SCIENCE
A CONNECTOME, NOT A MIND.
Human vs Brain uses MaleCNS v1.0, a reconstruction of the adult male fruit fly central nervous system containing approximately 166,700 neurons.
The biological wiring is real. The activity you see in the game is a computational simulation built on top of that wiring.
This is not a digital copy of a living fly. It is not a model of consciousness, and it does not mean a biological fly understands the game.
What we are testing is much narrower:Can a simple stimulus travel through a network constrained by real biological connectivity and produce a meaningful left or right response?
THE EXPERIMENT
How one round works
You and the connectome receive the same left or right target.
For the connectome, that target is translated into stimulation of LC10a visual projection neurons on the corresponding side.
LC10a neurons carry visual information about small moving targets. Experiments in living flies have shown that LC10a activity is associated with tracking, and that unilateral manipulation of LC10a can influence turning toward a target.
The stimulus then enters our MaleCNS simulation.
Activity propagates through the mapped network according to a simplified neural model. We measure the resulting activity in the left and right DNa02 descending neurons.
DNa02 is strongly associated with steering. Recordings from both sides of the nervous system show that the difference between left and right DNa02 firing predicts the direction and magnitude of turning. Experiments also show that unilateral activation can bias steering toward the activated side.
Whichever side produces the stronger DNa02 response becomes the connectome's choice. Spike count is the primary readout; an exact spike-count tie uses mean simulated subthreshold voltage, with an exact voltage tie resolved by the fixed LEFT convention.
- Target
- LC10a input
- MaleCNS simulation
- DNa02 output
- Left or Right
There is no LLM choosing the answer. There is no trained game-playing policy, hidden classifier or learned readout deciding which side wins.
THE MODEL
Where the simulation begins
A connectome tells us which neurons connect to one another.
It does not tell us the complete electrical and chemical state of a living brain.
To make the network run, Human vs Brain therefore adds a simplified model of neural activity.
We use leaky integrate-and-fire neurons, normalized connection strengths and predicted neurotransmitter signs. These are computational assumptions applied to the biological wiring.
We also simplify the sensory system.
The game does not recreate the fly's eyes, retina, lamina, motion-detection circuitry or physical body. Instead, a left or right target directly stimulates the corresponding LC10a population.
Likewise, the game does not simulate muscles or legs. DNa02 activity is used as the readout for the model's steering choice.
This distinction matters.The wiring comes from biology. The neural dynamics are simulated. The experiment is designed by us.
THE CHALLENGE
Why the game can get it wrong
The public challenge uses a weaker LC10a input than our scientific reference configuration.
That is deliberate.
A very strong stimulus makes the left-versus-right distinction too easy for this particular model. The public game uses a stimulus strength of 0.27 model voltage units per simulation step, compared with 0.80 in the reference configuration.
These are model parameters. They are not measured biological stimulus intensities.
The weaker public setting makes activity spreading through the network less decisive. Sometimes the stronger DNa02 response appears on the wrong side.
Validation across the current 500-trial aggregate found a 6.4 percentage-point residual LEFT action advantage. We leave that behaviour visible rather than artificially correcting the result.
PERSISTENT LEARNING
How the Decision Game learns
The Decision Game has one shared global brain. Before public live learning, that lineage completed a separate 10,000-trial historical validation run. Historical validation is bounded and separate from live gameplay; it does not replay as public activity.
Each successfully completed authoritative simulation can produce one backend learning event. The learning label is objective: the brain is correct only when its LEFT/RIGHT decision matches the server-authoritative target.
Plasticity is restricted to a fixed whitelist of existing synapses on the LC10a-to-DNa02 pathway. Updates are small and bounded, connection signs stay fixed, new edges cannot appear, and the connectome topology, sensory input, output neurons and decision readout remain unchanged.
After an accepted event commits, future players use the latest brain version. This is global cumulative learning, not personal adaptation: the system does not use a player's answer, accuracy or reaction time as a training label.
Backend simulation failures, backend timeouts, rejected events and duplicate events do not update the brain.
GAME RESPONSE TIMEThe actual model runtime used by the Decision Game for the Human vs Brain competition. This is the game's normal BRAIN TIME metric.NEURAL DECISION LATENCYA simulated neural commitment metric used for scientific analysis of the model, not the game's competitive response time.Persistent learning is not evidence of consciousness, general intelligence or a biologically complete learning process.
THE CLAIM
What does the score actually mean?
If the connectome gets 9 out of 10 rounds correct, that does not mean:
The fly understands the game.
It means:
Under this particular sensory input, neural model and output readout, activity travelling through the MaleCNS connectivity produced the expected steering response nine times out of ten.
That is a smaller claim.
It is also the scientifically interesting one.
THE CHOICE
Why LC10a and DNa02?
The choice was not arbitrary.
LC10a visual projection neurons respond to small moving targets and form part of a visuomotor pathway involved in tracking. Experiments have shown that manipulating LC10a on one side can alter ipsilateral turning behaviour.
DNa02 is a descending neuron involved in steering. Bilateral recordings show that differences in left and right DNa02 activity predict turning direction and magnitude, while unilateral activation can bias turning toward the activated side.
Together, they give us a biologically motivated input and output for a simple lateralized experiment.
They do not mean every modelling choice between them has been experimentally validated.
A USEFUL DISTINCTION
The simplest way to think about it
Connectome
The biological wiring.
Simulation
Rules that allow activity to move through that wiring.
Experiment
The way we turn a game stimulus into neural input and neural output back into a choice.
Human vs Brain sits at the intersection of all three.Not a mind in a browser.
Want to see the anatomy behind it?
SOURCES
Sources and attribution
- Berg et al. (2026), “Sexual dimorphism in the complete Drosophila male central nervous system connectome”Cell 189, 5504–5526.e15 · DOI: 10.1016/j.cell.2026.08.015
- Official MaleCNS v1.0 datasetJanelia Research Campus
- Ribeiro et al. (2018), “Visual Projection Neurons Mediating Directed Courtship in Drosophila”Cell 174, 607–621.e18
- Hindmarsh Sten et al. (2021), “Sexual arousal gates visual processing during Drosophila courtship”Nature 595, 549–553 · DOI: 10.1038/s41586-021-03714-w
- Rayshubskiy et al. (2025), “Neural circuit mechanisms for steering control in walking Drosophila”eLife 13, RP102230 · DOI: 10.7554/eLife.102230
- MaleCNS data is licensed under CC BY 4.0.
MaleCNS attribution: FlyEM project at HHMI Janelia, the University of Cambridge, the MRC Laboratory of Molecular Biology, Google Research and the MaleCNS authors. Human vs Brain does not redistribute the raw or full MaleCNS dataset; it serves derived visualization assets produced from that source for this interactive view.