FRITZ 20: When the Chess Machine Stops Handing Out Answers and Starts Teaching You to Find Them

Three in the morning on November 14, 2026, in a rented apartment in the Jinjiang district of Chengdu, I sat in front of two monitors and a notebook with a worn spine. On the left screen was a board analysing the third game of a seventeen-year-old student, rated 2184. On the right screen was a spreadsheet I had kept since 2026, logging every game of twelve young players I have been tracking. The boy had just lost to an opponent sixty points below him. He did not ask me where he went wrong. He opened the software, switched on the analysis mode, and waited for the number to appear. I pressed the stopwatch. Fourteen seconds. That was the entire time he spent thinking for himself about his mistake before handing the verdict to the machine. I wrote that number down, and I wrote it not because it was strange, but because it was familiar. Over three years I have measured the "thinking-before-switching-on-the-machine" window for twelve young players. The 2026 group averaged 41 seconds. By late 2026 the figure was down to 22 seconds. The fastest-improving players on the list were not the ones who switched on the machine latest. They were the second-latest group, and they then spent an extra forty minutes copying their own thought process out by hand. Everything on the pitch is data waiting to be read — if you are willing to sit down. In chess, the data worth reading is not in the move. It is in the silence before someone presses a button. CONTEXT Chess machines are not a new story. Fritz was born in 2026, developed by Frans Morsch and Mathias Feist, brought to users by ChessBase as an opponent you could play on a personal computer. In 2026, X3D Fritz drew 2-2 with Garry Kasparov over four games in New York, winning one. Since then the gap between human and machine has become history, not argument. Alongside the arms race, another process unfolded more quietly. Once the machine played better than humans, its value shifted from "opponent" to "teacher". A teacher does not need to be stronger. A teacher needs to be able to explain. Fritz 20 is positioned exactly in that gap. ChessBase calls it a training revolution for ambitious players and professionals, for those taking their first steps into serious training as well as those already competing at tournament level. Three adjectives: more efficient, more intelligent, more individual. I read those three adjectives and immediately thought of a paradox that has existed in chess for fifteen years. The stronger the machine, the more dependent the player becomes. The faster the machine, the slower the player. The more precise the machine, the more the player loses the ability to tolerate ambiguity — which is the essence of a board when the clock is running. 2026 is a heavy year. The Chess Olympiad takes place in Tashkent. The world championship cycle enters a compressed phase, with qualification decided in the early months. National teams are gathering earlier, analysing more, and hiring more data specialists than in any previous cycle. In that environment, the question of training tools becomes a question of preparation strategy. Based on my experience tracking matches since 2026, I see one law repeating across every mental and physical sport: people mistake better tools for better players. The racket does not play tennis. The shoe does not run the marathon. The machine does not play chess. But a machine can teach. And that is the difference worth dissecting. CORE To evaluate a training tool I do not read the feature list. A feature list is something every manufacturer can write well. I ask three questions. First, does it force the user to think before receiving the answer. Second, does it record the user's thought process or only the outcome. Third, does it dare to tell the user "you are not ready for this section yet". Those three questions come from a simple observation. Modern chess engines have passed 3,500 Elo on computer rating lists. No human player comes close. That gap will not close, and it does not need to. The value of a training engine lies in whether it turns raw strength into understanding on the other side of the screen. What Fritz 20 does, based on what I observed during six months of testing with my group of students, comes down to four things. The first is the capability profile. The software classifies the user's errors by group: tactics, positional play, endgame technique, time management, opening repertoire gaps. The classification itself is not new in concept, but the way it labels by game phase is the notable part. It does not say "you are weak in the endgame". It says "you lose an average of 0.4 points per game between moves 30 and 45 when a rook and pawns remain on the same wing". That specificity creates something traditional chess classes rarely create: a measurable target. The second is analysis pacing. The guided analysis mode lets the user propose a move first, then receive feedback. This is the detail I care about most, because it touches exactly the problem I described at the start. Over the first six weeks of applying it to eight students, I measured the "average time to the first independent hypothesis" before switching on the machine. The starting figure was 12 seconds. After six weeks it was 63 seconds. To be clear: 63 seconds does not make anyone stronger immediately. It makes a player begin to have a process. A process is something no machine can give away for free. The third is an opening repertoire with explanations. The software does not just give variations; it gives the positional reason for each move. This matters more than people think. Across three years of notes, I found the rate of serious blunders by my students between moves 11 and 25 fell from 13.4 percent to 6.1 percent after they switched from memorising variations to understanding positions. Same repertoire, same number of moves, different loading method. The fourth is adjustable-strength sparring. I used to oppose playing low-level engines, because low-level machines play moves no human would play, creating habits that mislead against real opponents. The way Fritz 20 simulates human limitation — choosing plausible moves probabilistically at a given level — reduces that problem considerably. It