pities you, never lets a mistake slide because you look tired. That is its real advantage and its real limit. It can hold the discipline. It cannot hold the fire. After two hundred repetitions, what keeps a human going is not an algorithm but a reason that is private, usually silly, and almost never programmable.

MY NOTEBOOK: FORTY-SIX NAMES

From March 2026 to December 2026 I kept notes on forty-six young players, aged eleven to seventeen, starting between 1400 and 2100 Elo, mostly in Vietnam and some in China. This was not research. There was no control group. But within that group, twelve passed 2200 within three years. What struck me was not which software they used. All twelve used software. The fastest improvers were not those who analyzed longest with an engine, but those who had a real person asking questions about their games every week, using the engine as a verification tool rather than a judge.

The heaviest engine users shared a habit that worried me. They opened their own game, ran the analysis, saw a red mark at move eighteen, nodded, closed the game, opened the next one. They accumulated a long list of known errors. They did not accumulate a habit of fixing them.

THE CONTRARIAN ANGLE: WHERE COLLECTIVE MEMORY IS WRONG

The story chess has told itself for thirty years goes like this: machines got stronger, humans learned more, humans got stronger, machines got stronger again, upward forever. It is a comfortable story. The data at the elite level does not clearly support it. The human rating peak has been flat for a decade. Draw rates at the top have risen. Engine-standardized opening knowledge has made the marginal advantage of one more novelty nearly worthless at the very top, while remaining enormous for learners.

And here is the under-discussed consequence: when a tool becomes universal, it confers no advantage. It only imposes cost. A fourteen-year-old in Hanoi today must absorb far more opening theory than Lê Quang Liêm's generation did when he won the World Blitz Championship in 2026. That volume does not make the boy proportionally stronger. It only shortens his chess childhood.

The second blind spot is this: adaptive systems optimize what can be measured — tactical accuracy, opening recall, endgame speed. They do not optimize courage, taste, stamina at move forty of a six-hour game, or the capacity to endure ambiguity. Metrics like move accuracy are convenient and they flatten the value of a game, because a practically correct gamble can score lower than a harmless safe move. A training culture that treats that number as a moral standard produces players afraid to err. Players afraid to err do not win major events.

MATE IN ONE, ANNOUNCED IN ADVANCE

Back to Bonn, winter 2026. Collective memory recalls a defeat: the world champion allowed a machine to mate him in one. What deserves remembering is that after the machine announced mate, Kramnik stayed seated and played on, forcing the ending to be executed on a real board. I do not romanticize a blunder. It was a blunder. But it had a quality no software possesses: it was the mistake of a man who knew he was losing and stayed anyway. My most romantic mistake was believing a chess piece knows how to write poetry. I was wrong about the piece and right about the hand that moved it.

SEVENTEEN, AND UNKNOWN TO HISTORY

In 2026, when tournaments moved onto screens, I sat in an empty hall in Shanghai with a recorder. No audience, no applause. In a club there, I watched a seventeen-year-old play games that an engine would mark in red — unclear sacrifices, moves scored low and highlighted. He won some with them. He lost more. He kept his style all season. When a seventeen-year-old prodigy does not yet know he is about to become a legend, he plays with everything the machine will later teach him to discard.

THE INTERFACE IS WHERE THE REAL MATCH IS PLAYED

If I were betting on the next decade of training software, I would not bet on the engine. I would bet on three things in the interface layer: error classification by cause rather than severity; the ability to reproduce competitive pressure; and interface design that makes studying the path of least resistance even when the user is tired and wants the answer. A license is only a piece of paper, and a dream is signed in sweat.

One aspect rarely discussed in reviews: data. A training program stores a player's entire error history, and for a young professional, weaknesses are the most valuable thing they own in the most negative sense of the word. Top national teams guard their training plans like military secrets.

READING THE THREE SLOGANS AGAIN

Trainer: the truest and hardest part. A system can organize lessons, schedule review, detect knowledge gaps. It cannot understand a person. It cannot know a student is yawning not from boredom but because his parents argued last night. Toughest opponent: technically easiest, and partly devalued, because the toughest opponent today is a free download. Strongest ally: the line I love most and doubt most. An ally is not something that always answers fast and always correctly. An ally is something that stays in the room when you have just lost an important game.

WHAT I WOULD TELL A FOURTEEN-YEAR-OLD

Can you sit with the machine at least forty minutes a day for six months, including days you do not enjoy it? Can you review your losses without opening the engine for the first ten minutes, saying aloud where you went wrong before the machine judges? Do you have someone to tell what you found, someone who asks questions instead of nodding? If all three are yes, any software will help you. If the third is no, none will.

CONCLUSION

After three decades of writing, I realize I have never written about sport. I have written about people standing before a limit, and how they choose to live with it. FRITZ 20 is a good tool, perhaps one of the best available for serious players at every level. It will help many people improve faster. But if the next generation of training software keeps perfecting itself, if it fixes every error before the learner commits it, we will eventually face a question chess has never had to answer: when there are no errors left to correct, what remains of becoming a player?