Exactly this. At a recent hackathon, I spent the bulk of my time actually engineering, building custom model architecture in Lua, wrangling a dataset, and waiting on training loops. The winning project? A slick UI wrapped around a basic LLM API call.
It’s incredibly frustrating. The incentive structure at these events has completely shifted away from actual technical complexity toward whoever can build the prettiest ChatGPT wrapper in 6 hours..