LLMs are on a very steep improvement curve, in general. Claude is good today. Something else from some group that figures out the next useful optimization will be better tomorrow. Repeat/rinse for at least 10 years..
A lot has been written about LLMs having reached a plateau with regards to improvements. They still all produce garbage way too often. LLMs have fundamental limitations that can't really be fixed. Garbage in / garbage out also applies, and that is only getting worse with LLMs being trained on ever growing volumes of "AI" slop that is permeating everything lately.
That group will never be Microsoft, and if it is, that model won’t be the one they’re shoehorning into every enterprise product for free.