Anthropic ships Opus 5 at unchanged pricing, Andrew Ng releases a desktop coworker that hands you finished work, and a Rust tokenizer hits 24.53 GB/s. And the open OCR and ASR fields get honest benchmarks. Datalab's Marker 2 vs MinerU, Docling & LiteParse: The Doc-Parsing Benchmark
Marcel Røed released GigaToken, an open-source Rust BPE tokenizer that encodes text at 24.53 GB/s on a 144-core AMD EPYC 9565 — vs 24.8 MB/s for HuggingFace tokenizers, a claimed up to 989x speedup (and ~681x over tiktoken). It's a drop-in replacement, 66% Rust / 33% Python.
So what: Tokenization is invisible until it's your bottleneck — at data-prep and high-throughput inference scale, it absolutely is. A drop-in that turns MB/s into GB/s reclaims real wall-clock time from every pipeline that touches text.
A head-to-head of open PDF-to-markdown parsers. Marker 2 scores 76.0 on olmOCR-bench at 2.9 pages/sec — matching Gemini/MinerU quality while running ~5x faster, beating MinerU and Docling on score and throughput at once.
So what: Document parsing is the silent failure point of most RAG stacks — bad extraction poisons everything downstream. Winning quality and speed simultaneously makes Marker 2 the new default front door for turning PDFs into clean model input.
sponsored
Andrew Ng released OpenWorker, an MIT-licensed, local-first desktop agent that returns finished deliverables instead of chat — a polished doc, a filled spreadsheet — working across 35 app connectors while keeping data on your machine.
So what: The "chat with your AI" paradigm is giving way to "hand me the output." A local-first agent that produces deliverables — not turns — is a real bet on privacy plus autonomy, and open-sourcing it sets the template everyone else copies.
Anthropic released Claude Opus 5, its new frontier model tuned for agentic coding and computer use — and it holds Opus pricing at $5 / $25 per million tokens. It tops SWE-bench Verified, extending Anthropic's lead on real terminal-based coding tasks.
So what: Frontier capability jumps while price holds flat — that's the whole story of this cycle. When each Opus generation gets meaningfully better at agentic coding without raising the meter, the economics of shipping software with an agent in the loop keep improving.
A hands-on build: a stateful agent that runs a venue-operations workflow end to end. MongoDB Atlas for data + vector search, Voyage embeddings for retrieval, and LangGraph to orchestrate the planning / tool-calling loop into a durable graph.
So what: A clean reference for the emerging "database + embeddings + graph orchestration" agent stack — the pattern most production agents are quietly converging on.
sponsored
A practical roundup of the top open ASR models on WER, languages, latency, and license. NVIDIA's Parakeet RNNT 1.1B tops the Open ASR Leaderboard, beating Whisper large-v3 on average WER, with Canary 1B leading multilingual transcription + translation.
So what: Open ASR has quietly caught and passed Whisper. If you're building voice agents, the right pick now hinges on the tradeoff table — WER vs latency vs language coverage vs license — not on "just use Whisper."
|
Document AI How AI Learned to Read
|
Part 1 · The Problem
Every day, businesses process billions of documents. Invoices. Receipts. Contracts. Insurance claims. Most start as paper or PDF — unstructured information a computer can't use. A human reads each one, finds the important parts, and types them into a system. One invoice takes 3-5 minutes. Multiply by thousands per day, and you've got entire teams doing data entry.
|
💵 The Old Way Human opens invoice → reads it → finds "Total: $4,820.00" → finds "Invoice #INV-2024-0815" → types each into accounting software → repeats 2,000 times. 3-5 min per document. 96-98% accuracy. Entire teams of people. |
The problem isn't that humans are bad at this. It's that there are too many documents and not enough humans.
Part 2 · What is Document AI?
Document AI automatically reads documents, understands what's in them, and extracts the important information — turning unstructured paper into clean data your systems can use.
Think of it like a new employee who's incredibly fast at reading. When you hand them an invoice, here's what happens:
|
1. Classification — "This is an invoice, not a contract" Sorts mixed documents into piles. You can't extract "invoice number" if you don't know it's an invoice. 2. Layout Analysis — "Total is bottom right, date is top left" Like reading a restaurant menu — you see sections and prices, not just words. This is what separates Document AI from basic OCR. 3. OCR + Extraction — Reads the text, pulls out the 5 things you care about Like your phone reading a license plate in a photo. Then NLP finds invoice number, total, date, vendor — understanding "Total Due" means the same as "Amount Payable." 4. Validation + Human Review — Checks the math, flags anything uncertain Line items don't add up to the total? Flag it. Messy handwriting? Show that one field to a human — 5 seconds of review, everything else automated. |
The difference? This "employee" processes a document in seconds, never gets tired, and handles the boring part so humans focus on exceptions that need judgment.
Part 3 · Real-World Examples
|
🧾 Invoice Processing 5,000 invoices/month in PDF, email, and paper. AI reads each one, extracts vendor, amount, line items, tax — routes to accounting. 50-70% faster. Humans only review the ~5% with issues. |
|
📝 Contract Analysis A 50-page lease agreement. AI reads the whole thing, identifies key clauses (termination, renewal, rent escalation), flags unusual terms, creates a summary — in under a minute. What took a junior associate 2 hours now takes 30 seconds. They review flagged items instead of reading every page. |
|
💉 Insurance Claims A claim with a photo of a damaged car, a police report PDF, and a handwritten repair estimate. AI reads all three, cross-references damage descriptions, extracts the estimated cost, routes to the right adjuster. Claims that took 5-7 days now get triaged in hours. |
Part 4 · The Numbers
|
53% of companies that implemented Document AI reported measurable productivity improvements. Processing time drops by 50-70%.
Part 5 · How We Got Here
|
1990s — Rule-based OCR: Rigid templates. One format change and it broke. 2010s — Deep Learning OCR: Learned to recognize characters across fonts and angles. But only read text — didn't understand it. 2020s — Document AI: Combined OCR + NLP + layout analysis. Now it understood structure and meaning, not just characters. 2026 — Vision-Language Models: GPT-5 Vision, Claude, and Gemini that "see" a page as an image and reason about it. They handle charts, handwriting, and messy layouts that would break older systems. |
The main platforms today:
Google Document AI — high accuracy, pre-built processors for invoices and receipts (~82% field accuracy)
AWS Textract — strong at tables and forms (~78% field accuracy)
Azure Document Intelligence — excellent layout analysis, multi-page PDFs
LLMs (GPT-5, Claude, Gemini) — newest, can reason about documents, best for unstructured or complex docs
Part 6 · The Aha
|
“A scanner copies. A reader understands. |
Old OCR turned pixels into characters and stopped. Document AI understands that the number in the bottom-right corner is the total, the signature means the contract is executed, and the checkbox means the user accepted the terms. Not faster reading. Actual understanding.
$ ./partner --with us
# Reach 1M+ AI developers, researchers & builders.
# Sponsor a future issue of the Intelligence Brief.
> apply: forms.gle/CY1eqZzuWFQBp7dH9