Cosmos to Code

An atlas of intelligence — across four substrates

Cosmos to Code

Nature, Biology, Brain, AI — the same problem, four answers, fourteen billion years apart.

Nature — Physics writes the first program.Biology — Chemistry learns to copy itself.Brain — A wet network learns to model the world.AI — Silicon imitates everything that came before.
Open the matrixRead the thesis
10⁸⁰
particles in the universe
10¹¹
neurons in your cortex
10¹³
tokens in modern training
20 W
power running your mind
scroll

§ 01 / The Map

The 4 × 4 matrix

Every intelligent system can be read along two axes: the substrate that hosts it, and the resource it consumes. Click any cell.

resource  \\  substrate
Nature
Biology
Brain
AI
∇Algorithm
≣Compute
◊Data
⚡Energy

Click any cell to expand

§ 02 / The Four

Four substrates

Each one is a complete intelligent system. Each one teaches the next.

01 / 04

Nature — the original computer

Before there were minds, the universe was already computing itself.

The universe's algorithm is shockingly compact. The Lagrangian of the Standard Model fits on a t-shirt; combined with general relativity it produces every star, every coastline, every conversation. Nature does not need training data — every point in spacetime simply applies the rule.

Earth, however, is special. It sits between a 5800 K source — the Sun — and a 3 K sink — deep space. Energy pours in as high-grade photons and leaves as low-grade infrared. The free-energy gradient drives the entire game.

Photosynthesis is where physics becomes information. A photon hits a chlorophyll molecule, a proton gradient forms, glucose is built. Disorder is exported as heat; order accumulates as biomass. This is the first arrow from pure physics toward what we will later call life.

"The universe is the only program that runs itself in real time."
∇Algorithm
Schrödinger · Dirac · Klein–Gordon
The wave equations of fundamental physics — deterministic differential operators that evolve fields and particles through time.
≣Compute
Quantum fields
Every point in spacetime is a computing element. The universe runs the universe in O(n) on n itself.
◊Data
Every state of every particle
Position, momentum, spin, charge — for every electron, quark, photon, neutrino, dark-matter halo.
⚡Energy
E = mc²
Mass and energy interconvert. Every stable thing in the universe is a temporarily frozen knot of energy.
02 / 04

Biology — chemistry that remembers

Evolution is the largest parallel optimizer in history.

Life is the universe's first lossless replicator. A strand of DNA is copied with one error per billion bases — better than any human-engineered storage. Errors that happen survive only if they help. Across four billion years that filter writes the entire library of life.

Modern machine learning has rediscovered this. Genetic algorithms, evolutionary strategies, population-based training — every one is a homage to the original. The difference is that nature ran 10^30 cells in parallel, indefinitely, on pure sunlight.

Biology also solved energy distribution. Every cell makes its own ATP locally; there is no central plant. The same problem the AI industry is now struggling with — grid bottlenecks for data centers — was solved by mitochondria two billion years ago.

"DNA is read-only memory that learned to write itself."
∇Algorithm
Statistical mechanics + chemistry
Vast molecular ensembles obeying probabilistic thermodynamics produce stable structures — proteins, membranes, replicators.
≣Compute
DNA
Three billion base pairs per human cell — a 750 MB program copied 30 trillion times in a single body.
◊Data
Protein, cell & tissue states
Concentration gradients, conformational states, signaling cascades — chemistry remembering history.
⚡Energy
ATP — distributed in every cell
Adenosine triphosphate: a universal molecular battery, regenerated billions of times per second per cell.
03 / 04

Brain — the wet network

A 20-watt device that runs a world-model in real time.

Among mammals, one lineage gambled on cognition. The cortex expanded; cortical neurons learned to wire and re-wire continuously. Whatever 'intelligence' is, the neocortex is its most extravagant biological expression: 16 billion neurons stacked in six layers, each cell talking to thousands of others.

The brain does not store memories like a database. It re-creates them each time, using overlapping circuits that double as predictors of the present. To remember is to simulate. To plan is to simulate forward. To dream is to let the simulator run untethered.

