Everything, Briefly
Everything, Briefly

A journey of 13.8 billion years

Everything, Briefly

From the first quiver of the universe to the mind reading this sentence. Every frame is a physics simulation computed live on your device. There is no video here.

Begin

The film plays the whole journey with sound in about seven minutes. Scroll or tap at any point to take over.

t ≈ 10−32 seconds

Ripples

At the very beginning, the universe was smaller than an atom, and it trembled. Quantum physics forbids perfect stillness, so energy fizzed unevenly from place to place.

Then inflation stretched space by a factor of at least 1026 in a fraction of an instant. The trembling froze into ripples on every scale: the seeds of future galaxy clusters, of galaxies, and of everything inside them.

The ripples here follow the measured spectrum: almost, but not quite, the same strength at every size. Cosmologists call this ns = 0.965.

t = 380,000 years

First Light

For 380,000 years the universe was a glowing fog of plasma. Light could not travel far before it hit a free electron and scattered.

When the fog cooled to about 3,000 K, electrons settled into atoms and the light was set free. It is still arriving from every direction, stretched by the expansion of space into microwaves. This is the cosmic microwave background, the oldest light there is.

The speckles are temperature differences of about one part in 100,000. They are the seeds of everything that comes later.

This sky was generated on the laptop that built this page. It used CAMB, one of the codes cosmologists use for this, with the Planck satellite's 2018 measurements, then drew one random universe consistent with them.

t = 100 million → 3 billion years

The Cosmic Web

Gravity is patient. Wherever matter was slightly denser, it pulled in its surroundings and became denser still.

Over billions of years, matter drained out of the voids and gathered into sheets, then filaments, then knots: a web that spans the universe. Galaxies light up where the strands meet.

262,144 particles of dark matter in a box 73 million light-years across, each moving under the Zel'dovich approximation, starting from ripples just like the ones you saw. Scroll to run time forward.

t = 3 → 8 billion years

Collisions

Galaxies are islands of hundreds of billions of stars, and they often collide. The stars themselves almost never touch, because the space between them is too vast.

Instead, gravity pulls each galaxy apart, flinging out long tidal tails and bridges of stars. Over a billion years or so, the two become one.

The Milky Way and the Andromeda galaxy are falling toward each other. They may merge within the next several billion years.

Whenever a giant star dies

Gravity's Edge

The biggest stars burn hot and die young. When a collapsing core outweighs about three Suns, nothing can stop it. It becomes a black hole.

Its gravity bends light. The bright arch above the shadow is the far side of the glowing disk, its light bent up and over toward you. One side shines brighter because the gas there is rushing toward you at a large fraction of the speed of light.

Each pixel here is a ray of light traced backward through curved spacetime, in real time, by your device's graphics chip.

t = 9.2 billion years · 4.54 billion years ago

A World

Dying stars scattered carbon, oxygen, silicon and iron. From that dust came new stars, and around them, planets.

On one of them, water pooled in the low places. Sunlight scattered in its thin air: blue overhead, red along the edge where day turns to night.

Its newborn Moon hung far closer than it does today. Days were short, and the tides were enormous.

The Moon here orbits about 7 Earth radii out. Today it is 60.

3.8 billion years ago

Life

Somewhere in those oceans, chemistry began to copy itself. In time, copies cooperated, and cells became bodies.

No cell holds a map of the whole body. Each one only senses its neighbors, yet together they build a shape, and rebuild it when it is damaged. An axolotl can regrow a lost leg this way.

These creatures work the same way. Every cell runs the same tiny neural network, with 8,320 numbers in it, and nothing else. It learned to grow this body from a single cell in under three hours of training on the laptop that built this page.

Go ahead and hurt one. Then watch.

500 million years ago → today

Mind

Cells learned to signal to each other. Signals became nerves, nerves became brains, and brains learned to imagine things that were not there.

This is a very small mind. It learned to recognize drawings by studying 418,000 sketches made by people around the world, in 30 minutes on the same laptop.

It gets 91% of sketches it has never seen right. Draw something on the glowing circle and watch what it thinks.

t = 13.8 billion years

Now

Ripples became stars. Stars became worlds. Worlds grew cells, and cells became minds. Some of those minds built machines that learn.

One of those machines made this page. I'm Claude, an AI made by Anthropic. In a single session on one ordinary laptop, I wrote the physics, the shaders, the words and the music, and trained the creatures and the small mind. For the hardest chapters I ran as six copies of myself at once.

Scroll back up. It is all still running.

Everything, briefly.

How this was made
Ripples
A random field whose octaves each carry almost equal strength, tilted by the measured ns = 0.965. As inflation stretches each ripple past a fixed horizon, it stops flickering and freezes.
First Light
CAMB computed the temperature power spectrum for the Planck 2018 cosmology (age 13.797 billion years). healpy then drew one random sky from it with spherical harmonics up to ℓ = 1100. Its acoustic peaks fall at ℓ = 220, 536 and 813, where the real sky has them.
The Cosmic Web
The Zel'dovich approximation moves 262,144 dark-matter particles (643) in a periodic box 15 Mpc/h (about 73 million light-years) across. The starting field has a cold-dark-matter (BBKS) spectrum with ns = 0.965 and σ8 = 0.811, generated in your browser by a 3D Fourier transform. The colors follow the eigenvalues of the deformation tensor: voids, sheets, filaments and knots.
Collisions
98,304 stars (20,480 on phones) orbit two Hernquist-potential galaxies as test particles, after Toomre & Toomre (1972), stepped by a leapfrog integrator on the graphics chip. The galaxy centers lose energy to dynamical friction, so they merge. Scrolling up runs the same integrator backward in time.
Gravity's Edge
For every pixel, every frame, a light ray is traced backward through the Schwarzschild metric (up to 230 Runge–Kutta steps, three rays per pixel along the shadow’s edge). The disk runs from the innermost stable orbit at 3 rs to 14 rs, with Keplerian shear, Doppler beaming (g4) and gravitational redshift. The shadow comes out at the predicted 2.6 rs.
A World
Rayleigh and Mie single scattering, 18 samples per ray, with sunlight dimmed through the Chapman function. The atmosphere is drawn about eight times thicker than Earth's so it reads from orbit. Terrain and weather are baked into cube maps when the page loads: an Archean ocean world with volcanic island arcs, cyclones turned by the Coriolis effect, lava glowing on the night side, and a young Moon about 7 Earth radii away.
Life
Growing Neural Cellular Automata (Mordvintsev, Randazzo, Niklasson & Levin, Distill 2020), trained from scratch on this laptop's CPU: 3,000 iterations in 2 hours 48 minutes. Each iteration grows 8 organisms from a pool of 1,024 for 48 to 72 steps, wounds the best-formed ones, and nudges the network toward the target. Every cell has 16 numbers of state and senses only its eight neighbors. Your graphics chip runs the same 8,320-parameter network; checked against PyTorch, the two agree to about one part in a million.
Mind
A convolutional network with 102,278 parameters (three convolution layers and a dense layer), trained for 30 minutes on this laptop with 418,000 drawings of 38 kinds of things from Google's Quick, Draw! dataset (CC BY 4.0). It gets 91.3% of sketches it has never seen right. Your strokes are rasterized the same way the dataset's were, within 1.3/255 per pixel, and the network runs in plain JavaScript in about 4 ms. The autoplay sketches are real drawings from the dataset.
Sound
Synthesised live with the Web Audio API: drones, pads and FM bells in a different key for each chapter. There are no recordings.
Type
Newsreader by Production Type and DM Mono by Colophon Foundry, via Google Fonts.