When building your own custom Chatbot using ChatGPT, GPT-4, Eleven Labs etc, you might be running into issues with cost and latency. In this video, I share some thoughts and experiments building a new and upgraded chatbot, that does not have to rely on third party APIs for all of the decision making, just some of it.
In this video, will share with you my experience and successes (as well as failures) and even talk through the logic of some of the code and useful Rust libraries if interested.
The main point from this video should be that you do not always need LLM's to understand and convert text into logic. The Elephant-In-The-Room (hence the thumbnail), is that we have serious demand for this tech, but inadequate local hardware to have our much wanted personal AI assistant to run locally and at near-zero cost. Equally, as the saying goes, if you want to eat an elephant, we need to digest one piece at a time. Here we tackle the voice decoding and decision on whether what a user is asking for is based on a chat request, or a task request (which an AutoGPT would then handle). You can build your own basic deep learning model using the Transformer Encoder to do just that as you will see in this video. This is with special thanks to Corey Lowman who developed the DFDX library which I could not have done this without when wanting to use Rust.
Why Rust? Python comes with a lot of overhead, latency etc. Rust uses very little memory and is pretty fast. In fact, really fast. Or as the Primagean would say BLAZINGLY fast.
If you have any thoughts as to how Open Source LLMS like Llama 2 can be trained, saved and most importantly run very quickly on most desktops, PLEASE add your thoughts on that in the comments.
Credit to:
Whisper Burn: https://github.com/Gadersd/whisper-burn
DFDX (Corey Lowman): https://github.com/coreylowman/dfdx
Useful Resources:
Courses: https://coderaiders.com/site/courses
Rachel Chat: https://coderaiders.com/site/view?pkg...
Смотрите видео Building an Advanced LLM AutoGPT + Chatbot with Rust - Thought Share онлайн без регистрации, длительностью часов минут секунд в хорошем качестве. Это видео добавил пользователь Code Raiders 11 Сентябрь 2023, не забудьте поделиться им ссылкой с друзьями и знакомыми, на нашем сайте его посмотрели 1,51 раз и оно понравилось 5 людям.