Fruit Fly Brain Chess: How the Viral AI Works

By OpsMavix. Public project materials and research sources checked on 15 September 2026.

You move a pawn. An animated fly moves a piece back. A few turns later, you are taking the position seriously because losing your queen to something associated with the fruit bowl would be difficult to explain.

That is the appeal of fruit fly brain chess. Ernesto Lopez shared a playable chess experiment and reported that it had reached 700 Elo. The premise is immediately interesting: use a map of an insect’s nervous system as part of the machinery that chooses chess moves.

The opponent is software. Its biological connection comes from a connectome, a map of connections between neurons. Turning that map into a chess player requires a numerical model, a way to represent the board, and a way to turn the model’s response into a decision.

Follow those steps and the project becomes more interesting than its headline. You can see where the neuroscience ends, where the engineering begins, and what the result might help us investigate.

What is the fly chess website?

The viral demo is called Chess vs. a Fruit Fly. Its interface combines a chessboard with a view labelled as neural activity. The site’s explanation identifies the biological source as the male Drosophila melanogaster central nervous system, version 1.0, and says that training changes a linear readout of motor activity while leaving the underlying network fixed. Play the demo and read its explanation.

The demo’s public information endpoint reported 164,587 neurons, 8,865 sensory inputs and 2,129 descending and motor outputs when checked on 15 September 2026. These are figures reported by the running application. They should be distinguished from the size of the complete research dataset. Demo model information.

The central question is how activity in that network becomes useful for choosing a move. A list of neurons does not contain a chess interface, an understanding of checkmate or instructions for moving a knight. Those connections to the game have to be built.

What is a connectome, and where did the fly brain map come from?

A connectome records which neurons connect to which others. At synaptic resolution, it can describe the contact points through which neurons communicate, giving researchers a way to investigate possible routes through a nervous system.

Creating one involves reconstructing neural structures from microscopy images and checking the resulting map. The work behind a downloadable dataset includes imaging, computational reconstruction, annotation and substantial human verification. FlyWire’s whole-brain reconstruction paper.

Two resources are particularly relevant to the fly chess discussion.

FlyWire’s adult female brain map was described in Nature on 2 October 2024. The published reconstruction contains 139,255 neurons. It supports research into the organisation of a whole adult fly brain, including pathways between sensory inputs and outputs. Dorkenwald and colleagues, 2024.

The male CNS connectome covers the brain and ventral nerve cord. Google Research’s account, published on 3 September 2026, describes more than 166,000 neurons and 125 million synaptic connections. The project involved HHMI Janelia and collaborators, including Google Research and the Cambridge connectomics group. Google Research’s explanation, Janelia’s dataset and research credits.

A connectome gives a model an anatomical starting point. It does not, by itself, specify everything needed to reproduce a living brain’s changing activity. Researchers and developers still have to decide how signals behave, how inputs enter, and how outputs are interpreted.

Think of the difference between a road map and a traffic simulation. The map tells you where roads connect. To simulate traffic, you still need assumptions about vehicles, speeds, junctions and the journeys people are making. Different assumptions can produce different behaviour on the same map.

How can a fly brain model choose a chess move?

The easiest way to understand the engineering is to follow a position from the board to a decision. The following explanation describes the computational roles involved; it is not a claim that we have independently reproduced the viral demo’s training.

1. Represent the board as information

A chess position contains more than the locations of the pieces. Whose turn it is and which special moves remain available can affect what happens next. A model needs a numerical representation that preserves the information its designer wants it to use.

For a concrete example, the separately documented cesp99/fly-chess project uses 20 planes across an eight-by-eight board. Each plane records a different feature. The dimensions work out as eight rows × eight columns = 64 squares, then 20 planes × 64 squares = 1,280 input values. Its encoding includes piece positions, castling information and other game-state features. Board encoding implementation.

This is already a substantial design decision. An input representation can make useful information straightforward to extract or force the network to recover it through a more difficult route. The biological map does not choose that representation for the developer.

2. Produce activity in the network

The encoded input is connected to selected model units. Signals then pass through the network according to its mathematical update rules.

In the cesp99 implementation, those rules include recurrent connections and adjustable neuron parameters. A recurrent network can feed the result of one update into another, allowing activity to travel through connected units over successive steps. Network model implementation.

