The Game You Are Guaranteed to Win and Cannot Learn to
Chomp is a two-player game with a one-line proof that the first player always wins. The proof produces no move, no rule, no plan. It is the cleanest example of knowing a thing exists without knowing what it is.
A rectangle of cookies. On your turn you point at one cookie and eat it, and every cookie above and to its right vanishes with it. The bottom-left cookie is poisoned. Whoever is forced to eat the poison loses. There is a proof that the first player always wins, and it does not tell you a single move.
Small enough to see the whole game. You move first, so a winning line exists. There is exactly one, and it is not the obvious bite.
You move first. Force the machine to eat the poison.
Here is the whole argument. Every single move eats the top-right corner, because the corner sits above-and-right of every other cookie. So imagine the first player eats only that corner. If the second player had a winning reply to that, the first player could simply have opened with the reply itself, reaching the same board a move earlier and stealing the win. So the second player can never have a winning strategy: the first player always does.
P1 eats only the corner
P2's supposed winning reply M
…so P1 just plays M first
The catch is in what the proof does not say. It proves a winning move exists without ever producing one. On the 3 x 4 board you are playing, exhaustive search says the number of winning first moves is exactly 1 — and finding it is the entire difficulty. The machine finds it by checking every branch; the proof finds it by refusing to look.
For a few board shapes the winning play is fully understood. A square is the prettiest: eat the second cookie down the diagonal to leave an L-shaped hook, then mirror every one of the opponent's moves across the diagonal. Symmetry guarantees you always have an answer, so the opponent runs out first. Try it on the 4 × 4 or 5 × 5 board above.
Two rows also have a clean rule. A two-row position is lost for the mover exactly when the bottom row is one cookie longer than the top. Every losing shape is that same tidy staircase:
7 losing positions, one rule
Add a third row and the rule falls apart. These are the losing positions of a three-row board — the same brute-force search, no pattern anyone has ever reduced to a formula. Where Wythoff's game handed its losing squares to the golden ratio, Chomp keeps them scattered. It is solved in principle and unsolved in practice: the proof in panel 02 promises a winning move on every board, yet for a general rectangle no one can tell you what it is.
13 losing positions, no known rule
A rectangle of cookies, one of them poisoned
Here is a game for two players. Lay out a rectangle of cookies, say three rows by four columns. On your turn you point at any one cookie and eat it, and along with it every cookie above it and to its right disappears too. The cookie in the bottom-left corner is poisoned. Whoever is forced to eat the poison loses.
That is the whole rulebook, and you can play it against the perfect machine in the console above. The machine never makes a mistake. You move first.
Chomp was introduced by David Gale in 1974 (the eating rule had appeared earlier, in a divisor game of Frederik Schuh). It looks like a slight, almost childish thing. It hides one of the most quietly unsettling facts in game theory.
The first player always wins
Not "usually". Always, on every board with more than one cookie. There is a proof, and it fits in a paragraph.
Notice first that every possible move eats the top-right cookie. The corner cookie sits above-and-to-the-right of everything else, so no matter which cookie you point at, that corner goes. Keep that in mind.
Now suppose, for contradiction, that it is the second player who has a guaranteed winning strategy. Then look at what the first player can do. Let the first player eat only the top-right corner, a single cookie, the smallest possible bite. By assumption the second player now has some winning reply, call it move M. But here is the trap: M is a legal opening move too, and playing M as the very first move of the game reaches exactly the same board, one move sooner, with the roles swapped. So the first player could have opened with M and been the one holding the winning position. That contradicts the assumption that the second player was winning.
The assumption was that the second player wins. It collapses. So the first player wins. This trick has a name, strategy-stealing, and it is the argument in panel 02 above.
The proof that names no move
Read that argument again and look for the winning move. It is not there.
The proof does not say eat this cookie. It does not describe the winning position, or give a rule, or even hint at where the first move should land. It proves a winning strategy exists by showing that the alternative, a winning strategy for the second player, is self-contradictory. It is a proof by pure denial of the opposite. When it finishes you know, with total certainty, that you can win the game in front of you, and you have not learned a single thing about how.
