FRITZ 20: When a Chess Engine Asks to Be a Coach

Table seven, second floor of a chess club on Nevsky Prospekt, a fourteen-year-old girl named Vera stares at a screen. A number blinks in green: 0.47. She turns to her coach: "The machine says I lost almost half a pawn on move twenty-one. How much is half a pawn?" The coach stays quiet for a few seconds, then answers with a different question: "What were you thinking when you played it?" I have watched that scene more times than I can count in nearly twenty years sitting beside chessboards in Russia. Half a pawn is a unit of measurement, but for Vera it is a question about meaning. Last week, the newsroom inbox received a promotional document from ChessBase about a software package called FRITZ 20. Eight paragraphs. No publication date. No system requirements. No testing data. Not a single engine-strength figure. Only phrases: "personal chess trainer", "a training revolution", "efficient, intelligent, individual". A product from the oldest chess software publisher in the world, sold with adjectives. I read the document four times, and what made me stop was not the question of whether the product is good or bad. What made me stop was that a company with more than thirty-five years of history chose to describe its flagship brand with words that cannot be verified. ChessBase is the largest chess software publisher in the world, headquartered in Hamburg, Germany. Fritz appeared in 2026, developed by Frans Morsch and Mathias Feist. By 2026, Fritz had won the World Computer Chess Championship, and from then on it became the engine name most familiar to club players across Europe. In 2026, in Bahrain, the Deep Fritz version drew 4-4 with Vladimir Kramnik in a match the press called the Brains in Bahrain. Four years later, in November and December 2026 in Bonn, Deep Fritz beat Kramnik 4-2. Game six entered history: the world champion played a move he himself called an inexplicable mistake, and was checkmated immediately afterwards. Bonn 2026 closed a chapter. Over the following two decades, engine strength became a public utility. Stockfish, an open-source project launched in 2026, rose to become the strongest engine in the world. Leela Chess Zero appeared in 2026, learning through self-play following the AlphaZero method. Lichess offers free, unlimited analysis. Chess.com attaches analysis tools to every game played by more than one hundred million accounts. When the best analysis in the world costs nothing, selling a strong engine becomes the hardest job in the industry. ChessBase responded the way a publisher does: it acquired Komodo in 2026 according to the company's own announcement, pushed its database and magazine business, and attached the label "training" to the Fritz brand. The FRITZ 20 document follows that logic, and I read it the way I would read a financial report written in the language of poetry. A decent chess training program has to do four things. Teach openings through annotated variation trees. Drill tactics by theme, with spaced repetition to fight forgetting. Build endgame technique through exercises with answers and a rising order of difficulty. And review the student's games to identify systematic errors rather than counting isolated blunders. All four tasks have measurable indicators, and none of them appears in the FRITZ 20 document. Chess has had its yardsticks for a long time. ACPL, average centipawn loss per move, is the standard for comparing the quality of a human game against engine suggestions. Engine match rate measures how often a player follows the optimal move. Blunder frequency measures how badly thinking collapses under clock pressure. A manufacturer wanting to prove the word "revolution" only needs to publish a small study: a group of students using the software for three months, how many ACPL points they gained, what percentage their score rose in time-controlled games. The document offers no indicator at all. Not a weak indicator. Nothing. There is a larger problem inside the concept itself. According to community testing lists such as CCRL, Stockfish currently sits in the 3500 to 3600 Elo range. Deep Fritz in 2026 was rated around 2850. The historical human peak is 2882 Elo, reached by Magnus Carlsen in 2026. The gap between the strongest machine and the strongest human is now more than seven hundred points, equivalent to the distance between a grandmaster and a second-board club player. At that distance, almost every human move is an error, including the best ones. A teacher who says the student is wrong about everything will not teach anything for long. So the real engineering problem for training software is not strength but honest weakness. A pedagogically gifted engine has to play bad moves the way humans do. It must miss tactics probabilistically. Choose slightly inaccurate defensive moves at the right moment. Commit psychological errors when the endgame gets tense. This is the old heritage of Fritz, the thing that made European coaches choose it as a sparring partner in the 1990s and 2000s. It is also the hardest part of the problem, because simulating human error is far more difficult than calculating the optimal move. A computer is best at finding the perfect move and worst at pretending it cannot see it. The promotional document says FRITZ 