FRITZ 20: When a Chess Engine Learns to Be a Personal Coach

An Afternoon in Malleswaraam

The club sits on the second floor of an old row house, its chequered floor tiles bleached pale by the sun, the ceiling fan turning slowly enough that you can hear the steel blades scrape the axle. Every afternoon, about twelve children between the ages of eight and fifteen gather around four wooden tables. On each table lies a plastic chess set with worn edges; every King carries a scratch on its crown from being tapped down too many times by small fingers during blitz games.

The man who teaches them is Ravi, forty-seven, a former Karnataka state champion. He now runs the class to pay his rent and his daughter's school fees. On his desk sits a desktop computer bought in 2026. The cooling fan hums like a bee, and somewhere on the hard drive is an old version of Fritz, installed back when chess software still shipped on a CD with a two-hundred-page manual.

The afternoon I dropped by, a thirteen-year-old girl named Ananya was sitting in front of the monitor. She had just lost to the machine on its weakest setting, and instead of switching it off, she pressed the analysis button. The engine returned seventeen moves, each with a number attached. Ananya read every line, slowly, her lips moving as if spelling the words out. Then she turned to Ravi and asked: "Sir, if the machine says this move is bad, why do I still think it's beautiful?"

Ravi did not answer. He poured more tea, pushed the cup toward his student, and told her to replay the game from the start, this time without opening the computer.

That scene is what came to mind when I read the announcement for FRITZ 20, the latest edition of the chess software line published by ChessBase of Hamburg. The publisher introduces it with three short sentences: your personal chess trainer, your toughest opponent, your strongest ally. For someone who has sat in rooms like Ravi's for years, those three sentences are both right and slightly off. Right, because the machine has genuinely changed the way children learn chess. Off, because it has not changed the hardest part of learning chess.

From Wooden Tables to Digital Boards

Fritz is not a new name. The first version appeared in 2026, written by the Dutch programmer Frans Morsch. Four years later, Fritz 3 won the World Computer Chess Championship in Hong Kong, and from that moment the name entered the memory of at least two generations of players. By the mid-1990s, ChessBase had folded Fritz into its software ecosystem, pairing the engine with what is generally regarded as the largest game database in the world.

In 2026, in Bahrain, Garry Kasparov faced Deep Fritz over eight games and left the board with a 4-4 draw. Four years later, in Bonn, Deep Fritz beat Vladimir Kramnik 4-2, and the second game of that match entered history as one of the most famous blunders by a reigning world champion: Kramnik allowed the machine to deliver mate in a single move. After that, the question of whether humans or machines played better simply ended. The machine played better, and by a very wide margin.

But something else changed too, more quietly. Once the strength race was over, the value of chess software shifted from beating the user to teaching the user. In 2026, DeepMind's AlphaZero showed that a neural network could learn chess from nothing and play at a superhuman level. Three years later, the NNUE technique was integrated into Stockfish, turning an open-source engine into a tool that surpassed every commercial product of only a few years earlier. The market's problem was no longer who was strongest. Stockfish is free, Leela Chess Zero is free, and both are stronger than anyone genuinely needs.

So when ChessBase talks about FRITZ 20, it does not talk about strength. It talks about training. That is a notable pivot, and it reflects the market accurately: people who buy chess software today are not paying for a powerful engine, they are paying for an engine that knows how to talk to them.

The regional context is worth placing alongside this. India, where I live and work, is going through an unprecedented chess boom. The Chennai Chess Olympiad in 2026 marked the first time a major team championship was staged in South Asia. In December 2026, D. Gukesh became the youngest world champion in history by defeating Ding Liren in Singapore at the age of eighteen. R. Praggnanandhaa, Arjun Erigaisi and Nihal Sarin have all entered the world's elite group. Behind every one of those names is a small club, a coach, a computer, and an engine.

That is why the message behind FRITZ 20 is not unfamiliar to me. It is also why it deserves a careful reading.

Three Promises

The first promise is the personal trainer. The difference between a chess workbook and training software is that a workbook does not know who you are. A workbook contains a thousand puzzles, and everyone who buys it solves the same thousand puzzles. Personalised training software does the opposite: it takes your own games, sorts your errors into categories, and generates exercises from the mistakes only you make.

Technically, this is not magic. You import your games into the program, from the ChessBase database, from online platforms via the PGN format, or from handwritten scoresheets at tournaments. The software runs a deep analysis, tags each move, and then uses classification algorithms to answer one question: what kind of error does this player commit? A pure tactical oversight? A positional misjudgement? A collapse under time pressure? A deviation from prepared opening theory?

The answer determines the entire training programme. A player who blunders tactically needs fast-paced tactical drills. A player who plays the first eighteen moves well and then collapses on move thirty needs endgame work and resilience work. A player who keeps landing in difficult positions out of the opening needs a redesigned repertoire tree, not another thousand tactics puzzles.

