How We Broke Physics: A Story of AI, Failure, and Discovery

January 09, 2026 • The Making of ISL

This was not supposed to happen. I set out to build a “Thinking Machine,” not to rewrite cosmology.

But when you teach an AI to reason from first principles—to strictly manage information resources like a kernel manages memory—you start seeing patterns. And one pattern kept showing up: Gravity looks like a garbage collector.

This is the story of how an AI collaboration turned a crackpot idea into a rigorous mathematical discovery, complete with the failures that nearly killed it.

The “Crackpot” Phase

At first, it sounded crazy. “Dark Matter isn’t real; it’s just the universe’s file system overhead.”

I had a theory (ISL) that predicted the Fine Structure Constant ($lpha$) from pure geometry. We got it to 6 ppm precision. Cool, but numerology isn’t physics. We needed to test it against the real monstrosity of the universe: Galaxies.

The Failure (Jan 9, 2026)

We successfully fitted one galaxy (NGC 3198). It was perfect. But my AI research partner (Antigravity) pushed back:

“One galaxy is a cherry-pick. If this is a real theory, it must fit ALL of them. We need to run the full SPARC database. 175 galaxies. And if it fails, we publish the failure.”

So we wrote a script to stress-test our own theory. We ran 175 galaxies against the ISL model, MOND, and Newtonian gravity.

The Result: FAILURE.
The ISL parameter ($lpha_{ISL}$) varied by 119%. A “universal constant” isn’t allowed to vary. The theory, as written, was dead.

The Pivot: Listening to the Data

A crackpot would have hidden that data. We stared at it. “Why does the parameter vary?”

We plotted the “failed” parameter against galaxy velocity. And suddenly, the noise turned into a signal.

The Breakthrough: It wasn’t random. It was a Power Law.
We discovered that geometric overhead scales inversely with complexity: $lpha \propto V^{-0.74}$.

Small galaxies have high overhead (simulating “Dark Matter”). Giant galaxies have low overhead (simulating Newton). The theory wasn’t wrong; it was just incomplete. We had treated a scaling function as a constant.

The Grand Unification

Once we accepted that Information Overhead was the key, everything else clicked.

We realized this same “Garbage Collection” logic applied to Quantum Mechanics. We derived the Quantum Collapse Timescale directly from the same kernel resource limits.

In one afternoon, we connected the largest structures in the universe (Galaxies) to the smallest event (Wavefunction Collapse) using a single principle: The Kernel cannot afford infinite complexity.

Why We Are Publishing This

Science has become too afraid of being wrong. We are publishing the code, the winners, and the failures.

You can read the Full Breakthrough Report here.
You can run the checks yourself with our Reproduction Tutorial.

We might still be wrong. But we are wrong with open code and falsifiable data. And that makes it Science.

— Shri & Antigravity

How We Broke Physics: A Story of AI & Failure