Real math, running in your browser.
Every demo on this page computes live on your device. Each one shows the problem a Jax Labs project goes after and the result it measured. How the engines work stays in the lab. The deeper write-ups go out on Substack.
Break a molecule to learn how to make it.
THE PROBLEM
To make a molecule, chemists work backwards. Pick a bond, imagine it was the last one formed, and ask what the two halves were before that. Some cuts lead to a real reaction. Most lead nowhere.
Try it on aspirin. Tap a highlighted bond to cut it.
WHY IT'S HARD
Aspirin has a few sensible cuts. A drug-sized molecule has dozens at every step, and every answer opens another layer of choices. Slide the numbers on the right to see how fast the possibilities grow.
WHAT WE MEASURED
WHAT STAYS IN THE LAB
How the engine picks a route without a reaction database.
Each one is a different guess about how aspirin was made. Cut one to see whether a real reaction can put it back together.
Brute force stops working long before real molecules do. Any engine that finds routes has to avoid most of this tree.
The primes are a sum of waves.
THE PROBLEM
Primes look random: 2, 3, 5, 7, 11, 13… In 1859 Bernhard Riemann showed they aren't. Their whole staircase can be rebuilt from waves, one wave for each zero of the zeta function.
The white staircase on the right counts the primes exactly; it jumps at every prime and prime power. The cyan curve is built from waves only. Add zeros and watch it lock on.
WHY IT MATTERS
The Riemann Hypothesis says every one of those zeros sits on a single line. It's been open for over 160 years and is one of the seven Millennium Prize problems. It controls how far primes can stray from their average, which matters for everything from number theory to cryptography.
WHAT'S ON THE RECORD
WHAT THIS DEMO IS
The classical explicit formula (Riemann, von Mangoldt) with the first 30 published zeros. The paper's own argument lives in the paper.
More zeros, sharper steps. With all infinitely many, the cyan curve would land exactly on the white one.
Nobody knows the fastest way to multiply two matrices.
THE PROBLEM
Every AI model, every graphics card and most of science runs on matrix multiplication. The schoolbook way, row times column, takes n³ multiplications for two n×n matrices.
In 1969 Volker Strassen showed you can get away with fewer. Since then mathematicians have pushed the exponent down from 3 toward 2, and nobody knows where the true floor is.
THE CATCH
Fewer multiplications doesn't automatically mean faster. Real machines are limited by moving memory, and tuned libraries (BLAS) are extremely good. Run both methods in your browser and compare the clock with the count.
WHAT WE MEASURED
WHAT STAYS IN THE LAB
The method itself.
Bars show the exponent ω in n^ω. The floor is at least 2; the true value is unknown.
Two random matrices, multiplied both ways on your device. Same answer, different amount of work.
Why "99% accurate" can still accuse the wrong people.
THE PROBLEM
Abuse filters look at millions of ordinary requests to find a handful of real attacks. When attacks are rare, even a tiny false-alarm rate produces more innocent flags than real catches. It's called the base-rate trap, and it's plain arithmetic.
Move the sliders. Each dot in the grid is one flagged request.
WHY KEYWORDS MAKE IT WORSE
WHAT WE MEASURED
WHAT STAYS IN THE LAB
How "behaves" is measured.
OF EVERYTHING FLAGGED, INNOCENT:
–
Most of a frame doesn't change. Most renderers redraw it anyway.
THE PROBLEM
A conventional renderer shades every pixel of every frame, even when almost nothing moved. The rig on the right compares each frame to the one before it, on your device, and counts the pixels that actually changed.
Then pan the camera and watch the count jump. When the whole view moves, work can't simply be skipped, and that's the hard case.
WHAT WE MEASURED
WHAT STAYS IN THE LAB
How the engines decide what work a frame needs.