Fritz 20 and the Limits of Training Chess With a Coach That Never Tires

In track and field, no coach makes a 400-metre hurdler race a motorcycle to improve results. They measure every stride, every rest interval, every second of recovery, then design a programme for each muscle group. In chess, we have done the opposite for twenty years: we put a player in front of a machine several hundred Elo stronger, call it training, and then act surprised when the next generation plays so alike that the games are hard to tell apart. Last June I sat behind a fifteen-year-old at a chess club in Chengdu. He was preparing for the national junior championship. Four hours in front of a screen, running three variations of the Sicilian Defence over and over, each more than twenty moves long, taking notes in English, colour-coding in English. On move thirty-one of the real game the next day, he made the exact mistake in the line he knew best. His clock showed twelve minutes left. He had not forgotten the variation; he had run out of time to work out where inside the variation he was standing. That is why I read the ChessBase announcement for Fritz 20 more slowly than everyone else. The oldest chess software house in Germany introduced its new product with three short sentences: your personal chess trainer, your toughest opponent, your strongest ally. The release says Fritz 20 is built both for players taking their first serious steps in chess training and for those already competing at tournament level, helping them train more efficiently, more intelligently and more individually. Those three adjectives add up to a promise far larger than a software upgrade. Fritz is not a stranger to anyone who has followed chess since the 1990s. The engine was written by Frans Morsch, published through ChessBase in Hamburg, and it once stood for an entire era in which humans still believed they could hold their own against machines. In 2026, Deep Fritz drew 4-4 with Vladimir Kramnik over eight games in Bahrain. In November 2026, X3D Fritz drew 2-2 with Garry Kasparov. In November and December 2026, in Bonn, Deep Fritz beat Kramnik 4-2 and became the first engine to defeat a reigning world champion in a classical match. After Bonn, Fritz lost the role of opponent and took on the role of tool. It was a logical change of throne. Raw calculation became a commodity: Stockfish for free, Leela Chess Zero for free, and by 2026 the NNUE neural network brought even small engines to a level no player could follow. AlphaZero appeared at the end of 2026 and taught the whole chess world a lesson about positional play. The value of a chess engine itself fell essentially to zero. What remains sellable is the interface, the curriculum, the way the software talks to its user. By 2026, market pressure pushes that promise higher. The Chess Olympiad takes place in Samarkand in September, national teams are in camp, and every training product wants to be installed on federation laptops before the planes leave. I still keep my own tracking log, and in the twelve weeks before a major team event, discussion of training tools inside coaching groups rises sharply, while discussion of conditioning and recovery barely moves. That says a great deal about how the chess world thinks about itself. The three promises of Fritz 20 deserve to be taken apart separately. "More efficiently" is the easiest promise to measure and the easiest to misread. In endurance sport, the efficiency of a session is measured by how long an athlete can hold a target intensity. In chess, efficiency has to be measured by how much is still remembered three months later, not by how many variations were viewed in one evening. Software can walk you through forty variations in two hours. It cannot guarantee that in November you will still recall ten of them. I once spent an entire winter memorising an opening system, checked myself in March and found I retained only the first part of the ideas; the rest were correct moves with no reason left behind them. Software did not cause that. Software simply did not prevent it. "More intelligently" is the interesting part. An engine answers the question of which move is best. A coach answers the question of why you failed to find it. Those two answers differ in kind. The first is a statement about the position, true for every player. The second is a statement about you, true for exactly one person. Fritz 20 promises the second. To deliver it, the software needs a model of your errors, which means it must remember where you went wrong, in what manner, and at what point in the game. This is a problem of personal data, and engine strength is only a precondition. This is where I have to open my spreadsheet. I still log thinking time for every move in the games I watch live, and what repeats across the years is a very simple curve: after the ninetieth minute of a classical game, the rate of serious errors climbs steeply, regardless of the player's level. In a small group of players I have tracked continuously, that rate rises roughly two to three times compared with the first two hours. The cause is not opening knowledge. It lies in the ability to keep decision quality intact as cognitive reserves run down. A personal training tool, if it works properly, sees that curve and places the exercise in the right spot. It will not make you study another variation at the moment you need to learn how to save time. "More individually" does not mean more exercises. It means fewer exercises, of the right kind. The clock does not create mistakes. It only strips the mask off those pretending to calculate. I still remember the game in Bonn. When Deep Fritz met Kramnik in 2026, there came a moment when the world champion was mated in a position commentators everywhere called an accident. Looked at through the time log, it was not an accident. It was the result of having to calculate precisely with only minutes left, against an opponent that never tires and never overlooks anything. The human erred at the end of a long defensive sequence. The machine did not. The gap between them sat not in knowledge but in the ability to hold accuracy through the final ten minutes. Fritz 20 promises to fill both roles: toughest opponent and strongest ally. The opponent role is far easier. You set the Elo level, pick a style, pick a degree of aggression, and the machine produces just enough pressure to force concentration. This kind of training has real value, like a pace car in a speed session: it does not teach you technique, but it stops you from lying to yourself about your own tempo. The ally role is harder. A good ally has to say the most uncomfortable thing: you should rest today, or you should not open an opening book this week, or your problem lies in the endgame and not here. Software sells you convenience. A coach sells you discomfort. Those are not the same product. I was once asked by a women's basketball team to analyse their rebounding during an Olympic cycle. I agreed on two conditions: no television, no name on the coaching staff. After six games, what I found was not a rebounding technique problem but a habit of choosing the wrong directional position for a fraction of a second, costing the team an