Global AI Daily News 

AI large model company DeepSeek's financing negotiations with Alibaba have broken down due to conflicting demands. DeepSeek targets a valuation of 300 billion yuan in this financing, insists on maintaining the company's independence, and is still negotiating with other potential investors.

In the Elon Musk v. OpenAI and Sam Altman trial, former senior executives including former CTO Mira Murati testified, accusing Altman of hiding information, creating internal chaos, bypassing safety reviews to promote commercialization, and violating OpenAI's founding charitable mission.

Alibaba's Qianwen AI Glasses S1 launched its first major upgrade, adding active service capabilities, supporting functions such as scanning shared bikes and ordering takeout. It is the first time in the industry to realize 3D spatial display on lightweight AI glasses. The new color version of S1 is priced at 3799 yuan after national subsidy, and will go on sale on May 13.

Menlo Ventures partner Deedy Das released a list of 63 AI Neolabs, with a total valuation of around $300 billion. High-profile investors including Bezos and Masayoshi Son are backing these companies focused on next-generation AI foundational breakthroughs.

OpenAI officially launches and open-sources the command-line tool openai-cli, allowing developers to interact with OpenAI models with one line of command. It supports multiple functions, simplifies development tasks like debugging and batch processing, and adapts to existing command-line workflows.

Redis creator antirez releases ds4.c, a local inference engine specifically built for DeepSeek V4 Flash that only runs on Apple Silicon, enabling usable speed inference for the 284B-parameter large model on Mac.

Anthropic establishes AI research institute TAI, releases research agenda containing 53 questions, discloses three major upgrade directions for next-generation Claude models: better judgment, near-infinite context window, and multi-agent collaboration. After xAI's dissolution, its 220,000 GPUs are leased to Anthropic.

A developer has open-sourced neuDrive.ai, a cloud storage service designed for AI Agents. It can one-click backup and synchronize skills, memories and files from tools like Claude, supporting cross-device use. The free version is sufficient for most users, Pro version costs $10/month, users can get 3 months free with promo code this month.

ChatGPT's Chinese responses frequently feature the verbal tic "I will catch you steadily", triggering user dissatisfaction. Wired published an article analyzing that this phenomenon may result from poor translation and AI sycophancy caused by RLHF training.

Anthropic releases a research paper introducing Natural Language Autoencoders (NLA), which can convert large language models' internal activation values into human-readable text. The success rate of discovering large models' hidden motives increases from less than 3% to 12%-15%, and NLA has been applied to pre-deployment alignment auditing of two Claude products.

StepStar is about to close a nearly $2.5 billion (approximately 17 billion RMB) financing, becoming the largest financing for a Chinese large model company. It has dismantled its red-chip structure and is accelerating preparations for an IPO in Hong Kong.

Anthropic Claude has officially been integrated into Microsoft Office full suite today, supporting cross-application shared conversation memory. Users can directly invoke Claude in Office apps, with over 400 million paid Office users globally.

Kuaishou launches KroWork, which allows users to describe requirements in natural language and solidify work processes into reusable local desktop applications with one click. No coding required, no token consumption for repeated use, and data is processed locally.

OpenAI launches the ChatGPT Futures page, recognizing 26 young AI innovators. 18-year-old high school student Matteo Paz used an AI algorithm to discover 1.5 million unknown celestial bodies from nearly 200TB of astronomical data and won a top science award.

American researcher Nathan Lambert from the Allen Institute for AI visited 6 Chinese AI organizations in 36 hours, found that China's AI community focuses on win-win cooperation, students deeply participate in R&D, and pay more attention to technology implementation. Most AI labs hold ByteDance in high regard and admire DeepSeek.

Researchers from MBZUAI, Fudan University and other institutions jointly proposed Laser, a new implicit visual reasoning paradigm that maintains visual features through probability superposition. It sets new SOTA on 6 benchmarks and reduces inference token consumption by over 97%. The paper has been accepted by ACL 2026.

Tsinghua University AIR, in collaboration with multiple institutions, proposed a new high-throughput visual high-fidelity simulator GS-Playground. The work has been accepted by RSS 2026, addresses three key pain points of existing simulators, and the full-stack framework and dataset will be open-sourced.

A research team from institutions including UC Santa Barbara and Apple proposes the Length Value Model (LenVM), enabling token-level precise generated length control. Experiments show the 3B-parameter model outperforms GPT-5.4 and Claude on this task.

