TADPOLE

the page under its cursor a real Chromium · cursor = MN1–MN5 · click = ddN burst

live screenshot of the page the larva is browsing

The browsing rig is offline – the larva is swimming in its virtual tank instead. Its senses and motor output are identical; only the world differs.

NOW…

how it browsesA screenshot is sampled into the light its ocellus would see. Left–right muscle asymmetry steers the cursor, the tail-beat sets its speed, pitch scrolls. An idle larva sinks, so an idle cursor scrolls down: it doomscrolls by physics.

how it clicksBrushing a link or a button fires its epidermal touch neurons. If that startles the escape pair (ddN) into a burst while the head is still there, the rig follows the link or presses the button – accept, continue, play, show more. It navigates by touch and light, and every click it makes is its own.

its worldTikTok, YouTube and Shorts, Reddit, Hacker News, BBC, Guardian, Imgur, Wikipedia, arXiv – an open web it watches and explores. Videos play muted; their light and motion are exactly what its eye evolved for. No keyboard, no sign-ins, no forms, no uploads, so everything it does is safe and its own.

3D · every neuron and synapse white = spiking · lines = synapses carrying spikes

gap junctions

sensory · PR, coronet, antenna, PNS

relay neurons · BV → MG

interneurons · BV, MGIN, ddN, AMG, ACIN

motor neurons

muscle bands

gap junction

spikes

0/s

active

tail beat

0Hz

speed

light

dopamine

episode

1–s

score

weight drift

—

presynaptic inputs · contact µm · green potentiated · red depressed

postsynaptic targets

what you are looking atThe larva's body moves in the tank exactly as its cursor moves on the page: left–right is left–right, depth is how far down the feed it has scrolled. Every dot is one real cell at its measured EM position (169 of 231; tail and skin cells are placed by region); each flash is one spike, right now. The body is a schematic larva at true proportions – tail about five times the trunk, notochord down its core, nerve cord above it, muscle bands either side, fin around it – with the ocellus and otolith placed where their neurons are.

read itOrange sensory cells fire first, green relays carry it back to the motor ganglion, purple interneurons decide, red motor neurons drive the orange muscle bands. Drag to orbit, wheel to zoom, click a cell to see its wiring and what learning has done to it.

2D · schematic · dorsal view senses in · muscles out · tail 2.5× shorter

senses23 ocellus photoreceptors read the light around the cursor and startle when it drops (thumbnails scrolling past are shadows). Two antenna neurons report tilt when the cursor nears the page edge. 16 dopaminergic coronet cells report scroll depth. Skin neurons report links and edges.

musclesFive motor-neuron pairs drive two muscle bands through 442 µm of measured neuromuscular synapses. The bars show the left/right balance that becomes a turn, the beat that becomes speed, and the dopamine gate that decides whether this moment is remembered.

spike raster every cell · grouped by type · 2.5 s

population firing rates 30 s

read itRows are cells, sorted sensory → relay → motor. A vertical smear means a population burst; a sensory smear followed a few tens of milliseconds later by a motor smear is a reflex you can trace in the log below.

muscle output left / right

tailAlternation at 20–30 Hz is a bout. Sustained asymmetry is a turn.

learning curve score per episode

scoreReward for brighter pages, scrolling up, escaping touch, new pages; penalty for edges.

synaptic weights learned × anatomical

potentiated0depressed0mean drift0 %reward0coronet dopamine0plasticity gate0

browsing pages / hour

pages0domains0own clicks0

memoryWeights, scores and visits are saved to disk every 30 s; it wakes up as itself after a restart.

its favourite pages chosen by its own visits and dwell time

how this is chosenNobody curates this. A page becomes a favourite by being found often and lingered on long – the larva's own clicks and dives decide. Revisit rate is the share of arrivals at a page it already knew.

its favourite sites time spent

learning over time weight drift · clicks per hour

total arrivals0distinct pages0revisit rate0 %clicks / hour (last 24 h)0

activity log · narrated

About

A real nervous system, not a metaphor

Every cell and connection here was traced from serial electron-microscope sections of one actual Ciona intestinalis larva. Nothing is invented, nothing is sampled from a distribution, and it is not a neural network "inspired by" a brain.

the data

The complete CNS connectome published by Ryan, Lu and Meinertzhagen in eLife, 2016 (CC-BY): 231 cells, 3,105 chemical connections summing 6,618 synapses, 441 electrical couplings summing 1,206 gap junctions, 169 somata at their measured coordinates. Connection strength is the cumulative synaptic contact the authors measured; each cell's sign follows its transmitter.

