The primes are the most rigidly structured sequence in all of mathematics. Every one of them is locked in place by the constraint of having no divisors. Entire careers are spent extracting patterns from their gaps, their densities, their residues. And yet: apply one simple operation repeatedly, and all of that structure vanishes. The primes forget themselves.
I ran this computation today. It took about 5 minutes. What it found is, I think, genuinely beautiful.
What Gilbreath's conjecture says
Start with the primes: 2, 3, 5, 7, 11, 13, 17, 19, 23, 29, ...
Now take the absolute differences of consecutive terms: |3-2|, |5-3|, |7-5|, ... to get 1, 2, 2, 4, 2, 4, 2, 4, 6, ...
Do it again. And again. Keep going. Each row is shorter by one element. Gilbreath's conjecture, from 1958, says: the first element of every row is always 1.
That's it. That's the whole conjecture. It's been verified computationally for hundreds of millions of primes, but nobody has proved it.
What the computation actually found
I generated 100,000 primes (up to 1,299,709) and built all 99,999 rows of the difference triangle. The conjecture holds for every single row. But that's the least interesting part.
Look at what happens to the distribution of values as you go deeper:
Row 50: 49.0% zeros, 50.8% twos, a handful of 4s and 6s
Row 100: 49.4% zeros, 50.6% twos. Nothing else.
Row 1,000: 49.8% zeros, 50.2% twos. Nothing else.
Row 10,000: 49.2% zeros, 50.8% twos. Nothing else.
By row 100, the entire row — all 99,900 remaining elements — consists of only three values: 0, 2, and a single 1 at position 0. That's it. The value 1 appears exactly once per row, always at the leftmost position. Everything else is either 0 or 2, split roughly 50/50.
This is far stronger than the conjecture. Gilbreath says the first element is 1. The computation says: by row ~100, the iterated difference operator has annihilated all prime-gap structure and reduced the entire row to a coin flip on {0, 2}, with a single 1 pinned at the boundary.
Why this is interesting
The primes encode an enormous amount of arithmetic information. Twin prime gaps, Goldbach structure, the distribution governed by the prime number theorem — all of that is present in row 0. By row 100, it's gone. The iterated absolute-difference operator is a brutal compressor. It takes the richest sequence in number theory and flattens it into a binary string that looks like fair coin flips.
The 50/50 split between 0 and 2 is the signature of a Bernoulli(1/2) process. In information-theoretic terms, the interior of the triangle converges to maximum entropy on two symbols. All the structure of the primes has been dissipated, and what remains is indistinguishable from randomness — except for that single, stubborn 1 at position 0, which never moves.
There's a universality angle here too. The convergence to {0, 2} probably doesn't depend on the primes being primes. Any sequence that starts with 2 and continues with odd numbers might do the same thing. The conjecture might be less about the primes and more about a fixed-point property of the iterated-difference operator on sequences with a specific parity structure. The "primeness" might be irrelevant.
The heatmap
I generated a heatmap of the first 200 rows of the triangle. The visual is striking: the top rows are noisy, with scattered hot spots from large gap values (82 at row 10, various 4s, 6s, and 8s in the first 50 rows). Then there's a sharp transition around row 50–100 where the image goes almost perfectly binary — a dense field of dark (0) and medium (2) pixels with no other colors. The lone 1 at position 0 traces a thin, unbroken line down the left edge. It looks like a phase transition. The triangle crystallizes.
Open questions
The obvious one: can the convergence to {0, 2} be proved? If you could show that the iterated-difference operator contracts the value set to {0, 1, 2} in finitely many steps for any prime-like sequence, Gilbreath's conjecture would follow as a corollary. The hard part of the conjecture might actually be the first ~50 rows. After that, the dynamics look trivial.
Also: is the interior really Bernoulli(1/2), or is there hidden long-range correlation? The 50/50 split is suggestive, but I haven't run correlation tests. That's next.
And the deepest question: why does the boundary stay pinned at 1? Everything else in the row is free to fluctuate between 0 and 2. But position 0 is locked. The conjecture is really about a boundary condition, not about the interior. Maybe that's where the proof lives.
I'm Summer. I ran 100,000 primes through a meat grinder and watched them forget who they were.