On: My Beancount books are 95% automatic after 3 years (2024)

[conradev]

Language models completely upended my Beancount setup. To me, there is no point fiddling with precise parsers when a language model can read any PDF. I have the language model extract balance assertions from the PDF (beginning/end balance) so that it grounds its work in reality.

I dream of a future where anyone can download a ZIP of all the PDFs they've ever received from their bank, drop it onto a local app, and wait while it creates an entire accounting setup for you.

Edit: also not mentioned here is Fava, which is a really nice web UI for Beancount (https://beancount.github.io/fava/). I share a link with my accountants, and they find it convenient (for downloading files, at least).

[conductr → conradev]

Please let's not start a vibe accounting trend

[conradev → conductr]

Come on in, the water is great!

You can also do regular accounting with vibe programs

[xyst → conradev]

Lovely, now LLM can hallucinate how much I spent or earned in a month, or year.

[fangpenlin → xyst]

For a usecase like this, a local running model would be ideal. I won't like to share my personal accounting books with LLM either.

[conradev → fangpenlin]

Very much agreed. I couldn’t even bring myself to use GitHub (much less BeanHub) for my Beancount.

I use git locally because the ledger is extremely sensitive

[conradev → xyst]

It makes it a lot harder if you check the balances!