Global AI Daily News

Ding Ning, assistant professor at Tsinghua University, founded general physical intelligence company "Natural Will", focusing on embodied brain direction. The company has secured angel round investment from IDG Capital, FreeS Fund and ZhenFund, with a valuation of 4 billion RMB.

Tesla, Figure AI and other companies developing humanoid robots lack action training data. Companies like Micro1 recruit workers globally at $15 per hour, requiring workers to wear iPhone head-mounted cameras to record action videos for AI training.

Anthropic open sourced Claude's free interactive visualization feature on March 12, 2026. It can generate various interactive charts, available to free users on web and desktop, no preset themes required, supports third-party application linkage. Google's similar advanced feature requires a $200 monthly subscription.

OpenAI CEO Sam Altman publicly apologized to Tumbler Ridge, Canada, for failing to report the violent comments made by the mass shooting suspect in time. The shooting killed 8 people, and OpenAI has now strengthened its security review processes.

OpenClaw releases version 2026.4.24, sets DeepSeek V4 Flash as the default model, launches V4 Pro version. Both models support 1M token context window and are fully open-sourced, while adding Google Meet built-in plugin and other features.

United Imaging Intelligence open-sources uAI Nexus MedVLM, the world's first medical video understanding large language model, and simultaneously launches a 6,245 finely annotated test set and a leaderboard. Its surgical safety evaluation accuracy reaches 89.7%, outperforming general large models significantly, and can assist doctors in performing surgeries.

CapCut launches AI Assistant on its mobile app, users can complete batch editing, material adding, intelligent copywriting and other operations through natural language. The test completed a red carpet vlog rough cut in one minute, lowering the threshold for video creation.

IPADS research team from Shanghai Jiao Tong University open-sources SkVM, a Skill-oriented language virtual machine. It solves Skill-model adaptation issues, enables 30B small model to match Opus 4.6 accuracy, reduces 40% token consumption, speeds up up to 50x, supports mainstream Agent frameworks and Skill ecosystems.

OpenClaw founder and OpenAI engineer Peter Steinberger developed the ClawSweeper tool, using AI to close 5000 invalid issues in one day, taking only 1/365 of the time manual processing would require, with total cost under $1000, and was rate-limited by GitHub API due to its high speed.

Epoch AI and Ipsos release a survey on AI use in the U.S. workplace: 50% of American adults use AI, only 6% are heavy users; 76% of users with company-paid subscriptions use AI for work, far higher than 38% of free users and 58% of self-paid users.

Luoli Fu, head of Xiaomi's large model team, accepted her first in-depth interview, stating that AI has shifted from the Pre-train-dominated Chat era to the Post-train-dominated Agent era. A 1T parameter model is the entry ticket to approaching Claude Opus 4.6's level. The open-source OpenClaw framework has unlocked the upper limit of mid-sized models and collective intelligence. Xiaomi has launched the MiMo-V2 series models and opened public testing.

Visual AI company Collov, founded by Stanford-backed founders, has completed a $23 million Series A funding round, established Collov Labs Research. The capital will be used for R&D and team expansion. It focuses on visual AI systems and core观点 that cameras will become the new AI interaction entrance.

Dai Jifeng, associate professor at Tsinghua University, has publicly escalated his conflict with Chen Tianqiao, founder of MiroMind. Dai spoke via foreign media, MiroMind denied his claims, and the commercial dispute between the two sides has escalated into a legal confrontation.

Tsianhua University associate professor Jifeng Dai founded AI company Naive.ai, completed approximately $300 million in financing with a valuation of around $800 million, focusing on post-training of open-source models and AI Agent direction, core team includes former MiroMind members.

Anthropic Product Lead Kat Wu shared in a podcast that her team has compressed product iteration cycles from 6 months to as fast as 1 day, releasing over 30 features in 40 days, with core improvements coming from process optimization rather than the model itself.

APPSO publishes a self-reflection article pointing out that the current AI industry has the problem of phenomena replacing real experience. Companies like Anthropic frequently release new versions with improved benchmark scores but worse actual user experience, and media are involved in the collusion, boosting the cognitive bubble that affects the market.

Researchers from multiple international universities jointly released a survey paper on Feed-Forward 3D Scene Modeling, proposing a new problem-driven analytical framework that categorizes the field into five core research directions, systematically sorting out research progress and prospecting future trends.

ICLR 2026 announces award-winning papers, with 2 Outstanding Papers and 1 Honorable Mention. Two 2016 papers including Alec Radford's DCGAN receive the Test of Time Award. The conference received about 19,000 valid submissions with an acceptance rate of around 28%.

NVIDIA-backed AI storage company Vast Data has completed $1 billion financing, with valuation increased to $30 billion, expected to be ready for IPO by the end of this year. The new capital will be used for acquisitions and investments in AI ecosystem companies.

AI now generates nearly 80% of production code, but debugging remains critical. PlayerZero raised $15M Series A to develop World Model, a context graph that detects and fixes bugs before code deployment. Its Sim-1 engine achieves 92.6% accuracy across 2,770 scenarios, compressing weeks of 300-person QA work into minutes, cutting production issues by 50%, saving each enterprise customer over $2 million.

Feishu Multidimensional Table upgrades, adding four new AI capabilities: AI data query, AI chart generation, AI page building, and AI form generation, enabling non-technical users to complete data analysis through natural language, lowering the threshold of AI usage and empowering frontline workers across industries.

DeepSeek V4 Pro launches a limited-time 2.5-fold discount promotion, valid until May 5. Cached hit input drops from 1 yuan to 0.25 yuan per million tokens, cache miss input from 12 yuan to 3 yuan, and output from 24 yuan to 6 yuan. Users can experience it during the May Day holiday.