Unsiloed AI Takes First Place in PDF Parsing Benchmark olmOCR-Bench — Beating GPT-5.5 and Claude Opus 4.7

Unsiloed AI, a document parsing startup, has taken first place on olmOCR-Bench, the standard benchmark for PDF OCR. The company's parser v3.1 recorded an 88.0% strict pass rate on an evaluation made up of 1,403 PDFs and 8,413 unit tests, outperforming 18 OCR services including GPT-5.5, Claude Opus 4.7, LlamaParse, Reducto, Azure Document Intelligence, and AWS Textract.

Notably, a secondary evaluation using an LLM as the judge raised the effective score to 94.8%. That correction accounts for cases that are semantically equivalent but mismatch as raw strings — such as differences in LaTeX notation and whitespace — suggesting real-world accuracy may be even higher.

The parser was built specifically for the hard cases that show up constantly in real documents — complex tables, handwritten pages, mathematical formulas, and multi-column layouts — and the evaluation code is public on GitHub.

Anthropic Reaches Its First Profitable Quarter — AI Company Monetization Gets Real

Anthropic, the company behind Claude, is on track to post its first profitable quarter since its founding. With massive infrastructure investments still ongoing across the AI industry, it is a symbolic milestone showing that major AI companies have entered a genuine monetization phase.

Revenue is reported at around $10.9 billion, though it is not yet clear whether that figure is quarterly or annual ARR (Annual Recurring Revenue). Either way, Anthropic is clearly growing fast on both its enterprise API business and its consumer subscriptions, a sign that the foundation-model business model is proving viable.

OpenAI's "Codex Thursday" — Six Features Shipped at Once for the Codex Agent

OpenAI has announced "Codex Thursday," a major update to its coding agent Codex. Six new features shipped all at once, strengthening the agent's autonomous mode. Details are covered in Qiita's daily digest, but the practical utility of agentic workflows appears to have taken another step forward.

Also in the news: the official Claude Code plugin system and a Dell case study on production-grade agentic AI adoption. As competition among AI coding tools intensifies, agent-based development support is going mainstream.

Aomori Prefectural Government Moves Its Official Website Chatbot to Generative AI — Operating Costs Down More Than 70%

The Aomori Prefectural Government has deployed "AIデジタルスタッフ" (AI Digital Staff), a generative AI-based chatbot, on its official website. The previous scenario-based chatbot could not handle unexpected questions, which drove up operational workload, and the prefecture had hit the limits of managing the website in duplicate.

With the switch to generative AI, the chatbot can now respond flexibly to a wide range of inquiries from residents, and operating costs have been cut by more than 70%. The project is drawing attention as a real-world generative AI deployment by a local government.

AI Adoption at SMEs Stands at About 12% — "We Don't Know Where to Start" Is the Biggest Barrier

According to a survey by Leach, AI adoption among Japanese SMEs stands at only about 12%. The most frequently cited obstacle was "we don't know where to start" (62%). Usage is concentrated on efficiency gains in routine tasks such as document processing and data entry, and many companies have yet to reach genuine business transformation.

Even as AI technology spreads, building adoption guidelines and support systems for the SME segment in particular remains an ongoing challenge.

Pope Leo XIV Issues Statement Calling for AI Regulation — "The Risk of Warping Humanity"

Pope Leo XIV has released a statement warning of AI's impact on humanity. Citing effects on employment and the danger of "warping" humanity, he is calling for robust AI regulation. Reported by both the Associated Press and Axios, the statement stands as one of the clearest messages from the religious world on the societal impact of AI.

Meanwhile, a paper published in the Springer journal "AI & Society" argues that AI-driven deskilling is a structural problem, not an individual one. As reliance on AI tools deepens, how human expertise will evolve remains a critical debate.

Local AI in Practice: A Lawyer Drafting Legal Documents on a 12× V100 MoE Cluster

A detailed report by a lawyer running a legal document drafting AI on a home-built cluster of 12× V100-SXM2 32GB GPUs has been drawing attention on Reddit's LocalLLaMA community.

The standout finding: on the V100 generation, MoE (Mixture of Experts) models are decisively more advantageous than dense models. Qwen3.5-122B-A10B (MoE) hit roughly 50 tok/s on just four V100s, outperforming a dense 32B model. The lawyer orchestrated multiple models into a pipeline that assigns roles for drafting, reasoning, review, and gate decisions.

Against the hallucination problem — "local models confidently fabricate citations" — the lawyer built a custom verifier that cross-checks every citation, date, and exhibit number against source materials. That is an instructive lesson for anyone running RAG or agent pipelines.

Pitfalls in Designing Claude Code Automation Scripts

An analysis published on Qiita explains why Claude Code automation scripts (Skills) "work with one but break with thirty." It points to structural problems caused by the interplay of descriptions, hooks, permissions, and the system prompt — worthwhile reading for developers putting Skills to serious use.

---