May 17, 2026
Float drama, integer chaos
Schanuel's Conjecture and the Semantics of Triton's FPSan
A math-powered bug checker just dropped, and the comments are split between awe and “but does it prove anything?”
TLDR: A new Triton tool tries to check whether rewritten number-heavy programs are truly equivalent by converting them into strange integer versions, with its logic tied to a huge unsolved math problem. Commenters were torn between “this is beautiful” and “nice trick, but does it actually prove what developers need?”
A delightfully chaotic little corner of the internet just got fed a very nerdy grenade: a new tool called FPSan tries to check whether two floating-point programs are really doing the same thing, even when the usual rules of arithmetic get slippery on computers. The twist? It works by turning those programs into totally different integer-based versions, and its correctness leans on a famously hard unsolved math idea. Yes, really. This is the kind of story that makes normal people blink and makes programmers race to the comments.
And race they did. One camp was immediately in full wonder mode, basically saying, “this is bonkers, but in a beautiful way.” The loudest praise came from people dazzled that messy decimal-heavy code could be transformed into what one commenter called “weird integer programs” while still preserving the right notion of equivalence. In tabloid terms: the math fans were swooning.
But then came the pushback, and that’s where the drama kicked in. The skeptics weren’t attacking the cleverness so much as the usefulness. Their big complaint: it’s nice if equivalent programs stay equivalent after the transformation, but what people really want is the reverse test — if the transformed versions match, can you confidently say the originals match too? That “cool idea, but is it enough?” energy became the thread’s mini-feud.
So the vibe was equal parts astonishment, side-eye, and nerd comedy: a sanitizer that proudly makes programs slower, changes what they do, and depends on deep math? On the internet, that’s less a tool launch and more a popcorn event.
Key Points
- •The article introduces FPSan, a Triton compiler pass built to help verify algebraic equivalence of floating-point programs by replacing float operations with integer-based ones.
- •The stated equivalence guarantee is conditional on the real version of Schanuel’s conjecture and applies only to programs with control flow independent of floating-point inputs and a restricted set of operations.
- •The supported operation set includes constants {-1.0, 0.0, +1.0}, ring operations {−, +, ×}, and the exponential function exp.
- •The article says these restrictions still cover common machine-learning GPU kernels, including matrix multiplication and much of self-attention.
- •FPSan uses a bijection from IEEE-754 single-precision floats to integers modulo 2^32, then defines transformed addition, subtraction, multiplication, and exponentiation through that embedding.