but the hardware is not capable. it’s running a miniscule custom 260k LLM and the “claim to fame” is that it wasn’t slow. great? we already know tiny models are fast, they’re just not as accurate and perform worse than larger models, all they did was make an even smaller than normal model. this is akin to getting Doom to run on anything with a CPU, while cool and impressive, it doesn’t do much for anyone other than being an exercise in doing something because you can.
except i’m not wrong. the model they ran is 4 orders of magnitude smaller than even the smallest “mini” models that are generally available, see TinyLlama1.1B [1] or Phi-3 3.8B mini [2] to compare against. Most “mini” models range from 1 to about 10 Billion parameters, which makes running them incredibly inefficient on older devices.
but I can imagine it. in fact, I could have told you it would have needed a significantly smaller model in order to run at an adequate pace on older hardware. it’s not at all a mystery, its a known factor. i think it’s absolutely cool that they did it, but lets not pretend its more than what it is - a modern version of running Doom on non-standard hardware.
[1] https://huggingface.co/TinyLlama/TinyLlama-1.1B-step-50K-105b
[2] https://ollama.com/library/phi3:3.8b-mini-128k-instruct-q5_0
[3] https://www.thirtythreeforty.net/posts/2019/12/my-business-card-runs-linux/