AI News: July 6, 2026

1. Baidu Releases Open-Source Unlimited OCR for Constant-Memory Document Reading

Baidu researchers published Unlimited OCR, a system that reads dozens of document pages in a single inference pass while holding memory use and speed roughly constant. The core trick is Reference Sliding Window Attention, which caps the KV cache at a fixed size so generated tokens attend to all visual reference tokens but only the last 128 generated tokens, mimicking how a human copyist focuses on recent work. Built on DeepSeek OCR’s encoder with a 3B-parameter mixture-of-experts decoder (500M active), it scores about 93% on OmniDocBench, keeps error rates below 0.11 past 40 pages, and ships with code and weights on GitHub and Hugging Face. Source

2. Study Finds AI Search Agents Fail by Not Asking Clarifying Questions

Tencent Hunyuan and Tsinghua University researchers introduced DiscoBench and found that AI search agents rarely fail at retrieval itself, but at recognizing ambiguous queries and asking for clarification. Top models including Doubao Seed 2.0 Pro, Gemini 3.1 Pro, and Claude Opus 4.7 reached only 43%, 41%, and 40% end-to-end accuracy, even though individual checkpoint success was far higher, because a single unresolved ambiguity collapses the whole reasoning chain. When agents asked a clarifying question first, success jumped to 93.4% versus 51.9% for agents that kept searching and guessing, a gap that current benchmarks miss by treating queries as unambiguous. Source

3. Amazon Stops Taking New Mechanical Turk Customers

Amazon will stop accepting new customers for Mechanical Turk on July 30, 2026, leaving existing users on a platform AWS says it will keep secure but will not add features to. Launched in 2005 and repositioned in 2018 as a SageMaker data-annotation tool, MTurk became key infrastructure for AI training data, but it has been eroded by bots and fraud, with a 2023 analysis finding 33 to 46% of workers using language models to complete tasks. The wind-down marks a reckoning for crowdsourced human labeling as AI increasingly contaminates the very data pipelines it depends on. Source

4. Hollywood Moves to Ban ByteDance’s Seedance While Quietly Using It

ByteDance faces a cease-and-desist from the Motion Picture Association, which calls the company’s Seedance video generator “systemic infringement” after viral clips depicting actors such as Brad Pitt and Tom Cruise. According to the reporting, many studios oppose the tool publicly while tolerating its use internally on a “don’t ask, don’t tell” basis, with some creatives calling it the best video tool on the market. ByteDance is expanding regardless, posting 100-plus US job openings, courting indie filmmakers, and discussing funding for AI-generated films. Source

5. Mistral CEO Warns Proprietary Models Give Vendors a View Into Customer Operations

Mistral CEO Arthur Mensch argued that relying on closed AI models hands vendors visibility into a customer’s business processes, claiming some labs “have a track record of going after their most successful customers.” He urged companies to keep data in open systems, set their own access rules, and train proprietary models, summarizing the sovereignty thesis as “if it’s not in your hands, it’s not your growth.” The pitch serves Mistral’s commercial interest, but a Bridgewater and Thinking Machines Lab experiment lends partial support, showing a fine-tuned open Qwen3-235B beat frontier models on finance tasks at roughly one-fourteenth the cost. Source

6. AI-First Private Schools Court Wealthy Families With Personalized Tutoring

Alpha School, an Austin-founded chain, is expanding an AI-driven model that replaces traditional instruction with about two hours of adaptive AI tutoring per day followed by human-led project workshops, adding eight locations in 2025 and planning nearly two dozen more. Tuition runs as high as $75,000 a year, drawing criticism that the approach deepens an education equity gap and boasting supporters such as billionaire Bill Ackman. The trend collides with research showing students who lean on AI for homework often outsource thinking, with one study finding exam performance dropped by up to 24%. Source