Good morning. It’s an OCR-heavy day, with Mistral and Baidu dropping competing document-understanding models within hours of each other, and the Hacker News crowd already poking at both. Elsewhere, Anthropic is putting Claude inside Slack as a persistent teammate, and a long blog post is doing the rounds arguing the AI subsidy game looks suspiciously Enron-shaped.
Mistral OCR 4 lands, gets the benchmark side-eye. Mistral released OCR 4 with bounding boxes, block classification for tables and equations, inline confidence scores, 170-language support, and a single-container self-hosted option. The company claims an 85.20 on OlmOCRBench and a 72% win rate in human preference tests, but HN commenters noticed the price doubled from OCR v3 to $4/1k pages with little explanation of what changed beyond bounding boxes, and several called out the truncated y-axes on the benchmark charts. One user testing Malayalam handwriting reported the model misidentified it as Kannada — a problem Sarvam apparently nails at 99% accuracy.
Baidu’s Unlimited-OCR takes a different swing. Within a day, Baidu open-sourced Unlimited-OCR, built on top of DeepSeek-OCR and designed for one-shot parsing of long multi-page documents without the memory accumulation that usually trips up VLM-based OCR on big PDFs. The repo credits DeepSeek-OCR and PaddleOCR explicitly, which the HN thread appreciated. Use cases that came up: sheet music scanning with on-the-fly transposition, and citation-grade RAG over books. The obvious comparison is Infinity Parser 2, which has been quietly dominating olmOCR-bench — nobody’s run the head-to-head yet.
Anthropic puts Claude in Slack as a persistent teammate. Claude Tag is a multiplayer Slack integration where one @Claude shares context across an entire channel, builds memory from history, and proactively surfaces relevant info from connected tools. Anthropic says 65% of its own product team’s code now comes from an internal version — a number HN met with skepticism given the company’s recent product polish. TechCrunch frames it as Anthropic muscling into the organizational-context layer where Microsoft Copilot, Glean, Snowflake, and Databricks are all converging. The unanswered question is permissions: Claude inherits channel-level access, which rarely matches an enterprise’s actual data policy.
Is AI’s affordability problem actually an ROI problem? A widely-shared blog post argues OpenAI and Anthropic are subsidizing enterprise users by 40-70x, citing SemiAnalysis numbers showing gross margins go negative once users hit 25% of their rate limits. The HN discussion split the diagnosis: the cost-per-token side will keep falling fast, but companies that mandated aggressive AI adoption end of Q1 are now quietly walking it back because the ROI isn’t there. One commenter noted Anthropic’s eventual IPO filings should settle a lot of the unit-economics speculation — Chinese and open-weight providers are already competing on price in ways the US duopoly isn’t.
Qwen-AgentWorld: world models for agent training. Alibaba released Qwen-AgentWorld in 35B and 397B variants, trained on 10M+ interaction trajectories to simulate agentic environments across 7 domains. The pitch is two-fold: a decoupled simulator for scalable agentic RL, and a warm-up that improves downstream agent benchmarks. The most interesting HN comment pointed to verification rather than training as the killer use case — if a world model can reliably simulate state transitions, you can use it to check an agent’s execution against hard constraints before letting it touch production.
Hollywood quietly buries the Sam Altman biopic. Luca Guadagnino’s Artificial, the nearly-finished Altman drama, can’t find a distributor: Netflix, A24, Focus, and Warner’s Clockwork have all passed after Amazon MGM dropped it shortly following its $50 billion OpenAI investment. The Verge walks through the conflict-of-interest math, which is hard to miss. On the political side, the same outlet reports that AI-aligned corporate super PACs dumped $27.8M into a single NY-12 House primary — Schlossberg vs. Bores — with Schlossberg alleging coordinated bot networks were boosting his opponent.
That’s the morning. Watch whether anyone actually runs Mistral OCR 4 against Baidu’s Unlimited-OCR on the same docs this week — the marketing claims on both sides need someone to call them.