LLMs are incredible. I would argue that, at times, LLMs can feel like magic. Some would capitalize on that argument by pointing out that modern “magic is performed by tricksters and illusionists—implying that LLMs do nothing more than give the illusion of intelligence. The truth is LLMs are magicians, but their greatest trick is not writing code or perfectly formatting your e-mail, it’s convincing founders, executives, product managers, etc. to make their products less useful by adding LLM-based features nobody asked for. Many of these features are unnecessarily complex because they attempt to shoehorn LLMs to work in systems that were designed decades before the Transformer even existed.
LLMs deserve better. Your products deserve better. We all deserve better. To truly unlock the potential of LLMs, we need to start thinking about how to design systems that are LLM-native. In this talk, I will discuss what it means for a system to be LLM-native, what you need to consider when building LLM-based systems, and demonstrate LLM-native alternatives to decades old software solutions.