- Sammyuri built CraftGPT, a redstone-powered chatbot running entirely in Minecraft, trained externally on TinyChat.
- The build spans a 1020×260×1656 volume (≈439M-block footprint) and relies on vanilla redstone; Distant Horizons was used for filming.
- Specs include ~5.09M parameters, 240-dim embeddings, 1,920-token vocab, 6 layers, and a 64-token context window with mostly 8-bit weights.
- Even with MCHPRS boosting tick rate ~40,000×, replies can take about two hours; quality is limited and often goes off-topic.
For anyone who has ever wondered whether a chatbot could exist entirely inside Minecraft, the answer is now a quirky yes. A creator known as sammyuri has unveiled CraftGPT, a fully in-game inference setup that responds to prompts like a miniature ChatGPT, wired together with vanilla redstone components.
This isn’t a replacement for your everyday AI tools; it’s more like a technical art piece that proves a point. The machine’s footprint stretches to a vast in-world volume and, even when sped up, can take about two hours per reply. Still, as a feat of digital craftsmanship, it shows how far redstone logic can be pushed when patience and engineering collide.
Who built CraftGPT and what exactly is it?
CraftGPT comes from YouTuber and redstone engineer sammyuri, known for ambitious in-game computing projects. It’s a small language model that runs inference on a massive redstone computer, while the training happened outside the game in Python on the TinyChat dataset of simple English conversations.
The structure occupies a volume of roughly 1020 × 260 × 1656 blocks, a scope that required the Distant Horizons mod to capture wide shots. Despite the cinematic footage, the circuits themselves rely on vanilla redstone mechanics—no command blocks and no datapacks.
The creator cautions would-be testers to keep expectations modest. Outputs can be basic, sometimes off-target, and occasionally nonsensical, which is expected from a compact model wired into a world where every logical step travels through repeaters and comparators.
Even so, the point isn’t raw performance; it’s a working demonstration that LLM-style inference can be remapped to an unusual substrate inside a game, producing a chatbot you can literally walk around and inspect block by block.
How the redstone chatbot actually works
Under the hood, CraftGPT recreates the inference pathway of a transformer-like model through orchestrated redstone logic. You’ll find modules that mirror familiar AI building blocks—carefully wired circuits handle tokenization, embeddings, and the arithmetic behind matrix multiplications.
The system steps through a pipeline: input text is transformed into tokens, those tokens are embedded into vectors, and the redstone computer iterates across layers to produce the next token. The process is slow because each operation must propagate through long chains of signals.
- A tokenizer-like unit turns text into sequences the machine can process.
- Embedding arrays and lookup structures translate tokens into vector representations.
- Matrix multiplier arrays and associated logic perform the heavy-lifting computations.
- A KV-style cache supports reuse of intermediate results to save cycles where possible.
- Output logic converts computed tokens back into readable text.
In the showcase, sammyuri demonstrates brief exchanges—simple pleasantries, basic factual prompts—showing that the machine can produce coherent responses, albeit with limited depth and accuracy compared to mainstream models.
Alongside these modules, the build simulates memory and synchronization components, coordinating timings so signals don’t collide or decay, a recurring challenge in large-scale redstone contraptions.
Specs, performance, and limits
CraftGPT’s model weighs in at around 5,087,280 parameters, with 240-dimensional embeddings, a 1,920-token vocabulary, and six layers. Most weights are quantized to 8 bits, while the embedding and LayerNorm weights use higher precision (18 and 24 bits, respectively).
The usable context window is just 64 tokens, enough for very short exchanges. That helps keep the hardware footprint manageable but naturally constrains conversation flow and coherence.
As for speed, the setup relies on the Minecraft High Performance Redstone Server (MCHPRS) to boost tick rate to roughly 40,000×. Even then, a single answer often needs on the order of two hours. Without specialized acceleration, community estimates suggest it would take many years to generate a reply.
Visually, the machine’s massive volume—about 439 million blocks of space—necessitated Distant Horizons for filming, which renders distant circuitry at a lower detail level. Functionally, though, the redstone logic itself remains pure vanilla.
The creator repeatedly notes that results may drift off-topic or contain grammatical mistakes. The showcase clips emphasize best-case output, not the average behavior you should expect when you start experimenting with prompts.
Setup, availability, and community reaction
The world is available on GitHub for those who want to explore the circuits or attempt their own chats or learn to install AIML in Python. You’ll need horsepower: the developer recommends at least 32 GB of RAM just to load the server, with 64 GB or more preferred for smoother experiments.
CraftGPT sticks to vanilla redstone for logic and uses MCHPRS only to make the system tolerably responsive. The training itself happened outside the game, so Minecraft is hosting inference, not learning—an important distinction that keeps the build grounded in reality.
Reactions from the Minecraft and AI communities mix awe with pragmatism. Fans celebrate it as one of the most ambitious redstone builds ever released; skeptics point out it’s a spectacle rather than a practical assistant. Both can be true, and together they highlight Minecraft’s role as a playground for computing ideas.
Past redstone milestones include custom CPUs, memory banks, and even projects that rendered classic games in-block. CraftGPT pushes that lineage into LLM territory, offering educators, hobbyists, and the curious a hands-on way to discuss what models do—right down to the signals that carry the math.
What lingers after watching CraftGPT work is a reminder that computation is universal: once you have the right primitives and patience, logic can be rebuilt in surprising places. This project blends game tinkering with computer science, showing that an AI-like conversation can live inside a blocky sandbox—just not at real-time speeds.
