Over the weekend, a Minecraft maestro performed a striking feat: embedding a working language-model inference engine inside the immensely popular sandbox game. The catch? All of the “wiring” was done with Minecraft’s redstone system—no command blocks, no mods (beyond rendering assistance), just circuit logic built from in-game materials.
The result is quirky, painfully slow, but fascinating: a GPT-style chatbot inside a block world, built from digital redstone and analog logic.
To understand the stunt, you need to know what redstone is. In Minecraft, redstone is the game’s analog to wiring and electronic components: power sources, repeaters, comparators, logic gates, signal delays—all built with blocks. Redstone circuits emulate digital behavior (on/off, logic operations) inside the game world.
Players long ago built calculators, memory units, and even CPUs inside Minecraft using redstone—and now they’re pushing it into AI territory. Sammyuri, the person who build the GPT hack, even built a recursive version of Minecraft within Minecraft using redstone.
In this project, the creator first trained a small language model outside of the game (in Python) using a dataset called TinyChat. That model has roughly 5,087,280 parameters, an embedding dimension of 240, a vocabulary of about 1,920 tokens, six layers, and a context window of 64 tokens. Most weights are quantized to 8 bits, though embedding and LayerNorm weights use higher precision. The redstone build itself spans roughly 1,020 × 260 × 1,656 blocks (≈ 439 million blocks in total). To film the scale, the creator used the Distant Horizons mod, which allows distant structures to be visualized in a giant world.
When you input a prompt (via in-game interface), the redstone circuits carry out the inference step by step, embedding lookup, feedforward passes, matrix multiplications, and softmax approximations. According to a video demonstration, the elaborate redstone build took months to assemble.
But the system is glacial. Even with an artificially boosted tick rate (~40,000× faster than normal, via MCHPRS), the response time is on the order of a couple of hours. At standard Minecraft speed, some estimate it would take over nine years to generate an answer.
Still, the point of the project isn’t utility—it’s demonstration. At its heart, this build is a playful proof that neural inference can be recast into almost any logical substrate. It reminds us how much abstraction our software and hardware stacks hide: that architecture, medium, and speed are separate dimensions. In other words, this is a technical art piece, a conversation starter: what counts is computation, not necessarily where it runs.
Does this matter? Yes, it does
Obviously, this isn’t practical, but it’s cool because it demonstrates something profound: that logic and neural-style computation can be mapped into bizarre substrates—here, virtual redstone circuits. It’s a playful proof of the universality of computation, an artistic and educational showpiece, and a challenge to our assumptions about how and where “intelligence” can run.
It forces us to ask: what really matters in a model—architecture, medium, speed—and what other strange substrates might one try (optics, DNA, mechanical systems)?
Across the web, the build has ignited debates. Tom’s Hardware ran a piece breaking down the block count, the redstone logic, and the performance tradeoffs. On Hacker News, commenters marveled at the engineering but also pressed on limitations: “At normal redstone tick rate… it would take just over 9 years for a response.” Meanwhile, in the Minecraft subreddit, fans and skeptics alike debated how much of the achievement is spectacle vs. technical depth.
Some observers suggest the project borders on meme more than research. Indeed, the real training happened externally; Minecraft only hosts inference logic. It’s a showpiece more than a practical model. But that’s precisely what gives it charm—and purpose.
