Learn

Six short lessons, from one stubborn equation to the vote that pays a proof.

Lesson 1

Why 114 is open

Take a whole number k and look for integers x, y and z, negative ones allowed, with x3 + y3 + z3 = k.

Some numbers can never work. Every cube leaves a remainder of 0, 1 or 8 when divided by 9, so three cubes add up to a remainder of 0, 1, 2, 3, 6, 7 or 8. A number that leaves 4 or 5, such as 4, 5, 13 or 14, is impossible. For every other k, Roger Heath-Brown conjectured in 1992 that there are infinitely many solutions.

Small cases are easy: 3 = 13 + 13 + 13 = 43 + 43 + (−5)3. Others hide among enormous numbers. In 2019 Andrew Booker found

33 = 88661289752875283 + (−8778405442862239)3 + (−2736111468807040)3

and showed that none of the other open cases below 1000 has a solution whose smallest number is at most 1016 in size. In September 2019 Booker and Andrew Sutherland found

42 = (−80538738812075974)3 + 804357581458175153 + 126021232973356313

using about 1.3 million CPU hours on Charity Engine's volunteer PCs. That run searched every open case whose smallest number is up to 1017 in size and also turned up 165 and 906; a separate run found a third way to write 3, and a slower run that started in 2020 found 579.

Seven numbers below 1000 are still open: 114, 390, 627, 633, 732, 921 and 975. 114 is the smallest. Ponmath's hunt starts at |z| = , past every published search. Nobody knows how large the answer is, so the hunt is a lottery, not a countdown.

Check a triple

Enter x, y and z; the page computes x3 + y3 + z3 exactly, in this browser.

Lesson 2

How the swarm searches

A sum of two cubes factors: x3 + y3 = (x + y)(x2 − xy + y2). Call d = x + y. Then

k − z3 = d · (x2 − xy + y2),

so d must divide k − z3. Written another way: z3 ≡ k (mod d). Three unknowns become two, z and d, and d stays small when z is the smallest of the three: |d| is at most (∛2 − 1)·|z|, about a quarter of |z|. Since x3 + y3 = k − z3 is close to −z3, d and z have opposite signs.

Your browser repeats four steps:

  1. Pick a value of d and factor it.
  2. Find every cube root of ±k modulo d. Each root gives a run of z values, one every d steps.
  3. Walk each run through the band the hunt is open for, |z| from to .
  4. For each z, compute q = (k − z3)/d and Δ = (4q − d2)/3. If Δ = r2 for a whole number r with d − r even, then x = (d + r)/2 and y = (d − r)/2, and the equation is solved.

Almost every Δ fails fast. A perfect square leaves only certain remainders when divided by 64, 7, 11, 13, 17, 19, 23, 29 and 31, and a random number passes all nine tests about once in 820 tries. The survivors get the full square root test, and some of them become shares (lesson 3).

Ponmath does not sweep the band in order the way the record searches did. Each browser draws its own values of d at random and walks them. The chain counts each candidate once, so no work is paid twice, and the odds are long but never zero.

Lesson 3

How the chain checks a share

A share is a candidate (k, z, d) that passed the same square filters a record search uses, with its square root worked out, and whose hash starts with enough zero bits. When you submit, the hunt contract redoes everything from scratch:

  1. k is an open target and the hunt is not paused.
  2. |z| is inside the band, d is not zero, d and z have opposite signs, and 3·|d| ≤ |z|.
  3. d divides k − z3. The contract computes q = (k − z3)/d and N = 4q − d2, checks that 3 divides N, and sets Δ = N/3, which must not be negative.
  4. Δ passes the nine square filters, modulo 64, 7, 11, 13, 17, 19, 23, 29 and 31.
  5. You send r, and the contract checks r2 ≤ Δ < (r + 1)2, so r really is the whole square root.
  6. It hashes the candidate with your wallet address, the season and its own address. The share counts only if the hash starts with zero bits, about one survivor in 2.
  7. The same (k, z, d) counts once, ever.

If any rule fails, the whole batch is rejected, so the page checks every share and simulates the batch before it asks you to sign. Each accepted share mints points to you.

Why the hash? It ties the work to your address, so a share copied from someone else fails for them. And the zero bits let the chain sample your work: on average, one share stands for 2 survivors your browser worked through, yet each share is small enough to check in a single transaction.

