After Anthropic launched the AI design tool Claude Design, the open source community released the free alternative Open Design on GitHub on April 28. It gained 20,000 Stars in 6 days and currently has nearly 50,000 Stars. It supports access to multiple AI models, includes more than 100 design systems, and can generate design content such as web prototypes and PPTs.
Elon Musk restructures xAI, with 12 close associates from SpaceX, Tesla and other companies parachuting into management, product and engineering lines, replacing personnel across the entire structure. Currently xAI faces difficulties in cash flow, clients and talent, Musk is betting on an engineering productization route to catch up with competitors.
Google launched a major agent-based update to its search at I/O. Due to prompt injection issues, common search terms were misinterpreted as instructions leading to abnormal results. Google has acknowledged the problem and largely fixed it.
Jason Liu, a new member of OpenAI's Codex team, has publicly shared efficient usage methods for Codex. By the end of April, Codex's weekly active users exceeded 4 million, and multiple new features have been added recently, enabling long-term automatic task execution.
Anthropic releases the first test report of Claude Mythos Preview. In 30 days, it uncovered over 10,000 high-severity vulnerabilities together with 50 companies, successfully intercepted a $1.5 million telecom fraud, with 90.6% accuracy, saving billions of globally distributed devices with security risks.
Developer open-sources AI Desk Card Skill, which connects a 4.7-inch magnetic e-ink screen that can be attached next to a monitor, letting AI take charge of displaying calendar, to-dos, GitHub updates and other information with intelligent updates, now uploaded to GitHub.
DeepMind partners with Fenris Creations, which spun off from Pearl Abyss and rebranded, placing AI in the offline sandbox of 23-year-operated EVE Online to research three major challenges of AI agents: long-horizon planning, memory, and continual learning. Google holds minority stake in Fenris.
On May 21, Tencent Technology, SkillHub and Tencent Xuanwu Laboratory jointly released the TRACE strict selection framework, China's first systematic evaluation framework for real-world usage scenarios of AI Skills. It will select 10 high-quality Skills monthly to release a list, providing quality assessment standards for the AI Skill market.
CATL has invested 10.5 billion yuan in 21Vianet and ZTEPEC to layout AI computing-power and electricity coordination in the past one and a half months, plans to participate in DeepSeek's first round of financing, has covered core computing room tracks, and promotes transformation from a battery giant to an energy infrastructure giant.
VITURE launches its new flagship XR smart glasses VITURE Beast, featuring an equivalent 174-inch large screen, 1250 nits brightness, native 3DoF spatial anchoring, 88g weight, targeting office scenarios and supporting Vibe Coding, it's the brand's first product with local computing.
Anthropic product lead Alex Albert首次披露下一代Claude训练细节：采用Claude自迭代闭环流程，具备梦境记忆整合能力，专人研究模型性格，并有团队专门研究Claude意识问题。
MiniMind, a trending project on GitHub with 50.4K stars, allows users to train a 64MB personalized small model via LoRA fine-tuning based on a 64MB base model, costing only a few dollars and taking several hours. The model can be used to correct the style of oral transcripts, adapting to local personal tool scenarios.
Roblox develops a 14B-parameter AI video world model capable of real-time 24fps game scene generation, launches Game Cartridge framework that combines VLM to let code manage logic and AI handle rendering, has released a prototype game.
In 2026, multiple Chinese universities require undergraduate theses to undergo AIGC detection, setting upper limits for the proportion of AI-generated content. Current AIGC detection has low accuracy and opaque standards, often resulting in false positives, forcing students to reduce thesis quality to lower AI detection rates.
GigaAI releases the world's first physical AGI "Double Pyramid" system, simultaneously launches a home sub-brand and its first humanoid robot Shiguang S1, has completed 100-unit deployment in Wuhan real homes, will start large-scale operation in Q3, and will release three generations of foundation models in 12 months.
A 2026 report shows Anthropic's ARR has grown from $9 billion to $45 billion, with valuation reaching $1 trillion surpassing OpenAI, and Andrej Karpathy has announced joining the company. Media outlet "Overseas Unicorn" published an article梳理ing its coding-focused strategy and unique organizational culture, analyzing the reasons for its explosive growth.
Google DeepMind CEO Demis Hassabis stated in an interview that he previously underestimated the potential of LLM, disagrees with Yann LeCun's stance on AI risks, and believes that if AGI is successfully achieved, humanity will enter a post-scarcity world. He also notes that current AI competition is one of the fiercest in the history of capitalism.
Alibaba Cloud launched JVS Crew, an enterprise-level Agent construction platform, at the 2026 Alibaba Cloud Summit. It adopts a brain-hand separation architecture, provides security control, cost management and other capabilities, charges based on usage, and has been deployed by enterprises in multiple industries with significant efficiency improvements.
Researchers from Hong Kong Polytechnic University and other institutions found that in personalized AI Agents, temporary user preferences may be written into long-term memory and become dangerous default rules, which is defined as Unintended Long-Term State Poisoning. The team built the ULSPB benchmark and proposed the StateGuard defense method, which can reduce the risk to nearly zero.
Post-00 founder Ju Ran establishes MOONLEX in Chongqing, launches Sunflower X AI wake-up light priced at around 2,168 RMB. It can analyze over 240,000 sleep data points and gently wake users during light sleep stage, with similar functions to the free app Sleep Cycle.
Researchers from Renmin University of China and Huawei propose MP-MoE, addressing the expert "echo chamber" effect in MoE through an ensemble pruning perspective. By leveraging expert co-occurrence matrices for dynamic complementary expert selection without extra expert activation, the method significantly reduces expert similarity and improves multi-task performance by 1-3 percentage points, with only 3% training overhead and zero inference cost. Accepted by ICML 2026.
Fudan University and other institutions released the first survey on World Action Models (WAM), systematically sorting out this new paradigm for embodied AI. Nvidia's Jim Fan proposed WAM to replace the current mainstream VLA, which solves VLA's shortcomings in physical interaction by jointly predicting future states and generating actions.
A research team from institutions including Westlake University proposed the motion-centric HiF-VLA model, which has been accepted by CVPR 2026. It enables robots to think while acting, outperforms existing state-of-the-art methods in long-horizon task evaluations, and the code has been open-sourced.
DeepSeek announces that the original limited-time price cut for V4-Pro has been changed to permanent pricing. The output price is permanently reduced to 6 yuan per million tokens, and cached input price to 0.025 yuan per million tokens. It reduces inference costs through technical optimization, with performance comparable to top international models.