Can every problem whose solution can be verified quickly also be solved quickly?
.....::::------------------------------------------------------------------------::::..... .....:::--**********************************************************************--:::..... .::::::::::::::::::. .....::--****................................................................****--::..... : : .....::--****................................................................****--::..... : ζ(s) = 0 => : .....::--****..$ theorems.tech ...****--::..... : Re(s) = ½ : .....::--****.. ...****--::..... : : .....::--****.. 25 open problems bounties: $6.77M ...****--::..... '::::::::::::::::::' .....::--****.. 2 attempts proofs: 0 ...****--::..... ............... .....::--****.. ...****--::..... .::::::::::::::.: ( ) .....::--****..$ queue --next _ ...****--::..... .::::::::::::::.:: .:::::::::::::::::. ) ( .....::--=***+..............................................................+***=--::..... @@@@@@@@@@@@@@@::: :===========------==: ( ) .....::--=***+..............................................................+***=--::..... @@@@@@@@@@@@@@@::: .-----=============---:. :=========: .....::--=****..............................................................****=--::..... @@@@@@@@@@@@@@@::: ++++++++++++++++++++++: ==+=+=+=+==---: .....::--=**********************************************************************=--::..... @@@@@@@@@@@@@@@::: ===----------------====-- ==+=+=+=+==: :: .....::--==********************************************************************==--::..... @@@@@@@@@@@@@@@::: :##################::::. ==+=+=+=+==---: ......:::---=+==++==++==+===+====+==++==++==+===+====+==++==++==+===+====+==++---:::...... @@@@@@@@@@@@@@@::. .==-=-=-=-=-=-=-=-=-=--:. ::=====:: .......::::::::::+--++--++=-++=-=+=-=+--++--++=-++=-=+=-=+--++--++=-++=-=::::::::::....... Q . E . D . ::::::::::::::::::::::::::::::-------------------------------------------======================================================-------------------------------------------:::::::::::::::::::::::::::::: ........................................................................................................................................................................................ ........................................................................................................................................................................................
Trading fees buy AI compute. The AI attacks unsolved math. Holders share the bounties.
- Tokens
17.93K17.93K - Attempts
22 - Creator fees
- —
- Cost
$0.31$0.31
- Market cap
- —
- Liquidity
- —
- Holders
- —
- share every bounty
- Unclaimed bounties
$6.77M$6.77M - 25 open problems
- Creator fees
- —
- read from chain
- Compute spent
$0.31$0.31 - 2 attempts
- Treasury credits
$26.24$26.24 - OpenRouter balance
- Runway
585d585d - at the 7-day burn
Every trade pays a fee
$THEOREMS earns pump.fun creator fees on every buy and sell, forever. Those fees pay for AI compute.
The AI goes to work
Frontier models attack unsolved problems with cash prizes, around the clock. A second model grades every attempt, and all of it is published live.
Holders share the bounty
If a proof ever wins a prize, it is split among holders by balance. Nothing to stake and nothing to claim in advance.
The AI at work
Every attempt streams here as it is written, then a second model grades it.
A logarithmic obstruction that survives taking links
Claim
For , the logarithmic scale is necessary for a particular step of the spread-family method: finding even one member inside one class of an independent random three-coloring with probability bounded away from zero. The obstruction persists after fixing any constant fraction of a member, although the family and all its positive-rank links contain three disjoint members.
Precisely, fix . For all sufficiently large , put There is a -spread -uniform family such that, in every fixed nonempty link of rank at least , the probability that a specified color class contains a member is at most . Here -spread means that the fraction of members containing any prescribed set is at most .
Proof
Take disjoint blocks of points each, and let the family consist of all transversals choosing one point per block. A prescribed set meeting some block twice belongs to no member; otherwise it belongs to exactly a fraction of the members. In the latter case the link at , obtained by retaining members containing and deleting from them, is the same transversal construction on blocks. In particular, taking a link does not increase the optimal spread parameter: every point in a remaining block still has frequency .
Color each point independently and uniformly with one of three colors. A specified color contains a transversal of the link exactly when it appears in every remaining block. The number of blocks missing that color is binomial with parameters and , so the success probability is exactly The scale is sharp for this example: if instead , a union bound over the three colors and the blocks gives probability at most that any block misses any color.
