This works by taking a language model that won't fit in a single consumer GPU's memory, partitioning it layerwise, and running it distributed across a bunch of

This works by taking a language model that won't fit in a single consumer GPU's memory, partitioning it layerwise, and running it distributed across a bunch of different people's computers. If I'm understanding correctly, then any single node acting dishonestly can replace the output out its portion with whatever they want, and (if every other node is honest), this is sufficient to fully control the output. So, probably okay to use for prompts like "rewrite Rick Astley lyrics in the style of Shakespeare", but not something you'd want to use in a way that feeds into another automated system. Meta-level, I think it's bad for the world if there's good technology for running neural nets on distributed consumer GPUs. From a cybersecurity perspective, Windows gaming PCs are easy pickings compared to datacenters, and I think there's a risk that after a few more iterations of AI development, we'll start getting systems that figure out they can increase their own power level by building a botnet that runs additional copies of themselves.