An activity pattern is an intermediate result. It becomes useful to a chess application when another part of the system can interpret it consistently. Bright colours in a visualisation may help a person follow the computation, but they do not explain the decision on their own.

3. Learn how to interpret that activity

A readout converts the network’s response into something the application can use, such as a score for a position. Training can adjust that conversion so that particular patterns of activity become associated with better or worse outcomes.

A fixed recurrent network with a trained readout is an established idea in reservoir computing. The recurrent network transforms the input; the readout learns to use the resulting features. That provides a useful conceptual comparison for the viral site’s stated approach. It does not establish that the demo reproduces every property of a particular reservoir model. Moon and Lu’s research on reservoir computing.

This explains how a system can learn a task even when its central network stays fixed. The learning takes place in the part that interprets the response. Keeping the middle unchanged does not mean the complete application contains no learned knowledge.

4. Use the scores to make a decision

The viral site’s explanation says it evaluates positions resulting from legal moves and ranks them through the trained readout. It also discloses two external assists: taking an available checkmate and avoiding a move that allows immediate checkmate. Demo explanation.

That division of work matters. A system can rely on a model for judgement while ordinary software handles rules and specific exceptions. Playing a legal move therefore does not demonstrate that a neural network independently discovered the rules of chess.

When assessing an AI demonstration, ask which component produced which part of the result. The input, model, readout, search procedure and rule checks can all contribute to what you see on the board.

Fixed wiring and a frozen network mean different things

Imagine a network as a collection of connected units. There are at least two separate choices: which connections exist and how strongly each connection influences the next unit.

Training can preserve the first while changing the second. The connection pattern stays the same, but information moves through it differently.

This is why the phrase “the wiring was unchanged” needs an explanation. It might mean the model kept the same connection pattern while learning new weights. It might mean the central network’s weights were also frozen and only an output layer learned. Those are different experiments.

The cesp99/fly-chess model explicitly contains trainable connection strengths, neuron biases and leak parameters, alongside its input and output machinery. Under its documented sign-constrained configuration, learning changes a connection’s magnitude without reversing its assigned sign.

Calling that process “nothing about chess is stored in the connections” would be misleading. Adjusting weights through chess training is one way a network acquires task-specific behaviour.

It is also common for ordinary neural networks to retain their connection pattern during training. The interesting constraint here is the use of a measured biological graph, together with the modelling choices around it. Fixed connectivity alone is not a new learning principle.

Carlo Esposito’s cesp99/fly-chess repository offers a detailed implementation for readers who want to inspect how this kind of experiment is built. Its documentation describes a FlyWire-derived graph with 134,209 neurons and 2,700,513 directed neuron-pair connections.

Keep its specifications separate from those of the viral website. Their published descriptions differ, and the repository should not be presented as verified source code for Ernesto Lopez’s demo.

Detail	Viral fly chess website	cesp99/fly-chess repository
Stated anatomical source	Male CNS connectome v1.0	FlyWire adult female brain connectome
Training approach	Site says the readout is trained	Documented model also trains internal parameters
Execution	Frontend requests moves from an API	Includes an engine that executes in a browser worker
What the source establishes	The demo’s published description and interface	An inspectable implementation with its own configurations

The site’s public frontend contains API requests for moves. The repository documents a separate browser implementation, including exported model files and numerical comparison tests. “You can play it in a browser” does not tell you where the computation happens. Viral site’s frontend, repository browser documentation.

The repository’s code licence names Carlo Esposito and uses MIT terms. Its licence file separately identifies the FlyWire-derived data and models under CC BY-NC 4.0. Project licence.

For developers, the useful exercise is to trace one board position through the encoding, network and output. That makes the contribution of each component much easier to understand than a single headline about a brain learning chess.

Does the fly really have a 700 Elo rating?

700 Elo is the figure reported by Ernesto Lopez in the post promoting the demo. We have not independently established an equivalent rating in a recognised player pool. Creator’s original post.

A chess rating describes performance relative to opponents and a rating system. It is not an absolute measurement of intelligence. Lichess explains that ratings from different servers and federations cannot be directly compared because their systems and player pools differ. Lichess on chess rating systems.