The subtle point is this: personalisation is not about making learning easier, it is about making learning more accurate. A player can spend three years solving puzzles about good structures while their real problem lies in converting a small advantage in the endgame. Software, if designed properly, will detect that within weeks.

The second promise is the toughest opponent. This is the most technically interesting part and also the most misunderstood. A modern engine is so strong that it cannot play weakly in any natural sense. To bring its strength down to club level, you have to interfere with how it calculates: limit search depth, limit time, inject noise into the evaluation function, or deliberately skip certain good moves. Each approach carries a cost. Limit the depth and the engine will make bizarre errors no human would make, errors of a mechanical kind. Inject noise and the engine will miss moves that a player of that level would never miss.

Building a training opponent that feels human is far harder than building the strongest machine in the world. This is exactly where adaptive sparring modes and handicap modes become useful. Handicapping, giving material away before the game begins, is the classical training method of nineteenth-century chess, when masters toured Europe and played local opponents at odds. It forces the weaker player to learn from a bad position and the stronger player to learn absolute precision.

What software adds is real-time difficulty adjustment. A good sparring mode tracks your win rate and adjusts continuously, keeping you in the zone where you win roughly half your games. That is the most effective learning zone in any sport, and it is nearly impossible to create by playing humans, because a human opponent has no obligation to stay in that zone with you.

The third promise is the strongest ally. This is where ChessBase holds its biggest advantage and where it is hardest to copy. A strong engine can be written by anyone. A database of millions of games, a cloud analysis system, a file format the entire industry accepts, and an enormous library of training material take thirty years to build.

For ordinary users, the value lies in having everything in one program. You open the software, play a sparring game, review it with automatic annotation, look up a similar structure in the database, find the classic games that shaped that structure, and save it all into a training profile. That loop is closed, and the closure is the product.

So what is new in FRITZ 20? The publisher does not offer shocking Elo numbers, and that is sensible. For an end user, a few dozen Elo between top engines is meaningless. The novelty lies in the training layer: how the software classifies errors, how it builds multi-week plans, how it presents a game as a lesson rather than a report. It is a shift from data reporting to learning experience.

And Here Is Where I Doubt

The first thing that stopped me was the phrase "toughest opponent". A machine cannot be your toughest opponent, simply because it is not afraid. Fear cannot be programmed. The toughest opponent I have ever seen is a twelve-year-old boy rated thirty points below you, sitting as still as a statue, playing as slowly as if time did not exist, not blinking once while you attack. The machine can beat that boy a thousand games in a row, but it cannot teach you how to endure him.

What the machine does, and does very well, is remove the cost of losing. You can lose to it a hundred times and nobody knows, nobody records it, nobody remembers. That is why children have learned chess so quickly over the last fifteen years. It is also why some of them never learn to handle real defeat.

The second issue is the personalisation loop. If the software builds your training programme from your own games, it is teaching you to become a better version of yourself. That has value, but it is a closed loop. The biggest leaps in a young player's development usually come from being forced to play what they hate: a closed positional game, a grinding endgame, a system they do not understand. An engine optimised for comfort will not voluntarily push you out of your familiar zone. Only a coach does that.

The third issue, and perhaps the most important to me, is that the real bottleneck is not engine strength. In the rooms I have sat in, from Bengaluru, Kochi and Guwahati to Chinese cities before that, the limiting factor is rarely the machine. It is power cuts at six in the evening, a connection that dies mid-tournament, one laptop shared by three siblings, a coach teaching thirty children at once to make rent, and language. A training revolution that assumes a decent computer, a stable connection and English literacy is a revolution for the top five percent.

The fourth issue is stylistic homogenisation. Since engines became the universal teacher, elite chess has converged: the same structures, the same prophylactic moves, the same conversion technique. Draw rates in major events remain anchored at a high level, and much of the decisive content now comes from preparation depth and from late-tournament fatigue. Something has been lost. The idiosyncratic style of players like Mikhail Tal or Bent Larsen is harder to sustain when a machine whispers that the move is minus zero point four.

The fifth issue is integrity. Every powerful training tool is simultaneously a powerful cheating tool. The engine in your pocket while you study is the engine in your pocket at a tournament. The 2026 scandal involving Hans Niemann and Magnus Carlsen forced the whole sport to rethink anti-cheating, and software companies have had to accept that the power of their product is double-edged.

What Remains

Somewhere in a corner of this city, in a room with a slow ceiling fan and sun-bleached floor tiles, a thirteen-year-old girl has just asked her coach a question no software can answer: why a move the machine grades as bad still looks beautiful to her. The hometown pitch has no grandstand, but every time the ball rolls, a whole sky of memory comes rushing back. The hometown board is the same, even though no camera has ever visited it.

FRITZ 20 will be stronger than the version before it. That is certain. But the interesting question is not how many Elo it adds. It is whether the children in rooms like Ravi's will get to use it in a way that keeps them curious, or merely teaches them to memorise numbers. The true measure of a training tool is not the rating points it produces, but whether the person using it learns to ask better questions.