Minecraft player builds a working AI chatbot in the sandbox from 445 thousand command blocks without a single mod

The Minecraft community has spent years building graphing calculators, QR code generators, and even Tetris inside the game using redstone and command blocks.

A player known as Objz took things further and built a working language model, using 445,782 command blocks with no mods, plugins, or datapacks involved. The creator summed up the project bluntly, calling it "a headache."

The model itself is tiny by modern standards. It has a 64-dimensional embedding space, a 256-neuron hidden layer, and a vocabulary of just 2,048 words, trained on 11,118 conversations from the DailyDialog dataset. Players can chat with it through the in-game "/dialog" command, and responses come out one word at a time, similar to how typical chatbots work.

Objz is upfront about the limits here – the AI can only hold a conversation. It can't do math, has no broad knowledge of the world, and isn't particularly sharp.

The real achievement lies in the optimization work. The original version, with the same capabilities, took up nearly 2 million blocks, and without some clever tricks, a data and computation structure of this size would genuinely need millions of cubes.

Language model weights are normally stored as floating-point numbers, which would be far too complicated to pull off in Minecraft. So Objz stuck to ternary values instead: -1, 0, and +1.

Minecraft "scoreboard" command handles multiplication and division, but only with integers. Converting between integers and floats would have added extra operations, and quantizing everything to -1/0/+1 solved two problems at once – cutting unnecessary commands and shrinking the dataset significantly. Thanks to all those zero-value weights, each multiply-accumulate operation averages out to just 0.67 commands.

One key detail: Objz didn't just round off the weights of a finished model after training. Quantization happened from the start during the forward pass, and a straight-through estimator handled updates to the underlying floating-point numbers during backpropagation. Combined with additional tweaks, this dropped the perplexity – essentially, the odds of the model spitting out a nonsensical word in a response – from 48.7 down to 38.8.

Because Minecraft caps how many commands can run at once, the model had to be split into smaller groups. Even with that workaround, generating a single word takes about 1.8 seconds on a server running at 35 ticks per second.

Objz noted that scaling this architecture up would get expensive fast from a computing standpoint – a 135-million-parameter language model built the same way would end up roughly 200 times larger than the current build.