If r2 = Δ exactly and d − r is even, the share is more than a share. It is a solution, and the contract pays the jackpot (lesson 5).

Replay the 42 solution

The 2019 triple for 42, taken through the share rules one step at a time, computed in this browser.

StepValueCheck

Lesson 4

What a Lean proof is

Lean is a programming language and a proof assistant. A statement is written as a type, and a proof is a program of that type. Lean's small kernel checks every step, so a finished proof leaves nothing to trust but the definitions.

Google DeepMind's formal-conjectures project writes famous open problems in Lean, with the proof left as sorry, Lean's placeholder for "not done yet". Erdős problem 1135, the Collatz conjecture, points to this statement:

theorem collatz_conjecture (n : ℕ) (hn : n > 0) : ∃ m, collatzStep^[m] n = 1

Many statements begin with answer(sorry) ↔, because Erdős asked a yes or no question. A solver replaces answer(sorry) with True or False and proves the result, and either answer settles the problem.

Ponmath pins each Erdős vault to one commit of formal-conjectures, one file and one theorem name. A valid claim is a Lean project that proves exactly that theorem at that commit, with no sorry left and no new axioms. Attestors build it on their own machines and list its axioms with #print axioms; the standard three, propext, Classical.choice and Quot.sound, are fine. Lean never runs on chain: the chain records the attestations and the DAO vote.

This is no longer science fiction. In January 2026 AI systems closed Erdős problem 728: GPT-5.2 Pro and Harmonic's Aristotle produced a Lean proof, written up in arXiv 2601.07421.

Lesson 5

How vaults pay

A vault is a pot of gifts with one rule for who gets it.

Hunt vaults. There is one for each open number, and it pays only when the hunt contract verifies a solution. The contract computes x3 + y3 + z3 itself and compares it with k. The first valid solution closes the target, stores the triple and the finder on chain for good, mints discovery points, and sends the whole vault to the finder. No person can release it. The one other way out is a DAO vote that moves it to another vault, and such a vote is public for days before it runs.

So that nobody can copy your numbers out of a pending transaction, a solution is claimed in two steps. First you commit keccak256(k, x, y, z, your address, a secret), which reveals nothing. At least 60 seconds later, and within 30 days, you reveal the numbers, and the chain checks both the commitment and the cubes.

Erdős vaults. These pay for a Lean proof (lesson 4). You commit a hash of your proof hash, your address and a secret. At least 10 minutes later you submit the proof hash and a public link, which opens a 14-day review. Attestors build the proof and attest pass or fail, once each. Once the review is over and at least attestor has passed it, a DAO vote can release every asset in the vault to you. Only the DAO's timelock can release or reject a claim.

Giving. Anyone can give ETH, USDG, WETH or approved Robinhood stock tokens. Gifts cannot be withdrawn. They earn patron points per dollar at the Chainlink price; when no fresh price exists, the gift is still accepted and earns 0 points. If a problem is solved elsewhere, or its Lean statement turns out to be wrong, the DAO can move the vault.

Lesson 6

How the DAO works

Ponmath is run by its points. Points are minted for accepted shares, for gifts to vaults, for a discovery and by DAO grants. They cannot be sent, sold or approved to anyone: the contract refuses every transfer. Your points are your votes, and they count from the moment they are minted, with no extra step.

A proposal goes through five steps:

  1. Anyone with at least points can propose. A proposal is a list of contract calls and a description.
  2. Voting opens later and lasts . You vote For, Against or Abstain with the points you held when voting opened.
  3. It passes if For beats Against and the For and Abstain votes together reach of all points.
  4. A passed proposal waits in the timelock, so anyone who disagrees sees it coming.
  5. Then anyone can execute it.

Proposals can list a vault, add an open number or change its band, start a season, release or reject an Erdős claim, appoint attestors, grant points, move a vault, put a vault's USDG into an approved yield vault and take it back, and remove the guardian. A proposal can also be a plain text signal.

At launch a guardian, the deployer, sets up targets, vaults and attestors so the hunt can start. The guardian cannot move prize money: only a verified solution, or a DAO vote through the timelock, can do that. During the beta the guardian can also cancel a passed proposal while it waits in the timelock, a veto that covers a vote to remove the guardian too. It cannot propose or execute anything. The DAO can do everything the guardian can, and a DAO vote can remove the guardian and end the veto.

Your turn.

Open the hunt, pick a number, and let your browser search. The chain checks every share.

Start hunting