ChatGPT works seamlessly with photos and files, enabling numerous new use cases
twitter.com
twitter.com
This does not use any OpenAI service to analyze the image. Rather, it’s using some other image, recognition model to generate a description of the image, feeding that description into chatGPT, and having chatGPT hallucinate other likely details based on that description.
Eventually I hope we’ll be able to able get a single high end server appliance to run an instance in house for such use cases.
Do you know if that hardware is capable of running LLM or is that optimized for something else?
I do! Checkout llama.cpp [0] and related projects. The game changer with this was Gerganov's ggml format that makes inference on CPUs realistic. I'm currently running alpaca.cpp and llama.cpp, using 7B models 4-bit quantized (so they're the lowest powered, lightest weight models) on an M1 Pro with 16gig RAM. You could run same on an M1 Macbook Air with 8gig RAM.
Gerganov shows the Whisper model running on an Android phone.