Fritz 20: when the chess engine takes on the role of teacher

Fritz 20: when the chess engine takes on the role of teacher On 27 November 2026, in Bonn, in the second game of the match between Vladimir Kramnik and Deep Fritz, the world champion played 34...Qh5. The h7 square was torn open. The machine answered with 35.Qh7#. The strongest human on the planet was mated by a move any 1,800-rated player sees in three seconds, and the chess world immediately drew its familiar conclusion: the machine has won, humans may as well sit and listen. That reading misses the most important detail. The collapse in Bonn did not measure the strength of Deep Fritz. It measured the gap between what a player knows and what he can retrieve with two minutes on the clock. That gap has never been closed, and Fritz 20 is ChessBase's latest attempt to sell players the bridge across it. Thirty-three years, from Fritz 1 to Fritz 20 Fritz is not a new name. The first version appeared in 2026 at the hands of Frans Morsch, a Dutch programmer, and was brought to market by Mathias Feist and the ChessBase team in Hamburg. In 2026 Fritz won the World Computer Chess Championship in Hong Kong, a title that at the time said more about hardware than about algorithms. In October 2026, Deep Fritz drew 4-4 with Garry Kasparov in Bahrain. Four years later, in Bonn, Deep Fritz beat Kramnik 4-2. Those thirty-three years divide into three transformations. The first was the shift from hardware to algorithm, as handcrafted evaluation gave way to deep search. The second was the AlphaZero shock of 2026 and its long aftermath: Fat Fritz appeared inside Fritz 17 in 2026, Stockfish switched to the NNUE neural network in 2026, and roughly three decades of accumulated evaluation knowledge were rewritten. The third transformation is underway now, and it has nothing to do with computing strength. It is about the machine's position inside the training room. Fritz 20 approaches users with a very different pitch: personal trainer, toughest opponent, strongest ally. This is the twentieth version of the line, exactly thirty-three years after Fritz 1. ChessBase is not selling a stronger engine. It is selling a training environment. Engine strength has become a commodity To judge Fritz 20, one thing the chess software industry rarely admits must be stated plainly: engine strength no longer carries commercial value. Stockfish is free. Leela Chess Zero is free. Every engine on the CCRL rating list sits above 3,400 Elo, while Magnus Carlsen's all-time peak is 2,882, set in May 2026. A gap of more than seven hundred points carries no competitive meaning. Nobody buys chess software today because it is stronger than the free alternative, because it cannot be meaningfully stronger. The money in this market has moved from the engine to the curriculum. And a curriculum is only expensive when it contains data about the buyer. The real value of training software lies not in how well it evaluates a position, but in how accurately it classifies your errors. That is where Fritz 20 has placed its bet, and it is a bet worth checking against data rather than faith. Openings have lost their monopoly value Another consequence of engines becoming a commodity is that the opening has steadily lost its edge. Twenty years ago, a private opening family could be worth half a point a game at club level. Today every fifteenth move sits in a shared database that anyone can consult in thirty seconds. The advantage has not vanished; it has shifted backwards, to moves twenty through forty, where no memory can carry you. That is the territory training software wants to occupy, and it is also the hardest territory to train, because it demands judgement rather than recall. The data mirror and three repeating error types Across more than two thousand games by young players that I have recorded, one pattern repeats with uncomfortable regularity. Roughly sixty per cent of losses trace back to three error types, and those three types barely change from tournament to tournament. The first is a calculation error in the transition phase, when heavy pieces have left the board but the endgame has not begun, most common between moves twenty-five and forty. The second is time management: burning too many minutes on an irrelevant branch and then losing control at the final complex move. The third is evaluation error, when a player picks the safe option in a position that is already better. What the three types share is that none of them is a knowledge error. The players I follow can point out the right move when analysing without pressure. They lose it when the clock runs. Analysis software cannot fix that error; a training system can, if it is willing to repeat the exact positions the player once spoiled. This is where Fritz 20's promise becomes readable. The application advertises personalised training, self-adjusting difficulty and a sparring partner role. The core mechanism is not new: it takes the user's own game