Hearing the same question over and over again either in person or on the various networks, I decided to share my thoughts and of course, to get other points of view, suggestions and experiences.
"How to introduce LLMs and the rapidly evolving ecosystem to engineers in my company with no background on the subject?", they ask.
My quick answer (no presentation or template, sorry) is to try to build it as a workshop and cover the following:
1. Lightweight explanation about the mechanics behind how LLMs work (yes, the transform architecture as well - they're engineers, they should get it. At least high level). There's that cool GPT-2 excel sheet which is a great learning tool. Also mention the various LLMs out there, open vs closed source, and how they are currently benchmarked.
2. Prompt engineering, tactics and methods such as few shot learning, chain of thought, etc. Use a lot of live examples at this point to demonstrate how leveraging different methods improve the outcome. Since it's a workshop you can pause for a challenge aiming to achieve something with an LLM that is usually done by combining those methods.
3. Next I would suggest exposing them to embeddings, vector databases and the RAG concept. It has become such a common and cardinal practice, you shouldn't stop before going over this as well (but you might want to split the session into two). So after a short theoretical explanation about embeddings and vector DBs, followed by examples, introduce RAG and immediately start another hands on practice. There are so many resources online that will guide any developer to implement a system that takes documents, chunks them up, creates embeddings, stores them in a database and finally enables you to query the data, all in under one hour. Tip: Use company-specific data for this, such as your knowledge base, your blog posts, your product guides, history of a public Slack channel, etc.
4. Finally, talk about fine tuning LLMs (I wouldn't go too deep into this one), and if you haven't explained before, that creating embeddings is also done using a model, which can be a fine-tuned foundation model's embedding model.
To sum it up, talk about the opportunities for your company/products/services that emerge given such capabilities. Sometimes "newcomers" have the best ideas!
What would you add or do different?
Have you tried something similar, how did it go?
