I’m playing with the little models at home on a local GPU. So far I’ve fine tuned one and gotten some of the better ones (WizardVicuna) to work as a complete dr

I’m playing with the little models at home on a local GPU. So far I’ve fine tuned one and gotten some of the better ones (WizardVicuna) to work as a complete drop-in for a use case I was testing with OpenAI. The surprising thing to me was, at least for my fairly simplistic requirements (Parsing an API and making decisions) it’s actually responding with less latency than the OpenAI API and the same degree of accuracy. I think we are rapidly approaching the proliferation of good, local models. 99% of “AI” does not need a fleet of thousands of A100s.