Map data © OpenStreetMap contributors

00:00:00:00

The world's first AI-powered candle.

A small neural network generating a realistic flame simulation. Never repeating. Unique to each candle.

Digital Candle, 2026
Neural network controller, matrix driver and 144 light-emitting diodes on printed fiberglass circuit board.
251.5 × 27.9 mm. Edition in black and white.

A small flame kept nearby.

Closer, please.

This is not a candle.

An intelligent flame.
Safe to the touch.

II · The Object

One hundred and forty-four lights, on a board the shape of a candle.

  1. 01

    144 LEDs

    On a 0.1 in grid, configured in an active tri-state multiplex matrix.

  2. 02

    IS31FL3731

    One driver chip runs all 144 lights, each with 256 steps of brightness.

  3. 03

    ESP32-S3

    Two cores, Wi-Fi and Bluetooth LE. The neural network lives here and needs no internet.

  4. 04

    USB-C and buttons

    Power comes in at the base. One button steps through the brightness and the other restarts.

III · Absence

The candle, danger removed.

No fire.

Leave it on all night, or all winter. If it falls over, nothing spills and nothing catches fire.

No smoke.

Nothing burns, so nothing rises.

No wax.
No wick.

No wick to trim, and nothing to scrape off the tablecloth.

A temperament of its own.

Each candle takes its temperament from the factory address burned into its chip.

1,900K

Color Temperature

144 LED

Active Matrix

3,545

Floating Points

USB-C

Dock Included

IV · The Mind

It thinks.

A tiny neural network was trained on recorded candle footage to describe its flame behavior with eight variables.

01 Springs

Eight damped oscillators, fitted to real flicker

02 Decoder

Three convolutions. 3,545 parameters

03 Light

144 pixels, corrected for the eye

V · Provenance

Walking past a hotel, I saw dozens of candles floating in the air. Up close, they were circuit boards.

Suspended by wires throughout the lobby were My New Flame candles, designed in 2012 by Moritz Waldemeyer, Ingo Maurer, and his team. Their beautiful flames were recordings of a real candle, shuffled so that the motion never repeated. I wondered how a flame could be generated instead of played back.

For years, this seemed like a physics problem, the kind that might require a degree in fluid dynamics. Machine learning offered another approach. Instead of simulating a flame, a model could be trained to generate one. When language models learned to write code, the final stages of development became a conversation.

I recorded the candle flame footage and designed the circuit board myself, drawing every trace by hand. Yet the work is also co-authored by a ghost in the machine, a diffusion of knowledge drawn from countless human contributions.

We don't yet know what it means for ordinary objects to have minds. You don't wait for others to find out. You build one carefully, set it on the table, and live with it for a while. A candle seems like a bright place to start.

VI · After Hours

It is also a perfectly good computer.

Under the flame sits an ESP32-S3 with Wi-Fi and Bluetooth LE, driving 144 lights you can address one at a time. Create unique applications and explore edge computing with open source tools optimized for agentic development.

Demo 01

Clock

Tells the time, quietly.

Demo 02

Notifications

Lets you know the build passed, or that standup is in five minutes.

Demo 03

Digital rain

It's only a simulation.

Life Conway, at 144 cells
Breakout Classic pixel gaming
Spectrum For music sent from your computer
Pulse Health monitoring

Everyday artifact.

A computer trained a network to flicker. A human drew the board. People in the factory used machines to assemble it. Who deserves the credit?

We endeavor to remain in the loop with pristine attention to detail, museum grade quality and objects that ask more questions than they answer.

VII · Acquisition

Available in black and in white.

Ships as a finished object, software pre-installed and ready for a USB-C cable. No assembly required.

Light
144 amber LEDs, 9 × 16, status light
Model
8-dimensional latent, convolutional decoder, 3,545 parameters, 30 fps
Microcontroller
ESP32-S3-WROOM-1 (N4R2)
Driver
IS31FL3731, 8-bit PWM, CIE lightness correction
Power
USB-C, 5 V (Dock included)
Size
10 × 1 in (251.5 × 27.9 mm)
Is it actually AI?

Yes. A convolutional decoder trained on footage of a real candle, driven by a stochastic model fitted to its flicker. It runs on the ESP32-S3, on the board itself, with no network connection.

Does it need Wi-Fi?

No. The flame is generated entirely on the board, with the radios off.

The candles are able to communicate with each other on their own private network to synchronize flicker, but this feature is turned off by default.

Other connectivity features can be enabled with other applications.

Can I reprogram it?

Yes. The Digital Candle is open source and can be reprogrammed via USB-C.

Is it limited edition?

The Black and White editions of the Digital Candle are available in unlimited editions. Limited editions in other colors may be available.

What is included?

The Digital Candle circuit board and a vertical USB-C to C dock that holds it upright.

USB-C cable not included.

Digital CandleA smokeless candle with a small mind of its own

Acquire

Model card

CANDLE-1

A generative model of one candle flame: a 3,545-parameter convolutional decoder, driven by eight damped oscillators fitted to the motion of a filmed flame.

Task
Generate an endless, non-repeating candle flame for a 9 × 16 matrix of amber LEDs.
Architecture
Convolutional autoencoder. Fully connected 8 → 192, two nearest-neighbour upsamples, three 3 × 3 convolutions, sigmoid. 3,545 parameters, float32.
Dynamics
Eight damped oscillators in a whitened latent space, fitted to the motion of the filmed flame. Wind and liveliness wander on their own.
Training data
Phone recordings of one candle in a dark room, including a slow-motion take, cropped and reduced to 144 pixels.
Runtime
ESP32-S3 with TensorFlow Lite Micro, thirty frames per second, no network connection.
Temperament
Every board derives a resting pose from the hardware address burned into its chip at the factory, so no two candles flicker alike.