as a test I have spent quite a bit of time this week fucking about with copilot to see if "give me a week and copilot and I could do this whole task" that someone told me (about a complex analysis task) was remotely accurate. I have learned three things: 1. copilot is a piece of fucking garbage for any task that takes more than about 3-4 prompts. It drifts off-task almost immediately, almost never reads the documents you upload to it, and will just invent shit and pretend it has read it in your docs. 2. because of 1, using copilot for anything remotely complex is incredibly frustrating. I am working from an 8-step framework and so far I have run the first 4 steps about 5 times trying to stop it from drifting off-task. 3. using copilot like this is much more time intensive than just doing the work yourself. I've spent so much time writing and editing prompts, testing theml rewriting them, testing again, moving to the next step then finding that the edits to the prompts I did previously have now broken something and I need to go back and fix it. Then copilot shits the bed and I have to start in a new chat (after about 6 times doing this I created a framework step to re-start the analysis using previous outputs). My final comment is that I have learned a lot about my own framework this week, and a lot about the limitations of machine learning for analytical tasks.