Working 5M parameter ChatGPT AI model in Minecraft, made with 439M blocks (www.tomshardware.com)

Minecrafter and engineer Sammyuri released CraftGPT, a working small-language model implemented entirely inside Minecraft using Redstone logic. The in-game contraption occupies 1,020 × 260 × 1,656 blocks (about 439 million blocks) and implements tokenizers, matrix multipliers and other LLM primitives without command blocks or data packs. The model itself has 5,087,280 parameters, a 240-dimension embedding, 1,920-token vocabulary, 6 layers and a 64-token context window; it was trained in Python on the TinyChat conversational dataset. Most weights are quantized to 8 bits, with embedding and LayerNorm weights stored at 18 and 24 bits respectively. CraftGPT is a proof-of-concept and spectacle more than a practical chatbot: even running at an accelerated tick rate using MCHPRS (~40,000×), a single reply typically takes about two hours and outputs are often off-topic or ungrammatical. Its significance lies in demonstrating how LLM building blocks—tokenization, matrix ops, quantization and layer norms—map onto extremely constrained, alternative compute substrates. For AI/ML practitioners the project is a vivid, tangible educational tool about resource trade-offs, latency, and quantization effects, but it also underlines why conventional hardware remains essential for usable models: compute density and throughput, not just logical correctness, determine practical LLM performance.