If previous Phi models are anything to go by, they will be great at benchmarks, but not for real applications. There are probably better alternatives for production ready SLMs

Completely agreed. I would argue that they are an incredibly tangled box of wires that is open, but really hard to make sense of. Incidentally, I was also just reading a Mech Interp paper by your team which demonstrates just how difficult this is to do

Once, it was being renovated. Two signs kept making me (and a bunch of other people) go round in circles. I finally had to ask the staff at duty free after getting thoroughly confused.

Have to disagree. It’s CDG and it’s not even a competition

Wonder if this is related to their sudden site today 🤔

There seems to be a very strong anti-AI bias among this platform’s users. They say AI, but what they actually mean is LLMs

There was Inspect, by the AI Safety Institute github.com/UKGovernment... and there was also LangSmith docs.smith.langchain.com

LMStudio has been pretty decent for me. However, not all models are supported by the mlx community yet. It’s growing fast, though