Google has released Gemini 3, emphasizing "deep reasoning." Reports state that Gemini 3 Pro leads on LMArena with a 1501 Elo rating; it achieved 91.9% accuracy on GPQA Diamond and 37.5% on Humanity’s Last Exam, which improved to 41% after enabling Deep Think. Deep Think is now available to AI Ultra subscribers paying $250 per month, scored 45.1 on ARC-AGI-2, and will expand rollout over the coming weeks. These capabilities have been integrated into Search AI Mode, the Gemini App, AI Studio/Vertex AI, and the CLI.

Interactive Brokers reports that ByteDance plans to invest approximately $14 billion (100 billion RMB) in purchasing Nvidia AI chips in 2026, contingent on H200 chips being approved for export to China. The report also notes that Chinese companies have already reserved over 2 million H200 chips for 2026, while Nvidia's current inventory is about 700,000 units, indicating a significant supply-demand gap; a single H200 costs around $27,000. To boost production, Nvidia has contacted TSMC to expand H200 manufacturing, expecting phased deliveries with priority given to meet strong demand.

Baidu's AI chip subsidiary, Kunlunxin, has filed a confidential IPO application with the Hong Kong Stock Exchange. The fundraising scale and timeline have not been disclosed. Sources indicate its valuation is at least $3 billion. Kunlunxin primarily supplies AI chips for data center servers. Baidu stated that a spin-off IPO would help highlight Kunlunxin's independent value, attract investors focused on general-purpose AI computing chips and software/hardware systems, and expand financing channels. Following the news, Baidu's Hong Kong-listed shares rose 9.35%, closing at HK$143.8.

AWS announced the General Availability of S3 Vectors: a single index can store and retrieve up to 2 billion vectors, with common query latency under 100ms, and integrates with Bedrock Knowledge Base and OpenSearch. InfoQ reports it advocates for a "Storage-First" RAG architecture, managing vectors as S3 objects, eliminating cluster and shard management, and charging based on storage and queries. The company claims total cost of ownership can be reduced by up to 90%, targeting scenarios like large internal knowledge bases, Retrieval-Augmented Generation, and agent retrieval.

Microsoft Research proposes an LLM privacy protection method based on "Contextual Integrity": the open-source PrivacyChecker extracts and classifies sensitive information during inference and injects privacy guidance into prompts, requiring no model retraining. InfoQ cites PrivacyLens results: information leakage rate on GPT-4o decreased from 33.06% to 8.32%, and on DeepSeek-R1 from 36.08% to 7.30%. The team also proposed CI-CoT+CI-RL training methods, using chain-of-thought and reinforcement learning to penalize improper disclosures.

Yahoo Finance reports that Nvidia is acquiring AI chip startup Groq for $20 billion, aiming to gain its high-throughput, low-latency inference technology and core engineering team to strengthen data center inference capabilities and complement its existing GPU products. The report suggests this deal reflects intensifying competition in AI inference compute: vendors are accelerating the deployment of more energy-efficient dedicated inference architectures to address cost pressures and supply tightness from growing large model inference workloads, while also strengthening long-term competitive barriers through software-hardware synergy.

A study analyzed the impact of 500 news websites blocking AI crawlers in robots.txt files: large publishers saw an average 23% drop in total traffic after blocking AI bots, with human traffic dropping 14%, indicating AI search and summarization tools have become important distribution channels. Data is from Comscore and SimilarWeb combined with HTTP Archive records; some mid-sized sites saw growth but the mechanism is unclear. Some publishers later reversed their blocking decisions, reflecting the practical trade-offs between content being used for training and maintaining 'visibility' within an AI-dominated information ecosystem.

A team from North China University of Technology has open-sourced OCR-CrossLing for low-resource language OCR. They claim a systematic rebuild of data, models, and evaluation systems: they built and finely annotated their own low-resource language datasets, adopted a unified vision-language model to adapt to different writing directions and complex layouts, and established a multilingual evaluation benchmark. Reports indicate significant improvements in metrics like METEOR and CER, outperforming GPT-4o and Gemini 1.5 in some dimensions while maintaining performance for high-resource languages like Chinese and English. The project aims to provide a reproducible paradigm and open-source foundation for low-resource script recognition.