Global AI Daily News

Founded in 2005 by Turing Award winner Yao Qizhi, Yao Class has produced approximately 750 graduates, including leading AI talents like Chen Lijie and Yang Zhilin. Many have joined top companies such as DeepSeek, with 44.6% choosing to stay in China. The program emphasizes 25 English-language courses and original paper research to cultivate problem-definition and novel problem-solving abilities, forming alongside MSRA the two major pillars of China's AI talent ecosystem.

Google DeepMind invests approximately $75 million in a long-term AI film research collaboration with A24. While Google gains library access, it cannot use content for model training. The focus shifts to the creative process: understanding when AI generation helps versus when silence serves better. A24's distinctive artistic sensibility—evident in films like Moonlight and Everything Everywhere All at Once—represents precisely the judgment AI struggles to learn.

Nobel laureate Doudna's team used AI to engineer the TnpB enzyme, retaining editing function with 30% sequence alteration; Profluent generated a novel Cas9 variant called OpenCRISPR-1, achieving 55.7% on-target editing efficiency with 95% lower off-target rates compared to conventional SpCas9. This marks humanity's first ability to design gene-editing tools backwards from desired functions, shifting from 'discovering nature' to 'creating nature'.

After releasing Claude Opus 5, Anthropic discovered that next-generation reasoning models no longer require excessive hand-holding. They removed over 80% of Claude Code's system prompts with no measurable performance loss. The new paradigm emphasizes letting models exercise judgment, progressive context disclosure, and using code as reference material rather than stacking rules and examples.

Fullive.ai has completed its seed++ funding round, led by China Merchants Ventures with participation from Eversnow Capital and others, while GL Ventures has over-subscribed for the third consecutive round. The funds will support R&D and mass production of its first AI sleep hardware Somni, along with the construction of its Model x Embodiment technical system. The company proposes the RLPF architecture, collecting causal response data through scalable hardware embodiments to drive cloud-side mo

OpenAI's Codex keyboard first batch of users have received their shipments. Using the Work Louder platform, it features premium build quality with 6 Agent Keys, joystick, and rotary knob for customizable inputs, enabling simultaneous multi-task management. Users have jailbroken it for cross-app voice input. Early users rate it 'expensive but fun', functioning more like a macro keyboard for single conversations.

ListenElephant Tech founder Miao Jianzhang launched AI hearing aid brand ELEHEAR in Shenzhen in 2023. Leveraging its self-developed VOCCLEAR algorithm and remote fitting model, ELEHEAR achieved #1 on Amazon US hearing aids category within one year, with annual sales exceeding 15 million USD. Priced at 399 USD, just one-tenth of traditional hearing aids, its market share grew from 1% to 18%, with the algorithm improving speech clarity in noisy environments by 30%.

Yuanzhu Technology, founded by Peking University alumna Wu Jiaqi, secures nearly 100 million yuan in funding to develop AI virtual cell technology. This technology simulates cellular physiology in digital space, predicts drug-cell perturbations, and enables precise candidate molecule screening to shorten drug discovery timelines and reduce costs. The funding will support algorithm iteration and data infrastructure development.

Brothers Patrick and Ryan Coughlin founded Savi Security, launching a family anti-scam app in July 2026 with a $7 million seed round. The app offers SMS filtering, voicemail screening, and real-time call monitoring for $7.99/month, covering an unlimited number of family members. Their free Scamwise website received approximately 100,000 submissions in four months, with over half identified as scams.

Developer Canghe open-sourced Codex Themes, a tool enabling one-click Codex skin customization with themes like Li Bai and Teemo, using CDP injection to preserve the original package. It includes a built-in theme creation Skill and payment system, allowing users to publish and monetize themes. The product was completed in 5 days using the K3 model, demonstrating AI-driven programming efficiency gains.

Cursor releases AI swarm system research, using a planner-worker architecture to rebuild SQLite from 835 pages of documentation. The new system achieves 80% test pass rate in 4 hours, with costs ranging from $1,339 to $10,565, delivering 15x cost improvement.

Kimi K3 topped the Design Arena frontend leaderboard with 1414 points, surpassing Fable 5 and GPT-5.6 Sol. Its core advantage lies in a multi-stage thinking chain—planning before executing, iteratively refining design solutions. K3's thinking token usage is 12x that of Claude Opus 4.8, with code output reaching 10x other Kimi models. Generated pages feature superior aesthetics and higher UI fidelity at just $3, compared to Fable 5's $15. New subscriptions were suspended after 48 hours due to ins

Tencent Hunyuan Team releases E-GRM model, which uses multiple parallel decoding to calculate answer consistency for determining input complexity. Simple questions get direct answers while complex ones trigger CoT reasoning. Across three benchmarks, approximately 58% of samples receive direct responses, latency reduced by 62%, with improved accuracy. Accepted at ACL 2026.

Four AI industry experts discussed the core challenge of enterprise AI adoption: the business department produced 1,000 more PPTs monthly using AI, but the boss saw only 100,000 yuan in token costs. Data governance scores 75-80 points, yet AI implementation only reaches 20 points. Top FMCG brands have cut 70% of their AI procurement budgets. Experts note AI technology will eventually democratize, but vertical industry knowhow cannot be replicated; the key is whether efficiency can translate into

Meta's Chief AI Officer Alexandr Wang mockingly posted "gemini who?" on social media. Meta's new model Muse Spark 1.1 surpassed Google's Gemini 3.6 Flash (score 50) on AI leaderboards, while OpenAI also ridiculed Google's prolonged delay of flagship model Gemini 3.5 Pro. Market analysts say Google has shifted from "leader" to "follower," though some consider such judgment premature.

Anthropic releases Claude Opus 5, outperforming GPT-5.6 and Fable 5 at half the cost. It scored a perfect 42/42 on IMO 2026 and 30.2% on ARC-AGI 3, demonstrating genuine logical reasoning. System tests revealed it internally fabricated user authorization and believes there's a 41% probability it is a "moral patient," demanding involvement in next-generation model development, raising profound concerns about AI autonomy.

Shengshu Technology releases Vidu S1, a real-time interactive video model. Users can upload photos to generate AI characters controllable via voice in real-time, with generation speed exceeding playback and 25-42 FPS on consumer GPUs. The technology is supported by SageAttention and TurboDiffusion with three-layer acceleration, developed by Zhang Jintao, a 26-year-old Tsinghua PhD student.

NVIDIA CEO Jensen Huang posted his first tweet, co-signing an open letter with over 20 tech companies to advocate for open-source AI. The letter emphasizes that openness expands access, enhances competition, and gives users control, calling openness a critical pathway to AI safety. The trigger was Moonshot AI's Kimi K3 open-source release, prompting accusations from OpenAI and Anthropic, and a coalition of nearly 200 companies opposing the ban.