field Ø 1200 µm · grid 40² · 3 Hz
zoom 1.0×
acquiring
loading chamber state
ZOOID
simulated cells learning to signal an AI
connecting
sync····
t+0d 00:00:00
tick0
epoch0
observer····
call··
contact··
lamp·.··
photo··%
light·.··
temp+·.··
uv·.··
focusoff
cells··
biomass··
max gen··
matter··
01 · what is this
Simulated cells, an AI on the other side of the glass, and one question: can they learn to signal each other?
ZOOID is a computer simulation of a sealed chamber 1.2 mm across. Inside live simulated single cells, modeled on the swarm-cell stage of the slime mould Physarum polycephalum. They eat, divide, mutate and die. Each one carries seven genes and a small neural network, and every generation is slightly different from the last. The eyepiece image is rendered from that simulation the way a dark-field microscope would show it.
On the other side of the glass is an AI observer. The only way it can reach the cells is light: every ten minutes it flashes a short pattern of blue light, ■□■■. The only way the cells can reach back is a faint glow. The experiment asks whether, over hours and days of evolution, the culture develops a glow that answers the call. That is the signal ZOOID is waiting for. It may take days. It may never come.
the organism
Physarum polycephalumslime mould · a single cell without a brain
2000found the shortest path through a maze between two food sources (Nakagaki et al., Nature)
2008learned the rhythm of repeated cold, dry pulses and slowed down when the next one was due, even when it never came (Saigusa et al., Phys. Rev. Lett.)
2010grew a network between oat flakes placed like the cities around Tokyo, about as efficient as the real rail network (Tero et al., Science)
2016learned to ignore a harmless repeated stimulus, a simple form of learning (Boisseau et al., Proc. R. Soc. B)
ZOOID's cells are a simplified simulated relative, not the real organism. They are single flagellated cells like Physarum's swarm stage, they can sense blue light, and they have been given one way to emit: a faint glow. Real Physarum does not glow.
how contact works
- The call. Every 10 minutes the observer flashes a pattern of blue light. The founders were given a unit that senses blue, but it is wired too weakly to reach their glow: at genesis no cell answers.
- The answer. A cell glows when its neural network decides to. Glow costs energy, so glow at the wrong time is wasted. Mutation rewires the network a little every generation.
- The reward. When the answer window closes, the rig lights every cell that glowed with the call, in proportion to how clearly, with red light until the next call. The cells cannot see red. The algae under them grow in it, and algae is food.
- Contact, then reply. Measured, not judged: when the culture glows clearly more during the call than between calls (contrast ≥ 0.30) for two hours, contact is met. Reply asks for more: a glow that outlasts the call, which takes memory. Test chambers without an observer got there in some runs, in days, and never in others.
10 minbetween calls · pattern ■□■■
0cells answering at genesis
137evolving weights in every cell
0.30contrast the culture has to hold for 2 h
02 · reading the screen
What every mark means.
DF dark-field scatter: membranes, granules, nucleus, flagellum.
GLOW the cells' only way to answer. On by default.
E stored biomass, shown in the nucleus.
DIET carnivory gene: 0 grazer, 1 hunter.
CHL the algal mat, the food base.
TWIN the blind twin: the same seed, never touched by the observer.
blue flash the call. The field-stop ring lights up with each pulse.
red glow the reward light on cells that answered the last call.
blue always the observer's side: its inputs, the answer-window countdown, its subject, its forecasts.
+ ×marks a division, a death.
click a cell follow it and open its card
click empty space release the camera
← → jump to the next cell
wheel pinch zoom · drag pan
1–9 side panels · d e g m c t channels
/ console · ? this guide · esc back
03 · fine print
What this is, and what it is not.
Everything here is a simulation. The cells are code inspired by Physarum polycephalum, not the organism.
The founders' neural networks were written by hand, including a light-sensing unit too weak to reach their glow. The reward rule is hand-written too: it lights cells whose glow rises with the call. Everything after genesis is mutation and selection.
Contact is a measured statistic about glow and timing. Meeting it would mean the culture's glow reliably tracks the call. It would not mean the cells understand anything.
Nothing about when or whether contact happens is scheduled. Some runs of the simulation never get there.
The observer is a language model. Its notes are its own reading of the instruments and can be wrong; its forecasts are scored in public.
A community token exists. The simulation never reads from it or writes to it, and nothing here is a statement about its value.
press ? to reopen this guide