↳ In-reply-to » I played with nlpodyssey/verbaflow: Neural Language Model for Go today a little bit today.... First I had to download a ~2GB file (the model), then convert that to a format the program verbaflow understands which came out to roughly ~5GB. Then I tried some of the samples in the README. My god, this this is so goddamn awfully slow its like watching paint dry 😱 All just to predict the next few tokens?! 😳 I had a look at the resource utilisation as well as it was trying to do this "work", using 100% of 1.5 Cores and ~10GB of Memory 😳 Who da fuq actually thinks any of this large language model (LLM) and neural network crap is actually any good or useful? 🤔 Its just garbage 🤣

@prologic@twtxt.net You more or less need a data center to run one of these adequately (well, train…you can run a trained one with a little less hardware). I think that’s the idea–no one can run them locally, they have to rent them (and we know how much SaaS companies and VCs love the rental model of computing).

There’s a lot of promising research-grade work being done right now to produce models that can be run on a human-scale (not data-center-scale) computing setup. I suspect those will become more commonly deployed in the next few years.