tbh great vid also funny editing for some reason it looked so satisfying..the leap onto the enter key etc. besides well done on the project, presentation is classy :D looks a lot of fun!
there were no clickbaits there at all. no command blocks were used at all. if you were so certain, why dont you download the world and try it yourself?
Well, TinyChat. Still mind-boggling. From the video description:
I built a small language model in Minecraft using no command blocks or datapacks!
The model has 5,087,280 parameters, trained in Python on the TinyChat dataset of basic English conversations. It has an embedding dimension of 240, vocabulary of 1920 tokens, and consists of 6 layers. The context window size is 64 tokens, which is enough for (very) short conversations. Most weights were quantized to 8 bits, although the embedding and LayerNorm weights are stored at 18 and 24 bits respectively. The quantized weights are linked below; they are split into hundreds of files corresponding to the separate sections of ROM in the build.
The build occupies a volume of 1020x260x1656 blocks. Due to its immense size, the Distant Horizons mod was used to capture footage of the whole build; this results in distant redstone components looking strange as they are being rendered at a lower level of detail.
It can produce a response in about 2 hours when the tick rate is increased using MCHPRS (Minecraft High Performance Redstone Server) to about 40,000x speed.
This is outrageous! I see people build some basic circuits and logic gates using red stone, but this is unimaginable. Is Minecraft becoming a digital clone of the real world?
Are these kind of projects actually build manually inside of minecraft block by block or is there some verilog/vhdl to minecraft-level compiler toolchains used?
Aside from the technical feat, which is amazing in and of itself, I was completely mesmerized by that video! So well done, incredibly entertaining and perfectly presented.
Here is an earlier, longer video on someone explaining how they built a neural network to recognise handwritten digits in Minecraft: https://www.youtube.com/watch?v=DQ0lCm0J3PM
It should answer some of the questions/clear up some of the confusion raised here (e.g. how they get the weights in).
Funny, I bought that book and studied it because I heard about it in a Minecraft video. Was that you? Nand2tetris helped me land my first programming job.
> (which I don't think existed at the time anyways)
This surprised me at first because I remember using Python to create Minecraft maps early on, or at least in the beta.
But it seems like redstone was added in the alpha, and the earliest commits of pymclevel (which I think I used) also dates back to the alpha. So there might indeed have been a time window of a few months in which redstone was available but not tooling for creating maps.
Any system that can represent a NAND gate is Turing complete. In Minecraft, connecting two inputs to the same branch of redstone wire is an OR gate, and a redstone torch on a block acts as a NOT gate. Two inputs, both inverted, into an OR gate gives a NAND.
when people do these complex CPU designs in minecraft are they laying the blocks individually in real time and in 3d space - or are they scripting some sort of algorithm that instantiates the system in one go?
It's impressive either way but the manual version seems ... impossible.
Can‘t wait to test out the world download, this is so cool and also scary how much time this must have taken! Did you build it with some schematic editor?
If you like thinking about the possibilities of the worlds of Minecraft and LLMs colliding I recommend the YouTube channel @EmergentGarden (https://www.youtube.com/@EmergentGarden) it's not exclusively agents and Minecraft, but there's a fair amount of stuff tested over time.
38 comments
[ 2.8 ms ] story [ 56.6 ms ] threadOh. I thought this would be some cheesy command block curl to Chat GPT.
Surely this wasn't all done in a block editor by hand?
There was some code shown at beginning. Was that placing the blocks to build each section?
And, more info on the system that could run this.
Bravo to you, sir.
But no. The author actually embedded a small LLM in Minecraft using hundreds of millions of blocks, that generates 1 token per 2h at 40,000x speed.
Bravo. I wouldn’t have even thought to try.
It should answer some of the questions/clear up some of the confusion raised here (e.g. how they get the weights in).
4-bit is small enough that you can build it manually, without the use of external tools (which I don't think existed at the time anyways).
Highly recommended for children interesting in computing!
This surprised me at first because I remember using Python to create Minecraft maps early on, or at least in the beta.
But it seems like redstone was added in the alpha, and the earliest commits of pymclevel (which I think I used) also dates back to the alpha. So there might indeed have been a time window of a few months in which redstone was available but not tooling for creating maps.
It's impressive either way but the manual version seems ... impossible.
- Re-implements parts of Minecraft
- Runs 512x512 plots in different threads
- Compiles Redstone applying different kind of optimization passes https://github.com/MCHPR/MCHPRS/blob/master/docs/Redpiler.md
- It had Jit backends before, but seems they have been removed
Most recently: Can AI (actually) beat Minecraft? - https://www.youtube.com/watch?v=Wh4abvcUj8Q