The Minecraft community has long mastered building graphical calculators, QR code generators, and even Tetris inside the game using redstone and command blocks. A gamer known by the nickname Objz went further and implemented a full‑fledged language model, using 445,782 command blocks and without resorting to mods, plugins, or data packs. The creator himself described the work briefly, calling it a headache.

The model turned out to be miniature by modern standards. It has a 64‑dimensional embedding space, a hidden layer with 256 neurons, and a vocabulary of only 2,048 words, and it was trained on 11,118 dialogues from the DailyDialog dataset. You can communicate with the chatbot using the /dialog game command, and the response is displayed one word at a time, just like in conventional language models. At the same time, Objz honestly warns that the neural network can only maintain a conversation: it is incapable of calculations, does not have broad knowledge about the world, and is not particularly quick‑witted.

The main achievement here lies precisely in the optimization. The original version with the same capabilities occupied almost two million blocks, and without any tricks, the structure for such a volume of data and calculations would indeed have required millions of cubes. Weights in language models are usually stored as floating‑point numbers, which would be too difficult to implement in Minecraft, so the author limited themselves to ternary values: minus one, zero, and plus one. The scoreboard command in Minecraft supports multiplication and division, but it only works with integers. Quantizing to ternary values solved two problems at once: it eliminated unnecessary commands and significantly reduced the data volume. Thanks to the abundance of zero weights, each multiplication operation with accumulation costs an average of sixty‑seven hundredths of a command.

An important detail is that Objz did not simply round the weights of the finished model after training. Quantization was applied from the very beginning during the forward pass, and the direct estimator method was used to update the original floating‑point numbers during backpropagation. Together with additional optimizations, this reduced perplexity – that is, the probability of producing a meaningless word in the response sequence – from 48.7 to 38.8. Due to the game’s limitations on the number of commands that can be executed simultaneously, the model is divided into small groups. Even so, on a server with a tick rate of thirty‑five per second, generating one word takes about one and eight tenths of a second. The creator notes that scaling such an architecture is extremely costly from a computational perspective: a language model with one hundred and thirty‑five million parameters would be about two hundred times larger than the current setup.