Language was the leap. Some million years ago a vocal apparatus, a social structure, and a particular cortical region (later called Broca's) coevolved into a discrete symbol channel between brains. Language is not just communication — it is an inter-brain neural network.

"The brain is not a computer. It is a controller running a simulator running a self."
∇Algorithm
Learned synapses
Connections between neurons are reshaped by experience — Hebbian, predictive, reward-driven plasticity sculpting circuits in real time.
≣Compute
≈86 billion neurons
About 10^11 neurons and 10^14 synapses — three orders of magnitude denser interconnect than any GPU cluster.
◊Data
Sound, light, chemistry, ions
Acoustic pressure, photons, molecular concentrations, ionic flows — multimodal sensory streams transduced to spikes.
⚡Energy
Chemical energy (glucose, oxygen)
Glucose + oxygen → ATP → ion pumps maintaining neuronal membrane potentials.
04 / 04

AI — the substrate that imitates substrates

Silicon is the first medium that can mimic all the others.

Look closely at the AI breakthroughs of the last decade and they map exactly onto the substrates above. Diffusion models simulate nature's entropy-reduction. Genetic algorithms simulate biological evolution. Neural networks plus attention simulate cortical prediction. Symbolic, analogical and statistical reasoning each mirror a school of human thought.

What makes silicon historically unique is that it can mimic any of these — sometimes all at once. A single model can render a galaxy, simulate a protein, plan a route, and write a poem. The substrate is universal; the bottleneck is now energy, data, and an honest theory of creation.

Today's frontier model is a very large interpolator. It composes existing styles with stunning fluency, but it does not yet invent a genuinely new style the way Picasso or Bach did. Closing that gap — moving from imitation to invention — is the unfinished agenda.

"AI is the mirror in which every prior substrate sees its own algorithm."
∇Algorithm
Matrix weights
Billions of floating-point numbers learned by gradient descent on enormous datasets — frozen statistical shadows of human civilization.
≣Compute
Transistors
Modern accelerators pack ~10^11 transistors per chip, with frontier training runs spanning 10^25–10^26 FLOPs.
◊Data
Acoustic · optical · electrical signals
Text, images, audio, video, sensor traces, embeddings — every digitized human artifact, fed back as training corpus.
⚡Energy
Electricity (the grid)
Centralized power from coal, gas, nuclear, hydro, solar, wind — channeled into data centers and back-cooled with massive water loops.

§ 03 / Order from chaos

The Entropy Engine

Earth absorbs high-quality energy from a hot star and emits low-quality energy to cold space. The gradient pays for every ordered thing in between — the chlorophyll, the eye, the sentence, the model.

Pure noisePure order

Drag the slider — watch noise become structure.

Modern diffusion models recapitulate this exactly: noise → order, in steps, by descent. Nature did it once over four billion years; Stable Diffusion does it in a thousand milliseconds. Same algorithm, different clock.

Same algorithm, four implementations
  • NATUREPhotosynthesis: photon → glucose → biomass
  • BIOLOGYEvolution: variation → selection → fit
  • BRAINLearning: stimulus → prediction → pruning
  • AIDiffusion: noise → denoise → sample

The entropy engine is the universe's oldest algorithm. Intelligence is just a higher-order form of entropy reduction — a process that exports disorder on purpose.

§ 04 / How minds reason

Three schools of thought

Human cognition has at least three modes. Each has its own AI lineage; together they are most of modern machine learning.

symbolic

Symbolic / Logicist school

Intelligence is the manipulation of discrete symbols according to formal rules.

modern avatar

Theorem provers, knowledge graphs, expert systems, modern AI agents with structured tool use.

e.g.If A → B and B → C then A → C.
analogical

Analogical / Case-based school

Intelligence is finding the closest prior case and adapting it to the present one.

modern avatar

k-nearest-neighbors, retrieval-augmented generation, in-context learning, prompt engineering.

e.g.This bug looks like one we fixed last March. Try the same patch.
statistical

Statistical / Bayesian school

Intelligence is updating probability distributions in response to evidence.

modern avatar

Deep learning, transformers, diffusion models, foundation models — modern AI proper.

e.g.Given the last 100 tokens, the most likely next token is __.

