Helped my very non-technical classical pianist brother tonight by setting up a system where he pastes a well-constructed prompt into Claude, and it turns the 12-16 hours/month of accounting work he does for his music school into 30 minutes.
-
Helped my very non-technical classical pianist brother tonight by setting up a system where he pastes a well-constructed prompt into Claude, and it turns the 12-16 hours/month of accounting work he does for his music school into 30 minutes.
-
Helped my very non-technical classical pianist brother tonight by setting up a system where he pastes a well-constructed prompt into Claude, and it turns the 12-16 hours/month of accounting work he does for his music school into 30 minutes.
It took us about an hour from start to finish, connecting Claude to his calendar and a Google Sheet, along with monthly payment statements from Venmo, Zelle, and Cash App. It took a bit of iterating because he uses notations like NS for No Show or OS for Out Sick, but by the end it all worked great.
-
It took us about an hour from start to finish, connecting Claude to his calendar and a Google Sheet, along with monthly payment statements from Venmo, Zelle, and Cash App. It took a bit of iterating because he uses notations like NS for No Show or OS for Out Sick, but by the end it all worked great.
These are the kinds of things I get really excited about as AI products develop and become more widely accessible. We still have a long way to go when it comes to making AI easier to use — but at the same time something like this would have been functionally unimaginable 3 years ago.
-
These are the kinds of things I get really excited about as AI products develop and become more widely accessible. We still have a long way to go when it comes to making AI easier to use — but at the same time something like this would have been functionally unimaginable 3 years ago.
@mergesort Honestly probably even 6 months ago – at least from my perspective, the models are now good enough to do stuff like this with relatively little guidance where before they'd go nuts without being tightly constrained
-
@mergesort Honestly probably even 6 months ago – at least from my perspective, the models are now good enough to do stuff like this with relatively little guidance where before they'd go nuts without being tightly constrained
@kyleve I have to emphasize that while my brother is a brilliant creative he is highly non-technical. He himself says that he just doesn’t think systemically the way did in our “workshop”. Part of this was me hoping to teach a man to fish while we built this together for him, but this was noticeably different than the usual workshops I do. I still spend a lot of time teaching AI to non-technical people at the non-profit, and I feel confident we’re not there *yet* for adoption by *everyone*.
-
It took us about an hour from start to finish, connecting Claude to his calendar and a Google Sheet, along with monthly payment statements from Venmo, Zelle, and Cash App. It took a bit of iterating because he uses notations like NS for No Show or OS for Out Sick, but by the end it all worked great.
@mergesort I‘m very skeptical of a lot of the things people use LLMs for. But I‘m trying to approach your example with an open mind
Accounting seems like one of the things you really shouldn’t use an LLM for as you want 100% accuracy, you want determinism and you want to be able to understand how the system gets to a result.
How did you solve those problems?
Lets ignore „understanding the system“ and „deterministic“ for now: how did you make sure the result is actually correct (and not just „in the right ballpark“ but „accurate to the cent“, it’s about money and billing/paying people after all)? It seems like testing this and making sure it works correctly alone would take a lot longer than an hour? How did you approach this? How did you solve it?
-
@mergesort I‘m very skeptical of a lot of the things people use LLMs for. But I‘m trying to approach your example with an open mind
Accounting seems like one of the things you really shouldn’t use an LLM for as you want 100% accuracy, you want determinism and you want to be able to understand how the system gets to a result.
How did you solve those problems?
Lets ignore „understanding the system“ and „deterministic“ for now: how did you make sure the result is actually correct (and not just „in the right ballpark“ but „accurate to the cent“, it’s about money and billing/paying people after all)? It seems like testing this and making sure it works correctly alone would take a lot longer than an hour? How did you approach this? How did you solve it?
@cocoafrog @mergesort I don’t wanna speak for Joe and I‘m curious about his answer, but the (implied) critique of LLMs feels a bit „outdated“ based on my personal experience - as that while there is the fundamental problem of determinism, it doesn’t reflect my experience using LLMs in the past months or rather weeks.
One potential solution to this that comes to mind is instructing the LLM to write a script/software instead of feeding it the data directly - which makes things inherently more testable and verifiable.
-
R relay@relay.publicsquare.global shared this topic
-
@cocoafrog @mergesort I don’t wanna speak for Joe and I‘m curious about his answer, but the (implied) critique of LLMs feels a bit „outdated“ based on my personal experience - as that while there is the fundamental problem of determinism, it doesn’t reflect my experience using LLMs in the past months or rather weeks.
One potential solution to this that comes to mind is instructing the LLM to write a script/software instead of feeding it the data directly - which makes things inherently more testable and verifiable.
@myell0w @mergesort yeah, that’s why I‘m curious about Joe‘s approach. I could see „write me an app/script that does this“ work mostly reliably. But Joe mentioned connecting Claude to Google Calendar which sounds like piping the data directly through Claude. But maybe that was just a toot-sized simplification.
Then again, if you let it generate a tool, but you don‘t have the skillset to review that code (-> non-technical user) how do you ensure it does what you want it to do?
-
@myell0w @mergesort yeah, that’s why I‘m curious about Joe‘s approach. I could see „write me an app/script that does this“ work mostly reliably. But Joe mentioned connecting Claude to Google Calendar which sounds like piping the data directly through Claude. But maybe that was just a toot-sized simplification.
Then again, if you let it generate a tool, but you don‘t have the skillset to review that code (-> non-technical user) how do you ensure it does what you want it to do?
@cocoafrog @mergesort I‘d probably instruct another AI to try to break and verify the results, control them myself a couple of times and then trust it and move on
-
R relay@relay.mycrowd.ca shared this topic
