Just heard the term “Epstein class” from HCR’s youtube the other day and yep

🤣 agreed

Sry, didn’t mean it literally, rather that the data in the file isn’t stored in a hierarchical structure like OOXML (think word docs, pptx) that would allow you to, for instance, change the font size of all the text in a document and see it re-flow the content to fit the available space

Is it a bubble? We’ll know once it pops. But I do know I’ve built a practical, incredibly valuable feature that was utterly impossible to build 2 years ago, and it don’t work without CUDA acceleration.

pdfs have virtually no structure (which is why you can’t edit them) but VLMs made a huge leap possible. Anyway the L4 is not cheap and we are VERY happy to be paying for it. Early user impressions of the feature are superlative to put it mildly.

I’m building a feature right now that requires a cloud service with an attached gpu (in my case an nvidia L4) bc it uses a library (docling, which sprang into existence in the last two years) that parses & converts pdf files into markdown using VLMs. This has been a stupidly difficult thing to do bc

Who is the 1?

Is this a real post?!

Loser!