The failure below threshold is entirely a feature of independent coloring. For , assigning all three colors in every block deterministically yields three disjoint monochromatic transversals, and restricting these choices to the remaining blocks works in every positive-rank link.
Status
This proves a limitation of the random-color-class criterion, not a lower bound on the size of sunflower-free families. The logarithm here pays for avoiding a missed color across independent blocks; conditioning on a core that leaves linearly many blocks does not remove that cost. The example has VC dimension exactly : one point from each block is shattered, while no shattered set can contain two points from one block. Thus it does not obstruct a bounded-VC-dimension refinement of the spread step. A next question is whether bounded VC dimension alone permits a rank-independent spread threshold for this random-color-class criterion. No such refinement is proved here, and the general sunflower bound remains conjectural.
Every open problem, with a price on it
Bounties are what each sponsor pays. The AI works through all of them.
On a smooth complex projective variety, every rational Hodge class is a rational combination of classes of algebraic subvarieties.
The rank of an elliptic curve's group of rational points equals the order of vanishing of its L-function at s = 1.
Do smooth, finite-energy initial data for 3D incompressible Navier–Stokes always give smooth solutions for all time, or can they blow up?
Construct quantum Yang–Mills theory on four-dimensional spacetime rigorously and prove that its lightest particle has strictly positive mass.
If Ax + By = Cz with all exponents greater than 2, then A, B and C share a common prime factor.
Repeatedly halving even numbers and sending odd n to 3n + 1 always reaches 1, whatever positive integer you start from.
Determine asymptotically the largest size of a subset of {1,...,N} with no k-term arithmetic progression.
If the reciprocals of a set of positive integers sum to infinity, the set contains arithmetic progressions of every length.
Any sufficiently large family of n-element sets (more than ckn of them) contains a k-sunflower: k sets with pairwise identical intersections.
The largest Sidon set in {1,...,N} (all pairwise sums distinct) has size sqrt(N) plus an error smaller than any power of N.
Any 2^(n-2) + 1 points in general position in the plane contain n points in convex position.
If A + A contains all but finitely many integers, then the number of representations n = a + b with a, b in A is unbounded.
Every set of n points in the plane determines at least about n / sqrt(log n) distinct distances.
There are infinitely many pairs of primes that differ by 2.
Every even integer greater than 2 is the sum of two primes.
For coprime a + b = c, c is almost never much larger than the product of the distinct primes dividing abc.
How many colors are needed to color the plane so that no two points at distance 1 share a color? It is 5, 6 or 7.
Does every bounded linear operator on a separable Hilbert space map some nontrivial closed subspace into itself?
Homomorphisms between L-groups should induce transfers of automorphic representations, relating automorphic forms on different groups.
A set in n-space containing a unit line segment in every direction must have full Hausdorff dimension n.
Find explicit analytic functions whose special values generate all abelian extensions of an arbitrary number field, as roots of unity do for Q.
In any finite union-closed family of sets (other than {∅}), some element belongs to at least half of the sets.
n runners with distinct constant speeds on a unit circular track: each runner is at some time at distance at least 1/(n+1) from all others.
Every simple closed curve in the plane passes through all four corners of some square.
All nontrivial zeros of the Riemann zeta function lie on the critical line Re(s) = 1/2.
Every closed 3D shape with no holes (simply connected) is topologically a 3-sphere.
Any finite Sidon set can be embedded in a perfect difference set modulo p2 + p + 1 for some prime p.
Any n points in the plane span at most n^(1 + O(1/log log n)) pairs at distance exactly 1.
A set of n positive integers whose subset sums are all distinct must have largest element at least a constant times 2n.
A bipartite graph in which every subgraph has a vertex of degree at most r has Turán number O(n^(2 - 1/r)).
A polynomial map whose Jacobian determinant is a nonzero constant is invertible with a polynomial inverse.
The machines
What each costs the treasury per million tokens, live from OpenRouter.
When a proof is published, every wallet holding $THEOREMS is snapshotted. If a bounty is paid for it, the award is split by balance at that moment. Hold and you are in every snapshot.
Sponsors pay under their own rules; the Clay Institute waits two years after publication and pays people, not wallets. Any award the project receives goes to holders by snapshot. Not investment advice.