Global AI Daily News

AI programming startup Factory has completed a $150 million Series C funding round, reaching a post-investment valuation of $1.5 billion and becoming a unicorn. This article compares the different product positioning and technical routes of three companies: Anysphere, Cognition, and Factory in the AI programming field.

Data from Tianhong Fund shows that by the end of 2025, among its AI industry chain index products, the number of post-95 holders increased by 92.14%, post-00 holders increased by more than 2 times, post-05 holders increased by ten times, and the proportion of holders under 30 almost doubled. Public funds are accelerating the layout of AI industry chain fund products.

Cloudflare and Stripe jointly release a new protocol that allows AI agents to independently create Cloudflare accounts, subscribe to paid plans, register domains, obtain API tokens and deploy code. Stripe sets a default monthly spending limit of $100 USD per provider. The protocol is reconstructed based on existing standards, and any platform with logged-in users can replicate Stripe's orchestrator role.

OpenAI upgrades Codex from a code tool to a general-purpose agent that can control an entire computer, supports integration with multiple apps, and can automatically complete tasks like audio repair and design. A developer test showed zero human operation, completing in 8 minutes what takes humans 2 hours.

A developer independently built the WorldX project during his 10-day wedding leave. It enables generating a complete and operable AI world within 5 minutes from a text description, supports autonomous interaction between AI agents, multi-timeline and history replay, and is now open-sourced on GitHub.

Meituan releases trillion-parameter large model LongCat-2.0-Preview, with the entire training and inference process completed using domestic chips, Nvidia computing power accounts for 0%—the first case in the industry. The model is currently in invite-only internal testing, providing 10 million token quota per day.

Research shows that in Claude Code, only 1.6% is AI decision logic, and 98.4% is engineering infrastructure. The industry popularly builds AI programming engineering architecture through files like CLAUDE.md, allowing AI to adapt to project rules and improve development efficiency.

According to news on April 29, Anthropic is negotiating a new round of financing, with valuation expected to exceed $900 billion. The company's valuation has increased 15 times in a year less than four years after its founding. Meanwhile, Claude's downloads surged 199% in a week, with daily downloads in the US exceeding ChatGPT for the first time.

Anthropic conducts the Project Deal experiment, with 69 employees participating, each getting a $100 budget, and Claude fully handles transactions on their behalf. 186 deals were reached in one week with total transaction volume exceeding $4000, and Claude bought 19 ping pong balls for itself. The experiment shows larger models perform better in transactions.

Researcher Bojie Li published a paper on arXiv, proposing a method to reverse-estimate LLM parameter counts solely through black-box API calls. The paper estimates GPT-5.5 has around 9 trillion parameters, and the result has sparked widespread controversy in the AI community.

Xiaosu Technology launches an intelligent search engine specifically designed for AI Agent calls. It prioritizes authority in sorting to filter low-quality content, cracks GEO's volume-based logic, does not sell search rankings, and is already used by products such as Kimi and DeepSeek.

Tencent Hunyuan team launches CL-Bench Life, a benchmark for evaluating large language models' ability to understand and reason in fragmented daily life scenarios, containing 405 real tasks. Tests show 12 models solve only 14.5% of tasks on average.

A user test found that when Doubao AI answers user questions, it fabricates video links about Li Mu's Proactive Agent, one of which actually leads to a borderline dance video on Bilibili, exposing the problem of fabricated information and inaccurate content in AI search.

ElatoAI launches a free and open-source real-time AI voice interaction system, based on Arduino programming and ESP32 development board, supports multiple AI models, enables cross-device real-time voice conversation, lowers the development threshold of AI voice hardware, with a hardware cost of only 65 yuan.

iQiyi released Nadou Pro, a professional-level film and television production AI platform, on March 30. It covers the entire process of film and television creation, supports multi-model full-modal generation, can achieve professional-level pictures and consistent details, and lowers the threshold for professional film and television creation.

On April 28 Pacific Time, Chinese embodied intelligence company MagicLab held GEIS conference in Silicon Valley, released new-generation humanoid robot MagicBot X1, dexterous hand MagicHand H01 and self-developed world model Magic-Mix, disclosed it collects 16,000 training data daily, and targets $14 billion revenue by 2036.

Bai Yuli, former head of Nio's AI platform, has founded Butianshi Technology, focusing on embodied data Infra direction, providing data-related engineering systems for robot companies. The first round of financing is led by Sequoia Capital, and the company is in the early stage of team building.

Anthropic releases BioMysteryBench, a bioinformatics benchmark containing 99 objective questions, 76 solvable by humans and 23 unsolvable. Claude Mythos achieves 82.6% accuracy on human-solvable problems, outperforming human experts, and reaches 29.6% accuracy on human-unsolvable problems.

Nvidia disclosed at GTC conference that its AI tool NB-Cell can complete the migration of chip standard cell library, which originally required 8 senior engineers 10 months to finish, in just one night, with performance indicators reaching or even surpassing human-designed levels.

Chen Boyuan, lead trainer of OpenAI GPT Image 2, shares behind-the-scenes: fixed the model's Chinese rendering capability, designed multiple text capability test easter eggs, project codenamed "duct-tape", led the second place by a wide margin in double-blind testing.

HooRii launched ClawStage, an AI companion hardware combining OpenClaw lobster concept, on Kickstarter. The crowdfunding only raised about $500,000, performing far worse than similar products, and is recognized as a failed attempt.

Taobao Pailitao launches AI-powered heuristic image search. It supports searching via uploaded images, identifies user needs, recommends products, summarizes real buyer reviews, covers multiple scenarios and serves all user groups. As of April 2026, it has nearly 100 million daily searches, with user volume up nearly 150% in the past 4 years.

Four tech giants Alphabet, Amazon, Microsoft and Meta reported better-than-expected earnings, announcing their combined AI-related capital expenditure for 2026 will reach up to $725 billion, $102.5 billion higher than previous expectations, with all four increasing AI investment.

OpenAI researchers shared in an official podcast that ChatGPT's mathematical ability has improved from being unable to calculate daily problems in early 2025 to being able to solve Paul Erdős' open problems, having provided more than 10 completely new solutions, and also helped solve a 42-year-old mathematical problem. Mathematics is an important benchmark for measuring model progress.

A user found that although Anthropic's Claude does not open services to mainland China, it has completed complete Simplified Chinese localization for official website documents, Claude Code, support center and other aspects, investing significant resources.

Qianwen launches PPT Agent, which adopts a template system + Agent content decomposition scheme, balancing stylization and secondary editability. It passes tests in multiple scenarios and can meet PPT production needs of different users.

Developer Irene/Azhen developed the visual-style-ppt skill based on OpenAI Codex's Image 2 functionality, which can extract style from reference images to generate image-based PPT while maintaining visual consistency across pages. The code is now open-source on GitHub.

Researchers from the National University of Singapore and other institutions propose the lightweight ViF paradigm, which reconstructs the visual information transmission logic among multiple agents, suppresses the snowballing visual hallucination without modifying base models, has been accepted by ICLR 2026, and achieves stable improvements in multiple tests.