The long awaited feature from OpenAI, “Structured Outputs”, is broken
My articles are free for everybody to read. Non-members can read by clicking this link.
The more I develop AI applications, the more I realize that noise on LinkedIn and TikTok doesn’t come from people who actually develop AI applications. It comes from wannabe influencers.
They love to talk about the latest advancements in AI… while simultaneously having never tried it out themselves. Or, they may have tried it with the smallest toy example, but haven’t created a real production use-case.
An example of this that I noticed recently is structured outputs from OpenAI. This release was championed as this huge deal for AI applications, despite being more of a bug fix.
OpenAI already had function-calling which forced you to supply terribly verbose JSON schemas; it just didn’t work. There was no guarantee that the response would conform to the schema; you were better off begging the model in the instructions to respond how you want it to respond.
And now, OpenAI is claiming with structured outputs, they’ve solved this problem.
I disagree.
