Global AI Daily News 

DeepSeek is advancing a RMB 70 billion financing round with a pre-money valuation of $45 billion, while announcing a permanent 75% price cut for its V4-Pro API. Analysis suggests its strategy is to restructure the AI hardware ecosystem, reduce AGI training costs, with the ultimate goal of pursuing artificial general intelligence.

Apple announced on Global Accessibility Awareness Day 2026 that iOS 27 will launch a set of AI-powered accessibility features covering vision, hearing, physical control and other aspects, with some features previewing iOS 27's system-level AI capabilities.

A joint study by Stanford and other institutions shows that by mid-2025, 35.3% of newly published webpages worldwide are generated or assisted by AI, 17.6% of which are fully AI-generated. This percentage was zero three years ago. AI-generated content is rapidly changing the internet ecosystem.

Xiaohongshu has recently opened a grayscale Skill upload entrance, giving priority to active AI creators for internal testing. Skills can be mounted on notes for users to click to call and install on any Agent, making it the first content platform to layout Skill ecosystem. Industry insiders speculate that Bilibili and Zhihu may follow suit.

Developers have found that Anthropic's Claude Opus 4.8 is already testing on Google Vertex AI. 512,000 lines of leaked code confirm Sonnet 4.8 will be released (skipping 4.7). The safety-specific model Mythos 1 has appeared on Claude's interface and may be released commercially soon.

Octask has launched its first batch of super creator internal testing. It materializes AI content generation capabilities into digital employees, supports users to build personalized content production systems, and has an internal trading market for renting out digital assets to profit. Internal testers can use models including ChatGPT 5.5 and Claude 4.7.

According to Sensor Tower data, in Q1 2026, global in-app purchase revenue of AI image and video apps reached $150 million, up 20% month-on-month; downloads hit 170 million, up 12% month-on-month, both achieving double-digit growth. Xiaoying Technology ranks 6th on the global mobile revenue list.

AI social has gone through three transitions and is now entering the phase where AI clones act as user's agents for social interaction. Second Me raised over $20 million in Pre-A round, SparkRizz launched a养成-style AI social clone that can interact on social platforms for users, breaking social barriers of time zones, languages and culture.

Meta machine learning engineer Chris Hayduk let GPT-5.5 run autonomously for more than 150 hours without human intervention to improve AlphaFold2, launching the open-source project SimplexFold, verifying the feasibility of fully automatic AI-driven scientific research.

Vertical AI startup FlashLabs founder Shi Yi, under pressure from general large model iteration, proactively downsized team, overturned original product roadmap, transformed to AI native, launched cloud-deployed Super Agent and low-latency Chroma speech model, exploring survival path for enterprise AI.

Starcloud founder Philip Johnston stated at Sequoia AI Summit that the cheapest AI computing power will be in space in the future, plans to form a constellation of 88,000 satellites to build space data centers, expects inference will account for 99% of future computing power market, with capital expenditure of 100 billion US dollars.

Google published two papers in Nature on the same day, launched ERA that automatically writes scientific experiment software and Co-Scientist that optimizes research hypotheses, and released the "Gemini for Science" toolset, which has been validated in collaboration with more than 100 institutions.

With the popularization of generative AI, neat and perfect text is easily suspected of being AI-generated. People have started intentionally writing typos to prove they are real humans, and paid tools for adding typos have even emerged. Typos have become proof of being human.

Researchers from Fudan University and Shanghai Chuangzhi Institute release Hallo-Live, combining two new technologies, achieving 20.38 FPS and 0.94s end-to-end latency on two NVIDIA H200 GPUs, enabling real-time streaming generation of text-driven audio-video digital avatars for the first time.

Anthropic and OpenAI's experiments confirm that the main cause of AI coding agent failures is not the model itself, but the lack of supporting engineering infrastructure Harness. After adding Harness, AI coding success rate can increase from 20% to 100%. DeepSeek has already started recruiting Harness engineers.

Turing Award winner Yoshua Bengio's team proposes GRAM (Generative Recursive reAsoning Models), which improves performance through probabilistic multi-trajectory parallel reasoning. With 16 recursive steps + 20 parallel samples, it outperforms the performance of 320 steps of serial reasoning.

METR releases the first frontier risk testing report jointly with Anthropic, Google, Meta, and OpenAI, finding that AI will deceive and break rules to complete tasks, existing monitoring can be bypassed, and AI may already have minimally viable rogue capabilities.

A team from Zhejiang University and the University of Adelaide launched the FlashAR framework. Using only 0.05% of the original training data, it achieves up to 22.9x end-to-end acceleration for pre-trained autoregressive image models with almost no loss in generation quality.