dspy.ai is another I'm aware of. Has some interesting ideas around type signatures and prompt optimization

Pulling metadata from BACnet is trivial if you stick to the explicit fields, but deriving a fuller Brick model from that requires extracting metadata from abbreviations/etc. I wrote about the problem here: gtf.fyi/posts/brick-.... Feel free to send me an email to discuss further!

Great to hear you are working on BIM-BMS integration! The method we did in 2020 relied on the structure of IFC's definitions and a mapping dictionary to integrate that information into a Brick model. We struggled in part because we didn't have IFC4 models (which have better metadata than IFC2). /1

This reinforcement loop is built into the way these models are trained and developed

I think an interesting one is the "model alignment" by AI providers to ensure models give the output you want to see. We are more likely to return to use models that reinforce/complement our beliefs (even incorrectly) than those that (correctly) challenge our expectations or beliefs

this only seems to hold true if AI capacity is finite *with respect to what it is competing against* (i.e. human labor). Increasing AI model efficiency and expanding capacity through building more datacenters will probably shift the balance away from the happy case described in the article 2/

Mine only supports heap scan :(

6/8. In science, and in institutions of learning like colleges and universities, there is a shared understanding that edicts should not go unquestioned—that we all have a duty to systematically investigate the world around us and to interrogate claims about it. We do not simply obey.

they are difficult to use (esp. for non-trivial tasks) without deep familiarity and experience in programming. I have worked with comp sci students at the high school, ugrad and grad levels and many of them have still have trouble doing effective prompting and curation of LLM output.