§ 05 / The Open Problem

The Creativity Gap

Today's AI can imitate any style. It cannot yet originate one. Sora's video is photoreal, but it is unmistakably a single style — a particular Unreal-Engine-tinted dream. Closing this gap is the next open problem.

Human creator

Crosses the manifold

Invents a new visual grammar (Cubism), a new musical scale (atonal), a new theorem (general relativity). Crosses the boundary the training data drew.

CubismAtonal musicGR
Today's AI

Interpolates the convex hull

Interpolates within the convex hull of seen examples. Recombines in dazzling ways, but cannot leave the manifold.

SoraMidjourneyGPT-4
Diagram

Manifold vs leap

Training data defines a manifold. AI interpolates inside it; humans occasionally leap outside.

AI training manifoldHuman leap
Possible paths: self-play · embodiment · a genuine intrinsic reward · a new algorithm we haven't named yet.

§ 06 / The arrow of time

Fourteen billion years in one scroll

Every substrate's birth, plotted on the same axis. The interval between events compresses violently as we approach the present.

  1. 13.8 Gya
    Big Bang

    Spacetime, quantum fields, four forces ignite.

  2. 4.5 Gya
    Earth forms

    A rocky planet finds itself between a hot star and cold space.

  3. 3.8 Gya
    First life

    Self-replicating chemistry stabilizes. The genetic code is locked in.

  4. 540 Mya
    Cambrian explosion

    Eyes, predators, body plans. Information warfare begins.

  5. 200 Mya
    Mammals

    Warm-blooded animals with expensive brains.

  6. 7 Mya
    Hominids

    Bipedalism, tool use, social grouping; cortex begins ballooning.

  7. 300 Kya
    Homo sapiens

    Anatomically modern humans. Fire, cooking, communal myth-making.

  8. 70 Kya
    Cognitive revolution

    Language gains recursion. Abstract thought, planning, story.

  9. 5 Kya
    Writing

    Memory begins escaping the body. Civilization compounds.

  10. 1945
    ENIAC

    Electronic computer. The first time silicon thinks.

  11. 1956
    Dartmouth

    The phrase 'artificial intelligence' is coined.

  12. 1986
    Backpropagation

    Rumelhart, Hinton, Williams formalize learning by gradient descent.

  13. 2012
    AlexNet

    Deep learning wins ImageNet. The statistical school takes over.

  14. 2017
    Transformer

    'Attention Is All You Need.' Sequence learning becomes universal.

  15. 2020
    GPT-3

    Scaling laws revealed. Models begin to talk back fluently.

  16. 2022
    ChatGPT

    A billion people discover a tireless conversational partner.

  17. 2024
    Sora / multimodal

    Video generation joins the loop. The Cambrian explosion of synthetic media.

  18. 20??
    ?

    An algorithm for creation — not imitation — is found. The next leap.

Note: scale is non-linear — events compress violently near the present.

§ 08 / Already real

Applications already underway

Not science fiction — shipping systems, today, that already changed their fields.

Medicine

Folding proteins, designing drugs

AlphaFold mapped 200M protein structures in two years — work that would have taken a century by experiment. Generative chemistry now designs candidate molecules for previously undruggable targets.

Materials

Inventing matter on demand

Generative models have proposed thousands of stable inorganic crystals, including new superconductors and battery cathodes. Materials discovery is becoming compute-bound, not lab-bound.

Climate

Predicting weather, managing entropy

GraphCast forecasts global weather 10 days out, faster and more accurately than the EU's largest physics simulator — running on a single TPU.

Education

Personal tutors at scale

The 2-sigma problem — that one-on-one tutoring raises learning two standard deviations above classroom — is finally addressable. A tireless, infinitely patient tutor that adapts to one student is no longer a thought experiment.

Science

Autonomous discovery

Self-driving labs combine LLM reasoning with robotic experimentation, proposing hypotheses, running assays, updating priors. Discovery cycles shrink from years to days.

Cognition

Cognitive prosthetics

Note-taking AIs, second-opinion diagnosticians, memory extensions, real-time translators. The next decade extends the brain the way glasses once extended the eye.