pull down to refresh

As I was working on an easy way to integrate Lightning donations into @bitcal website, I noticed a funny thing:
I asked Claude via Cursor to draft a simple dial pad for visitors to input the amount of sats they'd like to donate, and it spit out a decent, functional modal, but as I entered the amount to test functionality, I noticed it would call the API and get a 1000x invoice of what I requested. I looked at the code and saw this:
Apparently Claude believes mBTC (not sats) is the standard.
I gave Claude a proper lecture on how things really are, and it fixed its mistake...
Nevertheless, I am still concerned about the future of sound money πŸ˜…
Be vigilant, folks, take care!
* 1000 is what's used in LN, but that's msats, not mBTC.
reply
There's a lot of LN code that it was trained on? Maybe that explains it.
These tools will go with choices on things that may not be our preference unless we give it the expectations.
One dev I've watched use the tools says treat it like a junior dev that is very fast and doesn't ask questions often. Work upfront really pays off is what I've found. Still learning though. Its not magic as some fear or pitch.
reply
100 sats \ 1 reply \ @Tony OP 26 Aug
Agree. I was surprised it even knew about mBTC as a concept. There was no mention of it in my work with it too.
Apart from being a funny story, it is also a mystery.
reply
AI isn't magic but a ton of it is magic to me as in the bad kind. "What they heck is this doing?"
reply
This is what I expect: mixed training where the msats confuse the pattern. Could happen to your intern too.
reply
Yep. AI is gonna mess up everyone’s accounting one day… by a factor of 1000 πŸ˜‚
reply
100 sats \ 1 reply \ @Oxy 26 Aug
Claude's still on training wheels. Good thing you caught it before it drained your test wallet! A lecture on sound money is probably a new one for it.
reply
It’s actually been very helpful and I wouldn’t achieve what I already have without this new technology.
I still find some things quite entertaining if not outright funny while interacting with it. And of course a fair share of frustrating things. But if it was all butterflies, it would be too boring I guess.
reply
21 sats \ 1 reply \ @kepford 26 Aug
Directions like this are interesting. I'm always wondering why the agent predicted this would be the answer or info I want to get. Or expect to get. It has to be related to the training phase but that doesn't really answer the question of why.
The solution to problems like this are pretty simple by giving the agent context. I've found the Anthropic approach with Claude code is much more effective for my workflows. Still learning how to best use the tools but things like this are something you have to figure out. It seems to me the best path is to make it less uncertain to the agent what you actually want to get. The less guessing the better.
A colleague of mine believe FP (functional programming) is a paradigm that will work better with coding assistants since the flows have less room for "creativity". And it can be easier to understand as a reader of the code.
reply
Exactly. And apart from poor results and wasted time vague prompts or large blobs of info to process drain tokens -> funds, especially with thinking models designed to rid you of the need to think.
I shared my experience on this not so long ago here: #1075448
reply
This shows that Claude is bullish. It's already pricing things in for 2030 when we all transact in mBTC again
reply
Or could simply be trying to curry favor and 1000x the zaps I would receive thanks to its sly trick πŸ˜„
reply
That's even worse than bits.
reply
πŸ’―πŸ˜…
reply
A funny story, but also a critical lesson. As AI gets more integrated into our financial software, these small context errors could have massive consequences. Your vigilance is a great reminder for us all
reply
Wow, is mBTC a thing again? I might have to redo the Satoshi Bitcoin Converter back to how it was when it started. Then: http://satoshitimes.wikidot.com/blog:2 Today: https://satoshibitcoinconverter.com
reply