archive, extracts positions and throws them back as repeated drills. The theory behind it is the Ebbinghaus forgetting curve, which every language-learning app has used for years. For the first time it has been seriously packaged for chess at consumer-product level. The value of that mechanism depends entirely on classification quality. A system helps only when it names your errors correctly. If it labels a game lost on move forty-three to a time scramble as an opening mistake, it manufactures an illusion of progress. Data never lies, but it enjoys testing our patience, and it tests that patience first by forcing users to audit the labels themselves. Opponent, ally, and a psychological trap The most interesting part of Fritz 20's positioning is the idea of using the machine as both opponent and teacher. Technically, that requires an engine able to play like a human at low Elo levels rather than simply playing optimally. That is a far harder problem than increasing search depth. The reason is practical. Playing a 3,600-rated engine generates no learning signal. The player loses on move eighteen without understanding why, and the feedback is a string of perfect moves beyond human calculation. That feedback is noise, not instruction. To learn, a player needs an opponent who errs the way he errs, only slightly less often. ChessBase understands this, which is why Fritz 20's marketing stresses personalisation rather than Elo. What needs verification is the basis on which the system adjusts difficulty. If it weakens itself by cutting search depth, the result is a weak machine that still plays in an inhuman style. If it weakens itself by reproducing genuine human mistakes, the result is a true sparring partner. These two paths produce entirely different products under one name. A view from the Vietnamese market Based on my experience following tournaments and youth training centres in Vietnam, spending patterns here differ noticeably from Europe. Parents will pay for in-person coaching but hesitate to pay for a software licence. Software is treated as something a free version can replace. Le Quang Liem once sat among the twenty strongest players in the world, and that achievement convinced many that the road upwards needs no paid tools. That belief is half right: at the elite level, a personal coach matters more than software. At grassroots level, where nobody has a private coach, an automated error-classification system is the only tool available. The counter-intuitive point: correlation is not causation There is an easy-sounding sales argument: players who use training software improve faster. That is true in a meaningless way. The group using training software also plays more games, employs coaches, enters more tournaments and has more free time. Isolating the software variable from that bundle is work no manufacturer wants to fund. The second doubt is more serious. Analysis tools keep improving while retrieval under pressure keeps weakening. A generation raised alongside the on-screen evaluation bar can recognise good and bad positions but cannot assess them unaided, and when the bar disappears, meaning when they sit at a real board, their support is gone. This is the tool-addiction pattern every industry with decision-support software encounters. The third doubt belongs to the ecosystem. The better a training product becomes, the more of your data it holds, and the higher the switching cost. Game archives, error histories, edited opening repertoires: all of it belongs morally to the user but usually sits in a vendor's proprietary format. A sensible player should export raw data in standard formats, regardless of whether they use Fritz 20 or anything else. I place my bets on numbers before the world learns to read them. The other half of that statement deserves equal weight: numbers are trustworthy only when you know where they were generated. A progress report produced by the software itself is not independent evidence; it is evidence supplied by the seller. What to watch in the next round Chess has passed the stage of competing purely on strength. No engine on the market produces a meaningful difference in moves, and none ever will again. The next contest happens at the level of conversion: turning knowledge into reflex, turning analysis into the right move under clock pressure. Fritz 20 has bet on exactly that layer, and commercially the bet is sound. In an empty stadium, data is the only spectator left. In an empty training room, data is the only one that tells you where you just went wrong. The question is whether Fritz 20 tells you correctly, and that only becomes visible after about three months of use, long enough for a player to accumulate a few hundred games and long enough for the system to prove it classifies errors better than the user does. If after three months the error report is still just a list of moves the computer dislikes, then what the player bought is an expensive arbiter, not a teacher.