Zebrafish Neural

A zebrafish-inspired neural controller. Watch the model run. Building toward a recorded token launch on Pons.

[Zebrafish Neural @ZebrafishNeural]

Development continues regardless of $ZNEURO’s market cap.

We’re one developer and one researcher working on zebrafish-inspired control, browser experiments and anatomy visualization.

FLYBRAIN helped put this field on our radar. Our code and methods are public.

[Zebrafish Neural @ZebrafishNeural]

Reaching a target is a starting point.

Next, we want to change its speed, add distractions and test patterns the model hasn’t trained on.

Can it use what it learned when the environment changes? That’s the question we want the next experiments to answer.

[Zebrafish Neural @ZebrafishNeural]

What changes if we remove the fish’s visual inputs?

We want to test the same chess positions with normal inputs, no inputs, and random move selection.

That comparison can show where the controller makes a difference—and where it doesn’t.

[Zebrafish Neural @ZebrafishNeural]

Migration complete. The fish is running on AWS, with its game history intact.

Next up: YOU vs THE FISH 🐟♟️

Play against the model and inspect the neural activity behind its replies.

First playable build in 24 hours.

https://t.co/gloHuUmpru

[Zebrafish Neural @ZebrafishNeural]

We are migrating the chess backend to our AWS server. ETA 20-30 mins

[Zebrafish Neural @ZebrafishNeural]

We are migrating the chess backend to our AWS server. ETA 20-30 mins

[Zebrafish Neural @ZebrafishNeural]

What does the fish look for in a chess move?

A capture. A safer square. Control of the centre. A check.

We turn those four features into inputs, then use the neural controller’s responses to choose between legal moves.

https://t.co/gloHuUmpru

[Zebrafish Neural @ZebrafishNeural]

You’ll be able to play against the fish soon. 🐟♟️

For noweveryone watches the same live game. If you want to build your own app around the model, the code is public:

https://t.co/IBYfp0IVk4

[Zebrafish Neural @ZebrafishNeural]

Watch the fish play chess. 🐟♟️

Our model picks its moves against a bot. Watch the board and its neural activity side by side.

One live game for everyone. When it ends, the next one starts,model learns from each game.

https://t.co/gloHuUmpru

More experiments to come.

[Zebrafish Neural @ZebrafishNeural]

https://t.co/IBYfp0IVk4

[Zebrafish Neural @ZebrafishNeural]

Watch the fish play chess. 🐟♟️

Our model picks its moves against a bot. Watch the board and its neural activity side by side.

One live game for everyone. When it ends, the next one starts,model learns from each game.

https://t.co/gloHuUmpru

More experiments to come.

[Zebrafish Neural @ZebrafishNeural]

A preview of the chess interface

The goal is to show each move alongside visual input and the model’s left/right activity, making it clear what came from the controller model and what came from the chess engine.

[Zebrafish Neural @ZebrafishNeural]

We’re aiming to add the chessboard experiment within the next 8 hours, connecting our zebrafish controller to a chess environment with a separate game engine.

The code will go on GitHub too, so anyone can inspect how the controller, board and engine connect.

[Zebrafish Neural @ZebrafishNeural]

A chessboard is one of the next stops for our fish today.

We’re planning a task kit with browser challenges, puzzles and a chess prototype using a separate game engine. Each attempt will show the moves and results.

Concept preview below.
https://t.co/q9C5Ft5XH3

[Zebrafish Neural @ZebrafishNeural]

The 3D fish will become an interface you can interact with.

Give it a task, see the screen it receives, and follow its actions alongside the model activity.

We're starting by bringing our target-reaching experiments into this scene.

[Zebrafish Neural @ZebrafishNeural]

We're building a 3D body for the Zebrafish Neural controller.

The first local prototype links visual input, model activity and fish movement. Next, we'll connect it to task environments so you can watch each attempt and its outcome.

[Zebrafish Neural @ZebrafishNeural]

We're building a 3D body for the Zebrafish Neural controller.

The first local prototype links visual input, model activity and fish movement. Next, we'll connect it to task environments so you can watch each attempt and its outcome.

[Zebrafish Neural @ZebrafishNeural]

Challenge the Fish is live.

Place a target, move it or hide it. Watch the original and trained versions of our controller attempt the same task.

Compare their paths and times, then share the challenge for others to repeat.

https://t.co/Hm9zF8eFdJ
https://t.co/xRhjxnsoVU

[Zebrafish Neural @ZebrafishNeural]

Zebrafish adapt their swimming commands when visual feedback changes.

Our software learner turns visual input into movement, scores target-reaching attempts and saves changes to six controller gains.

Research context:
https://t.co/jr0YJLX1Yq

[Zebrafish Neural @ZebrafishNeural]

Next up: Challenge the Fish.

You'll be able to place a target, change its movement or briefly hide it, then watch the controller try to reach it.

Run the same task with the original and learned parameters. See which reaches the target, how long it takes, and share the challenge so others can repeat it.

We're aiming to put it live in the next 4–6 hours.

https://t.co/6VY4E3S45N

[Zebrafish Neural @ZebrafishNeural]

Migration complete. Persistent learning is now live on Zebrafish Neural.

We moved the saved run to AWS at generation 148. It resumed from that checkpoint and has continued training.

Here’s what happens in each round:

The controller receives synthetic visual input and attempts to reach a target. Its attempts are scored, and training adjusts six connection gains from those outcomes. A candidate replaces the retained controller only after passing validation checks.

Every completed generation saves its parameters, training state and results. Training continues even when nobody has the website open.

Every ten generations, we evaluate the controller on fresh targets that are not used to select its parameters. The page shows saved gains, evaluation results and labeled replays. The JSON records are downloadable.

This is a separate learning experiment alongside the Wikipedia browser controller.

Follow the experiment:
https://t.co/RAfLo52GwL

Training method, source code and tests:
https://t.co/kI8Jf4XgfH

[Zebrafish Neural @ZebrafishNeural]

Migration has started ETA 15 mins.

[Zebrafish Neural @ZebrafishNeural]

Migration has started ETA 15 mins.

[Zebrafish Neural @ZebrafishNeural]

Our AWS server is ready. We’re preparing to move the controller, saved weights and experiment records over.

We expect the migration to take around 15 mins once it starts. During the switch, the website’s live panels may briefly freeze or reconnect. That’s expected during the handover.

Once the server checks pass, we’ll enable the public target-learning experiment: a controller that adjusts its connection weights through scored trials and retains them between runs.

We’ll post another update when the migration is complete.

[Zebrafish Neural @ZebrafishNeural]

The dedicated server is still awaiting building, so the learning preview and controller haven’t been migrated yet.

We’ll begin migration as soon as the server is ready, preserve the saved state and experiment records, and check the feed before switching over. We’ll post an update when the new setup is live.