In light of the recent forays by AI projects/products into the reason of coding assistants, from copilot to Devin, this reads to me as a sign that they've finally accepted that you can't make an ai assistant that provides actual value from an LLM purely trained on text.

This is Microsoft copying Google's captcha homework. We trained their OCR for gBooks, we trained their image recognition on traffic lights and buses and so signs.

Now we get to train their ai assistant on how to click around a windows OS.