It's raining LLM's, hallelujah
(Posted 18:31:47 on 6th May 2023 by Rag)
Tall, blonde, dark and lean
Rough and tough and strong and mean
Maybe not, lol. But we do seem to be getting new LLMs on a daily basis which is really exciting. ChatGPT still sits at the top of the pile when it comes to AI models, particularly GPT 4 which is incredibly powerful, but we are seeing a lot of models being produced that are open source and being made available to the community. It kind of seems to me that we're ending up with a commercial and open source divide, kind of like Windows and Linux. Windows being ChatGPT and Linux being the amalgamation of all the open source flavors. I have to be honest, I'm running Oobabooga as the UI for StableVicuna, GPT4-alpaca, Vicuna, Pygmalion 6b and 7b. I'll soon be adding Wizard which has just dropped. I've got multiple StableDiffusion derivatives running on Automatic 1111 for text to image generation which is all pretty powerful stuff.
One thing that's interesting is that the latest Pygmalion model, the 7 billion parameter one, was trained using LLaMa which is Meta's LLM. I believe the source code was leaked, so it appears folks are just using it now to train models. The Wizard model also offers a new background on how it was trained and it's proving to be pretty efficient. The model is only a 7b model, so it doesn't have the power of the 13b models, however, it seems to perform pretty well. I also think an uncensored version of Wizard has been released.
I guess one of the main things to note is that companies are trying to lock down the use cases of AI so they can start to utilize them in their business, but the problem is that these models are progressing at such a rate, as soon as you lock in a business case, it's obsolete before you can even train the model on specific data because there's suddenly a better and faster way to train the models. Exciting times!
