Session abstract
What you’ll learn
Fine-tuning an LLM using parameter-efficient methods like LoRA takes a ton of effort. You're taking weeks curating datasets, configuring training runs, and burning through compute before you even know if the result is better than what you started with. Meanwhile, the rest of the industry has collectively decided that the answer to most problems is to just write a longer prompt, shove more context into the window, and hope the model figures it out....and somehow, that's kind of working. This talk walks through scenarios that could still warrant fine-tuning and covers the pros and cons of adapting model weights versus investing that same effort into better prompts and smarter workflow design. We'll walk through controlled experiments comparing LoRA-adapted models against prompted baselines to find where fine-tuning still earns a meaningful edge and where prompt engineering and context engineering have genuinely made it unnecessary.