StratoBuilds’ AI Fish Tank Packs a 14 Million-Parameter Fish Brain in Your Pocket

Real water and live fish refuse to share a pocket without making a mess. Strato Doumanis, who posts as StratoBuilds, built Pocket Tank instead: a 1.8 inch AMOLED aquarium you can hold in one hand, where every high level choice a fish makes comes from a language model that lives entirely on the board.

A new tank is opened, and it contains only two small fish (fry). You enter the bubble column, give them names, and select a body color. As the day progresses, the days of “active time” propel them through the various stages of juvenile, adult, and elder, and you will see the variations in their appearance. Adding extra fish to the mix is simple; you can add one at a time up to a maximum of five. As long as the adults drifting around near the bottom of the grass remain content, you may anticipate a new arrival in the tank the next time the lights turn on. Each of your fish will have a few stats that fluctuate according on how you treat them, such as hunger, energy, tension, curiosity, and boredom. Plus three attributes that float around depending on how you care for them: boldness, sociability, and sloth. Then there’s something called trust, which is a particular score that indicates how much they trust you. Stick your finger on the glass for around three seconds, and the trusting fish will come over to say hello. Tap the glass three times quickly, and all of the fish will scatter. Tap a fish on the glass to bring up its little information card, and tap again to see what milestones it has reached.

Feeding time is as simple as touching the surface or drawing a finger down from the top edge. Algae begins to grow on the glass; simply remove it with your finger. There is, however, a persistent problem with sea grass sprouting, which you can clip down to finger height with a sideways swipe, and if you leave some of it standing, it provides a good hiding area for the fish to perform their courtship rituals. Need to turn the lights on or off? Just double-tap. When you press the power button for a short period of time, the tank will simply go to sleep and the screen will turn dark, causing the metabolism to drop significantly.

Press and hold the button, and the entire system will decrease to tens of microamps. One of the things that will survive a complete power outage is the history of your fish, including the names you’ve given them, decorations, and so on. After a good long rest, the fish will gather around the customary feeding spot and wait for you to drop a pellet in to begin the day. If you’re going to be away for a while and one of the fish becomes hungry, an automatic feeder can help. You can earn sand dollars by caring for your fish, which you can then spend on items such as a sword plant, an algae grazing snail, a castle for the fish to swim through, some colorable coral, and a reef cluster that you can install in your tank.

All of this runs on a Waveshare ESP32-S3 Touch AMOLED 1.8″ display, which costs around $35 and includes the screen, battery, and a lovely tiny casing. A dual core processor, 16MB of flash, and 8MB of PSRAM are included, as well as a 368×448 display screen, a motion sensor, a real-time clock, a speaker, and power management, all housed in a small plastic shell. The standard code performs all of the swimming, steering, schooling, plants, and touch interface operations at a steady 25 to 30 frames per second. The second core then runs the model, making decisions and so on, which takes approximately 3.7 seconds. All of the fish have the same brain, which asks itself if it needs to do anything more frequently than necessary in order to conserve energy.

That brain, a 14.3 million parameter student with eight layers and eight heads each, was able to be compressed to a modest 7.56 MB in 4 bit form, which is quite an astounding effort. Doumanis spent a lot of time developing over 51,000 tank situations on their trusty old Mac Mini before handing over the data to Gemma 4 to assist label each one with a specific aim and urgency. The student then trained on all of those pairs of data, resulting in a 54-token vocabulary that is unique to this environment.

When a state arrives, think of it as a status update, as it’s simply one line with some basic information: what zone the fish is in, how hungry it is, its energy levels, the closeness of nearby food or pals, its personality, what stage of life it’s in, and how much trust it has in its surroundings. The model then concludes the sentence with something like “seek_food with a sense of urgency of 8.” It also has a list of approved goals to which it can respond, such as grabbing some food, following a friend, getting a closer look at the reef, checking out the bubbles, exploring, resting, or simply thrashing around in play.

One nice touch is that when the model’s top two responses are close together, the animation depicts some uncertainty before the fish chooses between them. When it came to testing the student’s abilities, it did an excellent job on a small test set, matching what it was taught nearly three-quarters of the time. The best part is that there is no need for Wi-Fi or online storage, and no data leaves the board because it is self-contained.
[Source]