New course with Lamini and Meta: Improving Accuracy of LLM Applications

I'm excited to introduce Improving Accuracy of LLM applications built in partnership with Lamni and Meta, taught by Sharon Jo and Amit Sanghani. Many developers have experienced frustration of inconsistent results when working with LLMs. Sometimes the results are just right, while other times they miss the mark. This course will guide you through a development pattern that systematically improves the accuracy and reliability of OM applications. I'm thrilled to introduce our instructor Sharon Zhou, who is CEO of Laminai, which provides integrated OM inference and tuning for enterprises and developers and is actually a company I've made a scene investment in, as well as Amit Sangami, who is director of Partner engineering for the Llama team at Meta. Both are returning instructors who bring a wealth of experience in optimizing LLM applications and using specifically Meta's LLAMA family of models. Sharon, you work with many customers on building production ready applications. Can you say a bit more about what you are seeing? Absolutely. Thank you Andrew. The approach that we found to be most effective starts with creating the initial application, stringing together all the different steps. Then we add in evaluation metrics or what I like to call evals to measure performance. After that we use prompt engineering and self reflection to encourage the model to perform better. Often this isn't enough however to reach production level quality, so we move on to fine tuning the model. This involves creating a fine tuning dataset and iteratively expanding on it. Fine tuning has become more rapid and affordable with techniques like lora. A form of fine tuning called memory tuning also systematically removes hallucinations and embeds facts in the model weights a kind of determinism within a probabilistic model. In this course we'll walk you through all of these steps with the llama3.8 billion parameter model on an example application for text to SQL of a custom schema. That sounds exciting. And Amit, you've obviously worked extensively with the Llama 3 family of models. What can learners expect to see in this course? Thanks Andrew, I'm excited to be back. In our previous course on prompt engineering with llama 2 and 3, we focused on how to best prompt these models. Now we are taking advantage of llama's open source nature to give learners hands on experience fine tuning a LLAMA model. I really appreciate Meta's work on releasing the Llama family of models as open source and I think what she just shared is an example of the types of work that developers can do almost uniquely on models whose ways are available. This short course will cover both some theory on how to think about building LLM applications and the practicalities of hands on building. We hope you'll join us in exploring these cutting edge techniques for optimizing your LLM applications. Yep. In fact, I hope you'll be ready to join us and fine tune your skills with this course. Sa.