I spent the entirety of Friday trying to understand how LLM's work, what goes on with these generative models, the difference between encoder and decoder models, why one might want to use encoder models to embed documents, and use decoder models to generate new content and so forth. It was all quite interesting since I'd zoned out on the happenings in the field for a few years, as things seemed to be moving so quick. Now it appears things have stabilized a bit.
The most exciting thing that I found out though was langchain. Well not the library itself which I'm made to understand is terrible, but the concept of chaining steps in models, and giving model the tools to accomplish its tasks. So you can tell a model: hey, call this function if you want to run google, call this function if you want to run calculator, run this function if you want to run wolframalpha, and it does all of that. It identifies the steps it needs to take to accomplish a task, and if the steps have been made available in terms of tooling, it uses them!
Which means it's not just a model now, the whole world is wide open. It's just an orchestration platform for other tools. Yes, the 'controller' is still text-based ui, and a cli in the end, but all of a sudden it's not a tool boxed up, it's an interactive tool that is able to leverage other tools, which gives it so much more strength.
No comments:
Post a Comment
Tell me what you think. I'll read, promise.