There are an emerging set of components that have proven useful in solving the common problems you run into when building applications powered by generative AI. We discuss the generative AI stack and explain from first principles why each component exists, and when they are appropriate to use. You'll build a working prototype of an AI product by chaining multiple components together using LangChain. The application we'll build is an AI blog post generator, which researches a given topic and summarizes the search results, before generating a comprehensive post based on the research, written in your writing style rather than sounding like AI. We’ll cover advanced topics in applied AI, such as multimodal vision and image generation, open-source models like LLaMA 3, RAG (Retrieval Augmented Generation), as well as the basics like building a simple user interface for AI applications. Basic experience with Python programming and access to an OpenAI developer account with a payment method is required to follow along, as well as a free Replit account to avoid any dev environment setup issues. You'll get access to all of the code to use in your own applications.

Location: Elastic, 35 W Wacker Drive, 5th floor

Trial and error can only get you so far when working with generative AI, because when you're running a prompt hundreds or thousands of times a day, you need to know when and why it fails. Prompt engineering isn't about finding the right combination of magic words that tricks the AI to do what you want, it's a process for building a production-grade AI system that delivers the results you need, reliably and at scale. We'll apply prompt engineering principles to a real-world AI use-case and make the strategic trade-offs needed to make your AI products economically viable. If you have tried prompting to automate a task, but couldn't get good enough results, this talk will give you actionable steps for closing that gap. You'll take away a checklist for optimizing prompts from idea to production, using principles that are transferable across models and modalities.