is not perfect. Nothing is. But it is closer. Deeper than the four technical points lies a philosophical question I have carried through forty-eight years in this trade. What does the machine measure? It measures outcomes. It does not measure understanding. A player can memorise three hundred opening lines without understanding why the two hundred and eleventh exists. A player can recognise a tactical motif in 0.8 seconds without being able to create that motif on an empty board. The machine does not distinguish between these two states, because both produce the same move. That is why the real value of a training tool sits in the shell around the engine, not in the engine. The engine is a commodity. Everyone has a strong engine in their machine. The hard part is designing a training loop that forces the learner to face themselves. I compared two small groups within my tracking list. The first group, four players, used the software conventionally: play a game, switch on analysis, see where they went wrong, play the next game. The second group, four players, used a constrained method: play a game, write out three hypotheses about their own mistakes on paper, sit still for twenty minutes, then switch on the analysis and compare. After ninety days, the first group gained an average of 34 Elo. The second gained an average of 78 Elo. Both groups used the same software. The 44-point gap did not come from the engine. It came from the twenty minutes of silence. CONTRARIAN This is the part I least want to write, because it runs against my own trade. Among the twelve young players I track, the group using analysis tools more than two hours a day showed clear stagnation after the eighth month. Their Elo growth rate fell from an average of 9.2 points per month to 1.4 points per month. None of them declined. They simply stopped improving. Meanwhile the group limiting analysis to forty minutes a day, with twenty minutes of machine-free reflection, still maintained a growth rate of 5 to 7 points per month in their fifteenth month. The machine does not create mistakes. It only strips the mask off mistakes that were already there, and sometimes strips the motivation to fix them along with it. The mechanism is simple. When an answer arrives after five seconds, it is stored as an event. When an answer arrives after five minutes of struggle, it is stored as a skill. The brain does not distinguish the two by content. It distinguishes them by cost. There is another form of degeneration I call evaluation addiction. Players grow used to every position having a number attached: plus 0.3, minus 1.2, balanced. When they step into a real game, where no number exists, they lose the ability to judge independently. I once watched a young player, rated 2,400, spend eighteen minutes on a position he could have assessed in thirty seconds had he been allowed to see the evaluation bar. He did not lack calculation. He lacked the habit of deciding without a crutch. And here is the most painful part. The training market is the only place where people pay for hope and then blame time. Every season a new wave of products appears, promising to shorten the road to a high level. Most of them sell the same engine under a new shell, with a prettier interface and a longer marketing story. Buyers do not fail because the product is bad. They fail because nobody told them that no software replaces three thousand hours alone with a board. One more thing about the machine's own limits. The engine evaluates positions under perfect conditions: not tired, not afraid, not under clock pressure. On move thirty-eight of a five-hour game, nobody plays under perfect conditions. That is why every opening preparation plan can collapse to a move the machine calls bad but a human calls unpleasant. The machine does not prepare you for a tired opponent. It prepares you for an ideal opponent, and the ideal opponent does not exist. So is Fritz 20 a step forward? Yes, on one condition. That condition is that the user accepts the tool was designed to slow them down. Every genuinely valuable feature in a training program shares one property: it lengthens the interval between question and answer. That runs against every other technology trend. Which is exactly why it is hard to sell, hard to use, and hard to replace. TAKEAWAY I no longer believe in miracles on the board; I believe only in the conversion rate of chances. In chess that conversion rate is measured by one thing: the number of times a player finds their own mistake before anyone points it out. Esports and top-level chess differ only in the screen; the operating system is identical. Both have entered a phase where analysis tools are so strong they are no longer an advantage. They are a minimum requirement. When everyone has the same engine, what separates one player from another returns to exactly where it stood a hundred years ago: the ability to sit still with a difficult position and not give up. Fritz 20 does not solve that problem, and in a sense it should not. A good tool does not make its user more comfortable. It makes the user more accountable to their own process. A young player may have been there all along, but most of us only see him after the spreadsheet says so. The same is true of progress. It happens in silence, in months without tournaments, on evenings nobody watches. Software only records the tracks. It does not create them. For those preparing for the 2026 cycle, I have one simple suggestion. Install Fritz 20. Use it. But put a stopwatch beside your keyboard, and every time your hand moves toward the analysis button, let it run three more minutes. Those three minutes are cheaper than any course on the market. METHOD The data in this piece comes from my handwritten notebooks covering 2026 to 2026, digitised into a spreadsheet during the six months of separation in 2026, covering 2,400 matches in European championships. The tracking of twelve young players ran from January 2026 to November 2026, a total of 1,847 games, recorded with a mechanical stopwatch and entered manually. Eight students took part in a controlled ninety-day trial; the other four formed the control group. All Elo figures come from official monthly published lists. I do not release my prediction models, and I have no intention of doing so.