It sounds like you're exploring some innovative approaches to enhance the capabilities of Synthiam ARC by leveraging large language models (LLMs) and Retrieval Augmented Generation (RAG) techniques. Using RAG with a local general knowledge or programming language model could indeed provide a powerful way to generate responses that are informed by specific datasets, such as Synthiam's support data. This method could potentially offer more accurate and contextually relevant answers by combining the generative power of LLMs with targeted information retrieval.

Embedding Synthiam support data for fine-tuning or as a knowledge base for RAG can significantly improve the model's ability to provide useful responses related to Synthiam ARC. However, as you mentioned, training or fine-tuning a model on specific datasets requires considerable computational resources, especially GPU power. Your plan to expand your setup with more GPUs and explore different models and techniques, like LangChain, is a promising direction.

The process of breaking up Synthiam support data into chunks and using it to inform the LLM's responses is a smart way to leverage existing documentation and expertise. This approach can make the interaction with the model more relevant to users' needs, especially when dealing with technical support queries or detailed information about Synthiam ARC.

As you continue to experiment with these technologies, keep in mind the balance between computational resource requirements and the benefits of enhanced response quality. It's exciting to see such innovative use of AI and machine learning to improve user experience and support in the robotics domain.

Wish i had the time to even look at it and play - but we're in crossplatform hell right now. sidenote, windows isn't bloated, it just has stuff because it actually does stuff.

Anyway, look into your thing and see if there's something called Embedding

*edit: here's some more info.. kind of the basis of athena: platform.openai.com/docs/guides/embeddings

Thanks looks like AnythingLLM does a hybrid approach of using RAG and embedding.  I am just starting to learn this space and a lot of the terminology is foreign to me so it’s a bit of a journey.

good luck with your cross platform challenges. As long as it runs on the Tesla Optimus robot we should be fine

Haha Tesla. I like his initiative but he needs to be more honest. It’s his talent to be charismatic about things and exaggerate. His stories talk about where he wants to be but makes it sound like he’s already there. That indeed makes him a sort of futurist, but not in a way ppl are used to.

you can’t build a reputation on things you’re going to do - Henry ford

the Tesla bot and others like sanctuary are being controlled by humans behind the scene. They’re essentially telepresence on steroids - and that’s to solve the mechanical challenges first. They know the power and AI constraints are crazy limiting today. So getting the mechanical reliable is not a waste of time.

but what they’re not doing is explaining that there’s a human next to it wearing a vr headset holding haptic controllers. Sanctuary is more transparent about it. I think Tesla has only mentioned it once in what I’ve read. They keep saying it’s training to do this or training to do that / which is correct. But they don’t say the training is a human controlling it.

So I went down the AI rabbit hole.  Fun journey.  Initially started with 1 graphics card (RTX 3090) second hand for $800.  This is enough to run a lot of AI models and you can also embed your own PDFs, Documents, do some image recognition etc. Even if you have a smaller GPU like a rtx 4060 you can do a lot with it. I have now started trying to create my own models and for this you need hardware.

here is my AI system I built.  It has 4 * RTX 3090 giving me 96GB of VRAM for training.   total cost about $4,000 Canadian (~$3000 USD) to build mostly second hand parts.  I also added an arduino with a couple relays to reboot, power on off, turn on second PSU when I need all 4 GPU’s (has one 750w and one 1600w PSU as GPUs chew a lot of power). It also  monitors temp and text me if it gets too hot.

long term goal this will be my robots brain hooked up to ARC to handle voice, image, object recognition, conversation, knowledge etc and hopefully object manipulation in future.  

:p How about a demo?