20 is intended for "ambitious players and professionals" and also for "players taking their first steps". Between a beginner of roughly 1000 Elo and a professional of 2700 Elo lies nearly two thousand points, which is to say two different mental sports. A beginner needs someone to show them how to look at the board in order: centre first, king safety second, material third. A professional needs a clean database, search speed, and the ability to match an opponent's games within seconds. One interface cannot serve both groups at the same quality, unless the product is split into two entirely different modes. Based on my experience watching training sessions of young players in Saint Petersburg, what makes a 1500-rated student improve over six months is not a stronger engine. It is playing more games, taking fewer but more regular notes, and having someone cross-check their conclusions. Software helps only when it plays the role of that cross-checker, not when it plays the role of the judge. That line is thin, and every manufacturer has an incentive to erase it. Chess lacks an independent testing body. There is no equivalent of Euro NCAP for training software. TCEC tests competitive engines, CCRL ranks strength, but nobody tests pedagogical quality. As a result, every training claim is validated by personal impressions on forums, by sponsored videos, or by the manufacturer itself. ChessBase is both the only source and the seller. That is an unavoidable conflict of interest, and readers need to know what kind of text they are reading before they believe it. The transfer market is where people sell the future to buy hope. The chess software market works exactly the same way, with a different currency: there, people sell adjectives to buy attention. The word "revolution" is cheaper than a controlled study, and in the short term it sells more. But one thing needs to be said in the opposite direction, because analysts tend to skip it. The absence of numbers does not prove the product is weak. It proves that we are reading an advertisement, and that this industry never learned to measure itself. For thirty-five years, chess software companies have sold players a belief: better tools produce better players. Nobody forced them to prove that belief with data, because no institution ever asked. Criticising a press release for lacking figures is easy; demanding that an entire industry build testing standards is hard, and nobody has done the hard thing. The real training revolution of the past twenty years happened somewhere else entirely. It happened in the volume of games. An average club player in 2026 played a few dozen serious games a year. An average club player in 2026 plays a few dozen games a week, each with instant feedback on serious mistakes. That change came not from smarter engines but from games becoming cheap and feedback becoming immediate. The coach's role has therefore shifted from information provider to information filter. Software that understands that shift deserves the word revolution; software that merely adds an analysis feature is an update. As for targeting both beginners and professionals, I would not rush to call it an identity crisis. In platform business, today's beginner is the paying customer of ten years from now; retaining them through each rating step is sensible strategy, provided the product truly has a ladder. The problem lies in collective memory. A whole generation remembers Fritz through the Bonn 2026 match, through the moment Kramnik collapsed before a machine. Very few remember that Fritz's real value during those two decades lay in its easy interface and its adjustable levels built for amateurs. On the Luzhniki night, the old bird understood that landing is also a way of flying. An engine that abandons the strength race in order to descend to the learner's level is not a loser; it is a calculated landing. The only question is whether the company can actually do it, or is merely borrowing the image for a press release. Before writing a single line of verdict on FRITZ 20, I want to see five things: a specific release date; a verifiable feature list; system requirements; price and update model; and a before-and-after ACPL comparison covering three months of use. Without those five, every review is just a press release translated into a different voice, and the reader is left holding nothing. In 2026, when tournaments in Russia stopped and the newsroom in Saint Petersburg stood as empty as a stadium without spectators, I wrote a series about old matches to keep the column alive. When the stadium no longer holds a ball, memory begins to score. Now, in a software market full of adjectives, I feel something similar: when data is absent, legend fills the gap, and the person who pays for the filling is the buyer. Vera is still at table seven. She does not need a stronger machine to understand her twenty-first move; she needs a machine humble enough to explain why that half pawn mattered. If FRITZ 20 can do that, the word "revolution" will find its own place, with no press release to remind us. If it cannot, that half pawn will remain a blinking green number, and the chess teacher will still have to be a human being. The engine that knows how to land at the learner's level is the only one that truly teaches chess.