The real skill is knowing when to bring in the LLM
Quote
I think the learning curve in all of this stuff is, is, is learning when to bring in the LLM. To do a specific, a very specific thing.
Episode 252: Joy Robber (episode page, opens in a new tab) 01:12:05 (play audio from this timestamp, opens in a new tab)
Voice-matched
Summary
Extracting dates, invoice numbers, vendors and amounts from 1,500 pages of invoices in different formats, Dave found that feeding everything straight to an LLM was a disaster. What worked: OCR with existing software (Apache Tika), pre-process the text into a predictable format, and only then hand it to a well-prompted LLM for one very specific job. They tried about ten models before one fit.
Also said
Episode 252: Joy Robber (episode page, opens in a new tab) 01:11:53 (play audio from this timestamp, opens in a new tab)
The first run of just feeding it to an LLM is a disaster. And yes, it will give you things. A lot of times it will give you a bunch of stuff, but that's not the right way to do it.
Voice-matched