I found this post by John Rush to be a fascinating read. Come for the vibe-coded infographics and confirmation that eggs are indeed more expensive, but stay for the deep dive into co-working with multiple AI tools. It’s also another reminder that, particularly around OCR, some tools are vastly better than others, so don’t get hung up continuing to use the one you’re familiar with if it’s not producing good results!
Some of the quotes I highlighted in John’s post:
- “First test on a sample bank statement: clean, accurate text in 2.1 seconds. Tesseract was faster but dramatically noisier. Second test on a tall Fred Meyer receipt: disaster. The model entered a repetition loop, hallucinating “TILL YGRT” endlessly.”
- “The models love regex.”
- “On the fourth attempt I said “I would have expected we start a new process per batch.” That was the fix… Codex patched it, launched it in a tmux session, and the ETA dropped from 12 hours to 3. Not a hard fix. Just the kind of thing you know after you’ve watched enough overnight jobs die at 3 AM.”
- “I asked Claude to build me a labeling tool – keyboard-first, receipt image on the left, classification data on the right, arrow keys to navigate, single keypress to verdict. It built the whole Flask app in 22 minutes. I sat down and hand-labeled 375 receipts.”
- “Real-world data is messy.”
- “Here’s what made the quality good: every time I caught something, I could show the agents what to look for and they’d go fix it everywhere.”
This reminds me, for the umpteenth time, of the aphorism: “AI won’t take your job, but someone using AI might.”
Leave a Reply