Parsing vs scraping: OCR and document intelligence as their own discipline
Scraping goes and GETS documents; parsing reads the ones you already have. Different verbs, different tools, different industries.
Scraping is the web-facing family — /crawl, /scrape, /search endpoints (Firecrawl), deep-search APIs (Perplexity Sonar). Parsing is Optical Character Recognition over a document: PDF in → tool → clean output (human-readable text, or token/structured forms an agent consumes). The key market fact: frontier LLMs DO have built-in OCR — his live demo hands a handwritten Declaration-of-Independence image to GPT-5.6 and it reads it correctly — but 'this is just by using a ton of tokens.' Models trained specifically on document intelligence (Datalab's family) or classic engines (Tesseract) do the same work cheaper, and often better on structure: tables, headers, H1s, images all TAGGED with positions.
Tagging is the bridge to the wider industry: what Datalab's playground shows is data labeling/annotation — 'there are companies like Scale AI... an entire company would just be doing data annotation.' Use cases stack up from the cohort itself: healthcare records, finance invoices, legal documents, contract management, manufacturing, teachers grading handwritten exams.
The Berkshire Hathaway income statement and a patent diagram in Datalab's gallery — structure-heavy documents where 'just dumping the PDF into the AI agent... might just be gibberish to it.'
Choosing the wrong verb costs real money: LLM-OCR on a book-length scan is a token bonfire; a parsing engine is a fixed, tiny cost.
Parsing is obsolete because modern LLMs read documents natively.
LLMs read documents EXPENSIVELY. Parsing engines do the extraction for near-zero marginal cost; the LLM belongs downstream, if anywhere.
This session pairs with your own token discipline: the reason your wrap scripts are Python and not model calls is exactly the parsing-vs-LLM economics he teaches here.
Go deeper
In one line: Parsing = OCR-based extraction from held documents (PDF → tool → text/structured output), distinct from scraping (web retrieval); powered by document-intelligence models or classic OCR engines rather than general-purpose LLM vision, which works but scales in tokens.
LLMs come with their own OCR — capability isn't the question, unit economics are ()
Tagging output = classified regions with positions (header/H1/text/image) — data labeling as a product ()
Playgrounds before API keys: every tool in the category ships a free playground 'just like platform.openai.com' — test before spending ()
Education use case he keeps returning to: handwritten exam-sheet grading ()
▶ Watch this taught:
Answer from memory first — the recall attempt is what makes it stick. Then reveal.
When is LLM-vision OCR still the right call?
Low volume, messy one-offs, or when extraction must flow straight into reasoning anyway — the moment volume or repetition appears, a dedicated parser wins on cost and consistency.
