People often ask me where this all started. I have spent twenty years in the AI industry, and at some point the phrase "the universe is a neural network" surfaced in my head. How it got there, given that I am neither a neuroscientist nor a cosmologist, is the subject of the short timeline below.
2017–2019: architecture as language
I had been working with transformers and attention mechanisms since the first papers came out. I worked with systems that learned from data and demonstrated properties no one had programmed in. By 2019 I had a stable intuition: architecture matters more than algorithm. The way the connections are arranged determines what the system can and cannot do.
There was nothing metaphysical about it: anyone who has run an LSTM and a Transformer on the same task has seen the difference first-hand.
2020: Vanchurin's paper
In September 2020, Vitaly Vanchurin posted a preprint on arXiv titled The World as a Neural Network. The title sounded like a pop-science headline, but inside was a proper formalism. Vanchurin showed that if you take a large neural network with the right learning dynamics, then in specific limits you can derive quantum mechanics (as the fast-learning limit) and gravity (as the slow one).
In 2022 Vanchurin, with Katsnelson and Koonin, published a follow-up in PNAS. Koonin is a well-known evolutionary biologist and Katsnelson a condensed-matter physicist; the journal is first-rank, and the work had nothing fringe about it.
It gave me an odd feeling. Someone had put into formal words what I had felt as an engineer but could not yet say as a scientist.
2020–2022: five cracks
After that, things piled up slowly. I started noticing related observations arriving from different fields, with no one stitching them together.
— Vazza and Feletti (Frontiers in Physics, 2020) showed that the brain's neural network and the cosmic web are statistically indistinguishable.
— Bérut and colleagues (Nature, 2012, widely discussed 2019–2020) experimentally confirmed Landauer's principle: information has a thermodynamic cost.
— Michael Levin (Tufts) published paper after paper: cells know form before genes, bioelectric patterns encode morphology, xenobots self-organise from frog fragments into new life forms.
— Tononi kept developing IIT and began working it out for computable cases.
— New experiments gave Penrose and Hameroff some support: microtubules may hold coherence longer than previously thought.
The five news items came out in different journals, but laid side by side they make one claim, stated five ways: information is the primary layer, matter and force are emergent, consciousness is a property of integrated processes.
2022–2023: the book has to be written
By 2022 I felt someone had to state this synthesis out loud, in Russian as well. In English it lived in fragments across academic Twitter and podcasts, and no book had yet collected it while keeping what is proven honestly apart from what is not.
I am neither a neuroscientist nor a cosmologist. But I had spent twenty years with systems where architecture determines what is possible, and I had talked with researchers from all five fields. It was an unusual position. A specialist has more depth and a journalist more breadth; what I had was an engineer's view of a science that had not yet become science.
2023–2025: Pointer Architecture
While I was writing the book, a working hypothesis of my own grew alongside it. I called it Pointer Architecture: a formal model of a discrete substrate in which everything else grows out of a first distinction and pointers, all the way up to an estimate of the dark-energy density.
In 2024 I assembled the preprint, and it has since gone through nine versions. The current one, v9.0, builds the substrate as working code — a small language called Sixth — with falsifiers written down in advance. The code is public.
This is a long way from proving Neural Cosmology. The preprint tests one of its consequences and says in advance which result would refute it.
Where I am now
Neural Cosmology is a working hypothesis with clearly marked edges. Part of it is being tested, part is speculation for now, part is personal interpretation. I try not to mix those three.
If the hypothesis is right, science gains a new field. If it is wrong, what remains is a concrete, testable construction, and whoever takes up the question next will not have to start from zero.
Which way it goes, I do not know. That is why I keep going.