Recently i have come across inception ai mercury dllm model and is really amazed by its speed with great quality response. a bit of research has also confirmed that due to the nature of diffusion that has a global view instead of token by token view, it can natively handle generating structure output so much better. Like it can in theory generate the json structure then fill in the data. Im just blantly thinking - with this emerging field (Google is joining also), if that becomes the new king, it would kind of defect the main selling point (IMO) of BAML on the structured output decoding. what is your take on this?

https://boundaryml.com/blog/structured-outputs-create-false-confidence This post likely explains it best! But on top of that, the DX and testing loop BAML makes possibles will also be helpful even in that world 🙂

really well written blog, thanks for pointing me to it!

So wait.. is this article in favor of, or a criticism of things like BAML?

it just poses the different options and their tradeoffs! we think SAP is the most robust way so far.