It can redraft emails for tone,
It catches semantics errors in documents that slide right past the usual spelling and grammar checks,
It can do a half decent job documenting other people’s undocumented code saving me a pile of time getting to grips with poorly documented code for whatever reason I’m doing that.
It’s like having an assistant that’s not proactive. I can throw tasks I feel it can handle at it and it comes back with work I can double check or fix up obvious issues with, sometimes I’ll have to ask it to redo the work if it’s not right but it also accepts corrective feedback, even if you cut it off in the first couple of sentences because you’ve realised it completely misinterpreted something like the tone you asked for.
Overall I’m very impressed and looking forward to future improvement in similar technologies. I’m absolutely going to buy whatever stupid expensive computer is needed the moment it’s possible to run an equivalent model locally, if I had to buy a top of the line Mac Studio to get enough unified memory to run this, done, no question, the only issues I’ve got with it are I’m sending information to outside parties which limits what I choose to use it for to situations where it’s ok to send that information, and that despite paying for GPTPlus I’m subject to a rationing system that doesn’t best align with my usage. I use it infrequently at most once or twice a day for a short burst then not again for a while, meaning I have to worry about the “every three hours” response limits forcing me to spread my work out over the day to meet their availability demands, if it’s that bad I’d rather a message queue to be honest then I can just fire my messages off and get replies when they’re done, but their real time chat interface is more useful to me than paying per use on the API directly.
Now, there are certainly consequences, but given how awful the track record of predictions about "technology [x] will have consequence [y]" has been, I doubt we will be able to predict it, and there are many previous AI waves (remember IBM's Watson, that beat Ken Jennings at Jeopardy, and Deep Blue, that beat Kaspersky at chess?) which ended up impressing a lot of people and then having little impact. Perhaps this will end up accomplishing more. But, given how excellent it as at Harry Frankfurt's specific definition of BS, we should expect it to be better at impressing people than at doing anything else.
I don't think LLM are powerful enough to produce independent work.
If they do, I consider myself smart enough to leverage the technology to augment myself instead of being replaced by it.