Two different claims hide inside one question
"The universe is a neural network" can mean two things, and they need different evidence.
The weak claim is structural: the large-scale universe and the brain are organised in similar ways, so the same mathematics describes both. This one can be checked with a microscope, a telescope and statistics.
The strong claim is dynamical: the universe learns. Its laws are the stable result of something like training, the way the weights of a network settle after gradient descent. This one needs a formal model and predictions that differ from standard physics.
Both claims have serious work behind them. Neither is established.
The structural evidence: brain versus cosmic web
In 2020 the astrophysicist Franco Vazza and the neurosurgeon Alberto Feletti published a quantitative comparison of two images: thin sections of human brain tissue and cosmological simulations of the cosmic web, the network of galaxy filaments and voids that fills the observable universe (Frontiers in Physics, 2020).
They measured the same quantities on both: how density fluctuations are distributed across scales, how many connections each node has, how clustered the network is. Several of those numbers came out close, even though the two systems differ in size by about 27 orders of magnitude. Both networks also hold most of their mass in a passive component: water in the brain, dark energy in the universe.
What this shows is that two very different physical processes arrive at comparable network statistics. Whether the cosmic web does anything like thinking is a separate question these data leave untouched. Bronchi and tree crowns share fractal statistics too, because both solve the same problem: maximise surface while keeping volume low. The interesting question the measurement leaves open is which optimisation problem the cosmic web is solving, if it lands on the same statistics as an evolved brain.
The dynamical proposal: Vanchurin's model
Vitaly Vanchurin, a physicist at the University of Minnesota Duluth, made the strong claim formal in "The World as a Neural Network" (Entropy 22, 1210, 2020). He describes the universe as a network with two kinds of variables, trainable ones (like weights) and hidden ones (like neuron states), and studies how the system behaves while it learns.
In one limit, the learning dynamics can be written in a form resembling the Schrödinger equation. In another, the dynamics resemble Einstein's equations of general relativity. Vanchurin's argument is that quantum mechanics and gravity may be two regimes of one learning process. He later developed the idea with Yuri Wolf, Mikhail Katsnelson and Eugene Koonin into a theory of evolution as multilevel learning (PNAS, 2022).
The model is a proposal. It has not produced a measurement that only it can explain.
A third thread: gravity from information
In 2010 Erik Verlinde argued that gravity may be emergent, a consequence of how information is distributed on surfaces, and in 2017 he extended the argument to galaxy dynamics. If gravity comes from information, and a network is a machine for moving information around, then the two ideas meet: the pattern of gravity would be the connection pattern of the network. That is the step this programme takes.
What Neural Cosmology adds: a working substrate
The programme on this site takes the dynamical claim one level lower and asks what the smallest network capable of producing physics-like structure would look like.
Its preprint, Pointer Architecture v9.0, defines a formal substrate: pointers, rewrite rules, commits, an append-only archive and observers. It implements that substrate as a small programming language, Sixth. Forty demonstrations, checked by 646 automated assertions, build up from a single distinction to arithmetic, time, space, observers and universal computation, and on to self-maintaining structures. On the same substrate, a holographic construction reproduces the standard estimate of dark energy in substrate language, within a factor of about 0.73 of the observed value.
What this establishes is constructive: a network of this kind can generate the structures a network universe would need. Whether our universe runs on one is a separate question, and the paper states in advance what would break its claims.
What would count against the idea
- A substrate demonstration failing its automated checks on the released code (falsifier F0 in the preprint).
- Area scaling breaking on other substrate topologies, which would make the dark-energy result an artefact of the chosen lattice (F1).
- Brain–cosmic web similarity disappearing under better statistics or appearing equally in random networks with the same density.
Bottom line
The universe being a neural network is a live scientific hypothesis with formal models, one striking structural comparison and a first working substrate built as code. It has not yet become a finding. It should be judged the way any model is judged: by the predictions it makes and by what could break it. The book The Celestial Code walks through the full argument in plain language.