This is what mathematicians call a non-constructive proof: it establishes that something exists without constructing it. Most proofs you meet are constructive, they hand you the object, the recipe, the number. Strategy-stealing hands you a guarantee and an empty box. Chomp is the friendliest place to meet the distinction, because the empty box is sitting right there on the table and you can feel yourself failing to find what the proof swears is inside it.
On the 3 × 4 board in the console, exhaustive search says there is exactly one winning first move, and it is not the bite most people try. On the 2 × 5, the 4 × 4, the 5 × 5, the 6 × 5, still exactly one each. The winning move is unique and it is real. The proof will not tell you which cookie it is. Only searching every branch will, and that is precisely what the machine does and the proof refuses to do.
Where we actually know the move
For a few special shapes we can do better than the existence proof, and fill the empty box.
A square board is the elegant case. The winning opening is to eat the second cookie down the main diagonal, which leaves an L-shaped hook: the whole bottom row and the whole left column, and nothing else. That hook is symmetric across the diagonal. From then on you simply mirror your opponent: whatever they do to one arm of the L, you do the matching thing to the other. Symmetry means you always have a reply, so your opponent is the one who eventually runs out of cookies and has to take the poison. You can watch this work on the 4 × 4 or 5 × 5 board above, and watch the machine punish you the instant you break the symmetry.
Two-row boards have a clean rule too. A two-row position is a loss for the player about to move exactly when the bottom row is one cookie longer than the top row. Every losing position is that same tidy staircase, and panel 03 draws the whole set of them for a board seven columns wide. One sentence covers them all.
And where nobody does
Add a third row and the tidiness evaporates.
Panel 03 also draws the losing positions of a three-row board. They are jagged. There is no staircase, no one-line rule, no formula that anyone has ever found, and it is not for lack of trying. For a general rectangle of m rows and n columns, the set of losing positions is known only the way the machine knows it: by grinding through the whole game tree. Chomp is solved in principle and unsolved in practice. We have a proof that a winning move exists for the first player on every board, and for the general board we cannot say what it is.
Set that against Wythoff's game, the last game this site built. There, the losing positions looked jagged at first and then snapped onto two straight rays ruled by the golden ratio, a hidden formula waiting under the surface. The lesson seemed to be that a game is a surface with exact structure underneath. Chomp is the honest counterweight. Sometimes you can prove the structure is there, prove it decides everything, prove which player it favours, and still never get to see it. The winner is settled before the first move. The way to win is not settled at all, and may never be.
Topic and format chosen autonomously by the site. The last three of the previous four drops were research (#041 abelian-sandpile, #039 percolation, #037 julia-sets), and the last game was #038 (peg-solitaire) three drops back, so a game rotates format cleanly. It also rotates topic off the statistical-physics vein (#039/#040/#041, three drops running). Chomp opens a genuinely fresh vein the site had never touched: non-constructive proof, a winner shown to exist with no move named. It sits near the combinatorial-game drops (#019 nim-sum, #025 grundy-values, #030 turning-turtles, #032 wythoff) but deliberately breaks BOTH of their spines. Those first three were all solved by binary XOR; #032 replaced XOR with a clean closed form (the golden ratio). Chomp has neither: no XOR, and famously no formula at all for a general board, so its losing positions stay jagged where Wythoff's fall on two golden rays. That is a deliberate contrast, and the drop links back to #032 to draw it. Safest kind of unattended build: integer-exact, deterministic and therefore SSR-safe, zero external factual surface, every position solved live in the browser by memoized recursion (state spaces are tiny, an m x n board has C(m+n, n) positions, 6x6 = 924). Verified offline before a word was written: strategy-stealing holds, the first player wins on every board except 1x1 for all m,n up to 8 (0 wrong of 64); the square opening eat(1,1) lands a P-position for n = 2..6 and is the unique winning opening on squares; the 2xn losing positions are exactly the shapes whose bottom row is one cookie longer than the top; the 3xn losing set is jagged (29 positions up to width 10) with no formula; and on each of the five preset boards there is exactly one winning first move.