To interpret an experimental engine’s rating properly, you would want to know its opponents, number of games, time or computation limits, starting positions and scoring method. You would also need to identify the model version and any search or tactical assistance used during those games.

Beating a random-move opponent establishes a different result from performing consistently against rated human players. Winning a few games against a deliberately weakened engine also needs that configuration stated alongside the result.

The same care applies to move-prediction accuracy. Matching a reference move in a collection of positions measures something different from winning complete games. A player can find many ordinary moves correctly and still lose through one decisive mistake.

For now, read “700 Elo” as the creator’s reported strength estimate. You can enjoy testing the opponent without treating the number as a certified comparison with your own account rating.

Why the experiment is worth paying attention to

The interesting question is what a biological connection pattern contributes when it becomes part of a computer model.

A playable demonstration makes that question approachable. It gives people an action to take, a response to inspect and a reason to ask what happened between the two. It can also motivate more controlled experiments.

For example, a useful comparison could train otherwise similar systems using the biological graph and a randomised graph. The comparison would need to control details such as network size, number of connections, input representation, training data and computation budget. Otherwise, several changes could explain the difference in performance.

A second experiment could remove the connectome component and train a readout directly on the board features. If the resulting system performed similarly, that would change how much credit we should give the biological network.

A third could compare the same model with and without extra search or tactical assistance. That would help separate the contribution of the learned evaluation from the software surrounding it.

These are proposed tests, not results established by the demo. They show how to turn an interesting application into an answerable research question.

The project also makes the value of accessible research data tangible. Janelia provides tools for exploring the male CNS connectome and downloading its data. A resource created for neuroscience can become material for experiments that its original audience would never have encountered through a paper alone. Explore the male CNS dataset.

Try it yourself: follow one decision closely

Open the fly chess demo and play a normal opening. Develop a few pieces before attempting anything elaborate. The useful question is how its choices change when the position demands a specific response.

Try leaving a piece available to capture. Does it take the material, create a threat elsewhere or miss the opportunity? Then consider a position where the obvious capture has a drawback, such as allowing a stronger reply. Recognising an immediate gain and evaluating its consequences are different demands.

If the interface lets you revisit a position, change one move and compare the response. Pay attention to the consequences on the board as well as the activity display. A dramatic visual response does not necessarily correspond to a better decision.

A single game is an exploration, not a benchmark. For a more meaningful comparison, use several positions, play both colours where possible and record the conditions. That gives you something concrete to discuss beyond whether the fly won once.

Frequently asked questions

Is a real fly playing chess?

The opponent is a computer system built around a model associated with a fly connectome. The biological source provides anatomical data. The game interface, numerical simulation and connection to chess are engineered software.

Does this mean someone uploaded a fly’s mind?

A wiring map and a functioning software model do not establish that an animal’s memories, subjective experience or complete brain dynamics have been reproduced. That would require evidence far beyond a model selecting chess moves.

Is the fly chess project the same thing as FlyWire?

FlyWire is a neuroscience reconstruction project and data resource. The viral chess website identifies the male CNS dataset, while the cesp99 repository uses FlyWire. The research maps and the applications built around them should receive separate attribution.

Can it beat Stockfish?

The sources reviewed here do not establish that the viral demo can beat full-strength Stockfish. Any such comparison would need the engine version, settings, computational limits and match results. A creator-reported rating of 700 does not establish that claim.

Does an unchanged connectome mean nothing was trained?

Training can happen in an output readout even when the central network remains fixed. In another design, the connection pattern can stay unchanged while connection strengths learn. Check which parameters the specific implementation allows training to alter.

Does this prove the same model can learn any task?

A successful result on chess would not establish that. Another task can demand different inputs, outputs, memory and training. Generality has to be demonstrated across tasks, and a biological graph’s contribution needs to be compared against suitable alternatives.

Play the fly, then explore the engineering

At OpsMavix, we build operations software and take an interest in how systems turn information into useful decisions. This experiment offers an accessible reason to examine that process closely, from the source data to the behaviour someone can interact with.

Start with Ernesto Lopez’s chess demo. Follow the research behind its stated connectome, then explore Carlo Esposito’s related open source implementation if you want to study the code. There is plenty to investigate after the first captured pawn.