It seems like the use of ChatGPT is something like "microtasks". Little things a given person could do but would rather not and so is able to delegate to an automatic thing whose output they can verify.

It seems like it's potential as of today is increasing or seeming to increase the productivity of a segment of white collar workers in the fashion that email and the web did (or might not have). A lot of researchers might not have need for this and so not understand this appeal of this.

Thanks for your comment, I think this is actually very helpful framing! It at least gives me something more tangible to think about—“micro tasks”.

I, too, have been experimenting a bit to see if/how an LLM might help me. This microtask framing jives with my experience.

One of the best examples so far for me (and it's truly micro) was at grocery store. Friend trying to figure out how big of a rice bag to get and avoid not finishing it before a long trip coming up. She knew she ate a couple of cups dry a week.

"I eat 2 cups dry rice per week. Can I finish a 25lb bag in less than 4 months?" "Yes" (it did show its work).

One shot, perfect response. I know this kind of computational thing is what WolframAlpha was for, but that wasn't nearly as reliable. I know I could figure it out myself, but I'd need to find a reasonable figure for the density of rice and probably do some imperial metric conversions and generally futz around for longer than one would want to stand in front of a pallet of rice bags.

50 cups of rice in a 25 lb bag, 4 months = 16 weeks = 32 cups at 2 cups / week. I think there will be some rice left.

Maybe this is one of those examples where this tool gives a confident and wrong answer even when it shows its homework?

Honestly this sounds more like a place for GPT/Wolfram API than just LLM alone.

Really? You're using ChatGPT to do basic math you could easily do in your head?

If we're characterizing LLMs as a kind of input interface I'd point out that the first GUI was released in 1973 by Xerox and the first commercially successful GUI was released in 1984 with the Mac 128K. It took 11 years for someone to answer this question. Sure things move faster these days, but we're still only a couple of months in.

With blockchains there was also a fundamental technological breakthrough (I'd argue less revolutionary than LLMs). The problem was that everyone jumped the gun and proclaimed the killer app had been discovered too soon: cryptocurrency's incarnations to date have yet to demonstrate much utility apart from being a vehicle for speculation. Nakomoto invented the first distributed blockchain in 2008...

It’s programming. The thing is really great at programming, and everything else is icing on the cake. They already got my $20 a month lol

Anecdotally, I've had more writers (screenwriters and copywriters) tell me they're using ChatGPT than programmers. I think people here might underestimate how big a deal it is in "the real world".

Not really programming but generelly producing text (or maybe generally certain structure like ASTs at some point). Programming is a subset.

I expect that this will move lots of things into the world of text that weren't before.

Bingo. No programmer seems to want to program without it anymore. Sounds like a killer app to me.