average of four points per game. I sent back a fourteen-page breakdown naming individual players. That is the kind of coaching a machine has not yet managed: naming a habit, not naming a move. In chess, that gap shows most clearly in the transition phase. Players learn openings with software, learn endgames with software, and then walk into the middlegame on vague intuition. There is a technical reason: openings and endgames are repeatable problems, while the middlegame is where everything depends on the specific position. Software teaches well what can be repeated. Humans must learn for themselves what cannot. Everything on the board is data waiting for a reader, if the reader will sit down. I do not believe in inspiration. I only believe in the conversion rate of an advantage into points. In my own database, roughly 2,400 games from European championships that I digitised in the first half of 2026, there is one number I track steadily: the move at which a game leaves known theory. In the 1990s, games on average left that zone around move twenty-two. By the 2020s, the marker had fallen to about move fourteen. Every eight moves compressed away corresponds to thousands of hours of analysis shared with everyone at once. That convenience has a price: the window in which a player may think unaided has narrowed considerably. Adjustable Elo is an attractive feature and also a trap. An opponent playing at your level gives you balanced, comfortable games. An opponent aimed at your specific weakness gives you uncomfortable, useful games. Tournament chess does not reward balance; it rewards the ability to endure discomfort for hours on end. Fabiano Caruana became famous for an enormous volume of theory, but what carried him into title matches was not the number of variations he knew, it was his ability to hold calculation quality on move thirty-five of a six-hour game. Measuring progress is another unsolved problem. Rating measures results, not the quality of decisions. A player can gain twenty rating points through luck in a few key games while the quality of his calculation does not change at all. A serious training tool ought to measure error rate by phase of the game, not merely display a summary figure at the end. If the software only tells you whether you won or lost, it is doing the job of a results feed, not of a coach. There is a cross-discipline comparison I find useful. In athletics, the year is divided into cycles: accumulation, specialisation, competition, recovery. A top 1,500-metre runner races at most fifteen to twenty times a year, and every race sits inside a calculated plan. In chess, a professional can play nine rounds of a Swiss event in ten days, each game lasting four to five hours, then fly to another continent and do it again. No phase is called recovery. No week is designed to reduce load. And training software, in most cases, is built to fill every remaining gap with new content. That is the biggest blind spot of an entire industry of tools. Load management gets turned into a romantic story about willpower, when in reality it is often the consequence of a tournament calendar and flights that cannot be refused. Software could measure this if it wanted to. It would only need to ask how many hours the user sleeps, how many games they have played in ten days, and then say plainly: do not open another opening line. But the market does not reward advice that says no. The market rewards content. The average amateur player buys more courses than he has hours to study, and every software update is presented as a new step forward. Fritz 20 sits inside that current, yet it holds an advantage the free rivals do not: it belongs to a brand that has survived several generations of users, and that brand is tied to the memory of a time when humans still played machines. For someone taking their first serious steps, the feeling of facing a famous engine still carries appeal. For those already competing at tournament level, the value lies in the structure of the exercises and the ability to track progress over time. Both groups have real needs. What deserves scrutiny is the price of convenience. When one machine is both opponent and teacher inside the same window, the line between training and competition dissolves. No session remains without an evaluation bar running alongside. The evaluation bar is a very effective painkiller: it tells you whether the position is improving or deteriorating, and in doing so it removes the habit of self-assessment. A player raised in an environment full of evaluation bars will be good at reading machines and weak at reading people. At the highest level, reading people is the decisive skill. The counterintuitive part sits here: the better the tools, the more uniform the style. When everyone trains against the same source of truth, decisions in ambiguous positions tend to converge. Fifteen years ago, two leading players could choose two different plans in the same position and both could be right in their own way. Today, the probability that two players choose two different plans in a balanced position has fallen markedly in the events I follow. In top-level games between players of different generations, for instance between Magnus Carlsen and Alireza Firouzja, the middlegame often begins from structures that have been analysed to exhaustion. Nobody ordered them to play alike. They simply read the same book. One more point is rarely mentioned: an opponent you can switch off at any moment is never the toughest opponent. The real pressure of a chess game comes not from the opponent's move but from the fact that you cannot undo, cannot request analysis, cannot ask again. Every piece of software has a pause button. That button is what keeps it forever a training tool, however many hundred Elo stronger it is than you. Based on my experience following games over many years, I have drawn a rather dry principle: the quality of a training tool lies not in its strength but in the quality of the questions it forces the user to answer alone. A machine that supplies answers only produces users. A machine that asks the right question at the right moment produces players. Whether Fritz 20 manages that, the real answer lies elsewhere: with the person sitting in front of the screen. Software cannot decide that you should spend the next two hours playing ten blitz games to test an idea instead of viewing thirty more variations. It can only suggest. The difference between a player who improves and one who stands still for ten years usually lies not in the tools they use but in whether they are willing to refuse those tools. The training room can be the most perfect laboratory chess ever accidentally created, but only when the person entering knows what they intend to test. A machine that does not tire, does not forget and does not forgive will be there again in the morning, in the next version, stronger than today's. The only thing that does not upgrade automatically is the person sitting opposite it. The question I leave for myself after reading that announcement is simple: after a hundred hours of training with Fritz 20, does the player who walks out of the room become more like the machine, or more like themselves.