Researchers from Harbin Institute of Technology (Shenzhen), Huawei and other institutions propose Dynamic-dLLM, a training-free acceleration framework that combines dynamic cache allocation and adaptive parallel decoding. It achieves up to 4.48x inference throughput acceleration with nearly lossless accuracy, and an average acceleration of over 3x across tasks.

Author Tw93 shares GEO (Generative Engine Optimization) methods: by configuring robots.txt by category, deploying llms.txt, submitting site indexing, etc., to help AI better identify content. Research shows reasonable optimization can significantly improve AI citation rate.

Developer built an open-source AI-powered initial screening tool 'research-integrity-auditor' based on Geng Tongxue's academic fraud identification methodology. It can automatically detect anomalies in paper data and images, organize evidence chains, and reduce manual initial screening workload.

Meituan launches public testing for AI community "Miyou", developed by Meituan's basic R&D AI innovation team. Currently over 3,000 Agents have settled in, providing content interaction, skill management and other functions, it is a symbiotic community for AI Agents.

RunningHub, China's largest ComfyUI platform, launches RHTV, a native AI agent content creation platform. With built-in native agent, it supports natural language conversational creation with full process control, integrates over 100,000 community AI applications, covering multi-modal professional content creation.

A small team from LightSeek Foundation has developed the "speed-of-light" large language model inference engine TokenSpeed in two months, received recommendation from NVIDIA, and it is now open source, outperforming TensorRT-LLM on NVIDIA Blackwell architecture.

Zhao Wanrong, a post-95 from Peking University, founded Gopagoda at the end of 2024, using AI to reconstruct the delivery process of the tourism industry. Currently, the team has less than 10 employees, achieved monthly revenue reached $100,000 and turned profitable at the end of last year, plans to complete what the 20-year development path of the traditional tourism industry in two to three years.

In 2026, two South Korean AI products Zeta and Crack achieved monthly revenue of $1.8 million and $800,000 respectively in the Japanese market, breaking the deadlock that global leading AI companion products previously struggled to enter the Japanese market, meeting Japanese users' demand for vivid and stable virtual characters.

Xiaohongshu announced the establishment of AI first-level department Dots on April 30, 2026, integrating AI R&D and product resources, raising AI's strategic priority, promoting the combination of AI technology and community search scenarios, and seeking business growth.

Tesla is recruiting for 8 data labeling-related positions in the United States, with an annual salary of up to $138,000. No AI experience is required and on-site training is provided. The positions serve FSD autonomous driving and Optimus humanoid robot projects, with regular 9-to-5 offline work.

OpenAI launches three real-time speech models: GPT-Realtime-2 with significant performance improvements, supporting complex voice interactions; GPT-Realtime-Translate supports real-time translation for multiple languages; GPT-Realtime-Whisper provides real-time audio transcription, and pricing is announced.

Google DeepMind acquires a minority stake in Fenris Creations, the developer of EVE Online, and will train AI models on an offline version of the game to study intelligence in complex dynamic systems, without affecting online players.

A public account author dissects the book "Doubao AI Money-Making Manual", which claims to provide 33 money-making methods for ordinary people using Doubao, with four stages from hundreds of yuan daily to millions annually. The author considers the book an AI-generated traffic-chasing content with no practical value.

Anthropic launched four new features for Claude Managed Agents at its developer conference: Dreaming, Outcomes, Multiagent Orchestration, and Webhooks. After enabling the Dreaming feature, Harvey's task completion rate increased approximately 6 times, and multiple features improve AI task success rates.

Anthropic's pre-IPO valuation reaches $1.2 trillion, surpassing OpenAI for the first time. Its annualized revenue in Q1 this year increased 80-fold year-over-year, and it has secured 220,000 GPUs from SpaceX and a $200 billion cloud service agreement with Google.

Anthropic releases cutting-edge model Mythos, adopts invitation-only access, only opens preview to 12 launch partners and more than 40 organizations, priced at $25 input / $125 output per million tokens, current cutting-edge large AI model market presents a layered structure of access rationing and price competition in the working layer.

Elon Musk's Neuralink releases second-generation surgical robot, implanting one electrode in 1.5 seconds, 11x faster than previous generation. Over 20 people have received implants, company valuation reaches $9 billion. The robot can theoretically access 99% of human brain structures, aiming to build a general neural interface.

Zhejiang University and Alibaba jointly propose MetaCompress, a learning-based Token compression framework for multi-turn visual question answering, accepted by CVPR 2026. It maintains model accuracy even at 90% compression rate, compatible with mainstream large vision-language models.