Find out how Bart and James learned to use local LLMs for test tasks, and play with what they have built.
Local LLMs are slower and weaker than frontier cloud models, but are private, unlimited and offline. Learning about local models not only gives you a useful tool for situations where sharing with cloud-based models is prohibited, but gives you skills and insights into the ways to get the best out of more powerful LLMs.
You’ll hear about what’s worked, what hasn’t, and what weirdnesses have emerged. We’ll show you how we have split test-related work between human judgement, LLM flexibility, and persistent / deterministic scripts – and how we’ve pulled the outputs back together. We expect to cover: isolating the environment (and getting instructions, code and data in and out of it), equipping the model with agentic tools and custom skills, curating context to tune behaviour, using tracking tools as a memory, and building tiny tools. You’ll see how we’ve worked with limited LLMs – and how you can reapply those approaches with cloud LLMs to reduce token use and increase dependability.
We will have LLMs running on our local machines. You’ll work on your own machines, using a browser-based cloud environment so that you can instantly access the same models as us. You’ll use the same tools, on the same projects, to find your own paths. We hope that you will share your experiences with your workshop peers, and we’ll learn together.