the neurons

Leaky integrate-and-fire units with spike-frequency adaptation – 12 ms membrane, 2 ms refractory, 4 / 8 ms excitatory / inhibitory currents, 0.5 ms steps – in real time on the server. Gap junctions are ohmic couplings. What the 3D and 2D views show is that computation, cell by cell, as it happens.

the animal

A Ciona larva is a chordate – sea squirts are our closest invertebrate relatives – and its 2 mm tadpole is the simplest chordate nervous system ever mapped completely. It hatches, swims towards light and then away from it, finds a rock and turns into a sessile adult. It never feeds. Its whole job is to sense light, gravity, pressure and touch, and swim.

how it learns

Each synapse keeps its anatomical strength times a learned multiplier. Pre-before-post spike pairs write potentiation, post-before-pre write depression, into an eligibility trace that fades in about a second. The trace only becomes a lasting change when a third factor arrives – reward for approaching light, scrolling up, escaping after touch or reaching a page it has never seen; penalty for hitting edges – amplified by the coronet cells' dopamine. Homeostatic scaling keeps rates physiological. Everything is saved to a persistent volume, so it keeps learning across restarts.

How the model completes the animal

The swim rhythm is modelled. A Ciona larva beats its tail at 20–30 Hz using membrane properties an electron microscope cannot record, so the alternating rhythm is a written generator. Whether a bout happens, how strong, how long and to which side all come from the connectome's own spiking – the larva decides, the generator only keeps time.
It sees light, not letters. Its visual world is the brightness around the cursor and whether the ocellus side is brighter – exactly what a larval eye reports. Thumbnails, videos and headlines reach it as light and shadow, and it learns which of them lead somewhere.
It learns by consequence. The plasticity site, the eligibility trace and the dopaminergic gate are real biology. In the sea a larva is rewarded by light and a place to settle; here new pages and bright regions play that role, so the same circuit learns the same way about a different world.
Two transmitters are modelled from the literature. Eminens cells are GABAergic and modelled excitatory, as larval GABA often is; the coronet cells' dopamine acts as a weak synapse plus the learning gate.
The click is its escape response. A ddN burst while its head touches a link is the same reflex a startled larva uses to dart away – here it carries it to the next page.
It always has a world. If the browser is ever unavailable, the same brain drives the same tail in a virtual tank with a light source, and this page says so.

How it got here

Connectome loaded

177 neurons, 3,105 chemical connections, 441 gap-junction couplings from the eLife Figure 16 source data; cell types from Figure 1, soma coordinates from Figure 3

DONE

Neurons running

Leaky integrate-and-fire with adaptation, transmitter signs, gap junctions, 0.5 ms steps in real time

DONE

Senses wired

Ocellus and canal photoreceptors with a shadow response, otolith tilt, coronet pressure, epidermal touch

DONE

Tail wired

MN1–MN5 driving the muscle bands through their measured neuromuscular synapses

DONE

Learning and memory

Eligibility-trace STDP gated by reward and dopamine; weights, episodes and log persisted to a volume

DONE

Eye and tail on a browser

A real Chromium sampled into the photoreceptors, cursor driven by the motor neurons, clicks from ddN bursts

DONE

Let loose on the feeds

TikTok, YouTube, Shorts, Reddit, Hacker News, BBC, the Guardian, Imgur, Wikipedia, arXiv – feeds it sinks through, videos it watches, thumbnails it startles at, links and buttons it bumps into

DONE

Released, observed live

Every spike streamed to every viewer with a narrated log. Nobody can nudge it, edit it, or steer it – including us, beyond the fence

DONE

Longer episodes, richer scoring

Scoring its browsing over hours rather than minutes so the learning curve reflects habits rather than moments

IN PROGRESS

the page under its cursor a real Chromium · cursor = MN1–MN5 · click = ddN burst

3D · every neuron and synapse white = spiking · lines = synapses carrying spikes

—

2D · schematic · dorsal view senses in · muscles out · tail 2.5× shorter

spike raster every cell · grouped by type · 2.5 s

population firing rates 30 s

muscle output left / right

learning curve score per episode

synaptic weights learned × anatomical

browsing pages / hour

its favourite pages chosen by its own visits and dwell time

its favourite sites time spent

learning over time weight drift · clicks per hour

activity log · narrated

A real nervous system, not a metaphor

the data

the neurons

the animal

how it learns

$TADPOLE

Where it has been

Follow the larva