About this project

Every fish in this desk aquarium asks a small transformer what to do. Once every 1.5 seconds — or immediately when food or a finger arrives — the model reads a 26-word line describing that fish's species, hunger, energy, stress and surroundings, and answers with one goal and one mood, such as seek_food eager. Nothing second-guesses the answer. The network has two layers, two heads and a 54-word vocabulary: 29,952 learned numbers stored at 8 bits, 30,522 bytes in total. A decision costs about 31 ms; caching recent answers and keeping weights in internal RAM holds the tank at 16 to 17 frames a second. Training is Python and PyTorch, distilling qwen3.5:4b over 12,000 situations. The board is a Hosyond/LCDWIKI ES3C28P with a 2.8-inch 320x240 screen, and an SDL2 simulator runs the same code on a Mac.