An Ending That Was Fixed and Still Unreadable
The rules never branched. Every program still guessed like a coin.
Chaos is the usual excuse when a forecast fails. Tiny differences at the start grow into different endings. Lars Koopmans, Elinor Kay and Hyun Youk at the University of Illinois built a system that does not have that excuse.
It is a simulation. The rules are local and known. Each starting pattern has one legal ending. The grid finishes either still or moving, and that finish is fixed from the first step. They ran a million simulations and handed six kinds of programs the easiest question: still, or moving? The programs had been trained on as many as 850,000 runs. Simple statistics and deep networks both landed near 50 percent. A coin flip.
The paper is Nature Communications, 11 September 2026. Earth.com wrote it up on 23 October. The EurekAlert is 7 October. Inside the feature window, and the claim has not been walked back.
Later in each run, vortices appear. Youk and Koopmans say they only learned to see them by watching the movies. The vortices are positive, negative or neutral, and their numbers fall as the run goes on. That emerging structure can be used as a predictor. Predictability shows up. It is not sitting in the initial conditions in any form the machines could read.
Key Evidence
- One million simulations. Six program types. About 50 percent accuracy on a yes-or-no question (Koopmans, Kay, Youk, Nature Communications, 11 September 2026).
- The system is deterministic and not chaotic.
- Vortices that appear mid-run become a usable predictor. The start does not.
The Rational Explanation
"These programs could not predict it" is not "no program could." A lookup table of every start and every end would score perfectly and explain nothing. The result is about a model. It is not a claim that nature hides its endings, and it is not an argument about free will.
What We Don't Know
Whether this kind of unreadability shows up outside the grid. The authors call it an accidental finding. They have a mechanism inside the model, the vortices, and no evidence yet that a shoreline or a cell uses the same trick.
The Rabbit Hole
Fixed is not the same as predictable. That is the whole discomfort. The answer is written. The ink only becomes readable after the page has turned itself.