Multiple sources reveal that OpenAI's next-generation foundational model, GPT-6 (internal codename Spud), has completed pretraining, showing approximately 40% improvement over GPT-5.4 in coding, reasoning, and agent tasks. The context window may expand to 2 million tokens, leveraging a natively multimodal architecture. OpenAI President Greg Brockman confirmed that Spud represents two years of research, positioned as the 'last mile' toward AGI. To consolidate computational resources, OpenAI has moved the Sora project into its robotics research branch and plans to integrate ChatGPT, Codex, and Atlas Browser into a unified desktop-scale super-agent application. Brockman stated that 'AGI will be achieved within the next few years,' with the company viewing compute capacity as a revenue center rather than a cost center.

Microsoft is significantly reshaping its AI strategy by establishing a 'Super Intelligence' division and a 'Frontier Models' initiative following revised agreements with OpenAI, led by AI CEO Mustafa Suleyman. The goal is to build domain-specific AI systems competitive with ChatGPT, Claude, and Gemini. In March 2026, CEO Nadella restructured the AI team to allow Suleyman to focus exclusively on model development. To support internal AI research, Microsoft is prioritizing data center compute for in-house projects over external leasing, leading to underperformance in Azure revenue and a 10% drop in stock price the following day. Meanwhile, OpenAI’s recent $122 billion funding round could further dilute Microsoft’s stake of approximately 27%. Microsoft has already launched three MAI-series in-house models (speech-to-text, text-to-speech, text-to-image), offered via Azure and Foundry.

The Cyberspace Administration of China has drafted the 'Administrative Measures for Digital Virtual Human Information Services (Draft for Public Comment)' and is soliciting public feedback until May 6, 2026. The document consists of five chapters and twenty-seven articles, defining digital virtual humans as virtual personas simulating human appearance, voice, and behavior using computer graphics, digital image processing, or AI technologies. Key provisions include requiring separate consent for using sensitive personal information of individuals and parental consent when involving minors; prohibiting诱导 of minor addiction and offering virtual intimate relationship services to minors; and safeguarding rights such as portrait, reputation, and privacy. Close relatives of deceased individuals may exercise legal rights over the deceased's personal information.

Kevin Gu from thirdlayer.inc developed AutoAgent, an open-source library that uses a meta-agent to autonomously iterate and optimize AI agents’ system prompts, tool configurations, routing logic, and execution frameworks without human intervention. Within 24 hours, AutoAgent achieved a 96.5% score on SpreadsheetBench, ranking first, and reached GPT-5-level performance with 55.1% on TerminalBench, surpassing all manually tuned approaches. The core architecture features two agents: humans only edit the program.md file defining objectives, while the meta-agent modifies agent.py and records experiment history in results.tsv for hill-climbing optimization. The system supports Harbor-format benchmarking with LLM-as-judge evaluation, applicable to any scoreable task, marking a shift from manual tuning to goal-oriented paradigms.

Multiple reports indicate that DeepSeek’s upcoming large language model V4 will run entirely on Huawei Ascend chips, marking a significant step toward decoupling in China’s AI industry at the compute level. This reflects broader efforts among Chinese AI firms to build independent technology stacks based on domestic chips amid tightening U.S. export controls. Concurrently, Meituan has internally restricted use of Alibaba’s Qwen models and is guiding business units toward its self-developed Longmao model, reflecting strategic shifts among Chinese tech giants toward model autonomy and supply chain security. However, analysts note DeepSeek V4 faces major challenges, needing to overcome technical gaps in efficiency where U.S. competitors have made substantial progress, all under constraints of limited access to high-end chips.

Foxconn Precision (Hon Hai) reported first-quarter 2026 revenue of NT$2.13 trillion (approximately $66.6 billion), a year-on-year increase of 29.68%, primarily driven by sustained growth in AI server demand. As the world’s largest electronics manufacturer, Foxconn holds a competitive edge in AI server assembly due to deep vertical integration, particularly benefiting from increased shipments of servers equipped with NVIDIA accelerators. Analysts believe ASIC server projects and deployment of the NVIDIA Vera Rubin platform could further boost growth. Chairman Young Liu warned of potential supply chain disruptions from geopolitical crises in the Middle East. The company has appointed Ji Ji Heng to succeed Yang Chiu-Chin as rotating CEO and is facing memory chip shortages, though executives say the impact on high-end products is limited.

NVIDIA is advancing photonic interconnect technology to overcome physical limitations of copper cables, aiming to build ultra-scale systems housing over one thousand GPUs by 2028. Over the past month, the company has invested billions of dollars in optical technology firms including Marvell, Coherent, and Lumentum to secure its supply chain. Current NVL72 systems still rely on copper backplanes to interconnect 72 GPUs but face limitations in reach and power consumption. Future systems like Vera Rubin NVL576 and Rosa Feynman NVL1152 will adopt optical expansion and explore co-packaged optics (CPO) for NVLink switch fabrics. CPO can significantly reduce power and latency, enabling multi-rack scaling. The next-generation Feynman architecture, expected in late 2028, may integrate CPO to create a single-layer computing network.

Multiple signals indicate the AI industry is experiencing a severe compute shortage. OpenAI's CFO said the company is forgoing commercial opportunities due to insufficient compute; Anthropic has imposed session limits on 7% of users; and H100 GPU rental prices have reached an 18-month high. AWS and Microsoft previously lost key contracts due to inability to guarantee compute availability. Meanwhile, Alibaba has shifted its leading open-source model Qwen to closed-source to protect its competitive advantage. Economists predict AI could boost annual U.S. GDP growth by 1–1.5 percentage points by 2050, but also eliminate 10 million jobs and exacerbate inequality. GitHub Copilot has rolled back a feature that inserted promotional content during code reviews after developer backlash.

Youi Robotics has officially filed for a listing on the Main Board of the Hong Kong Stock Exchange, with CICC serving as sole sponsor. The company focuses on industrial embodied intelligence, providing robotic products with perception, mobility, and manipulation capabilities for the semiconductor, energy & chemical, and manufacturing sectors. According to Frost & Sullivan, Youi Robotics ranked first in China’s mobile manipulation robot solutions market by 2024 revenue with a 7.1% share, and fourth globally; it leads domestically in semiconductor-related revenue. The company launched its humanoid robot solution in 2025 and has secured orders. The global embodied intelligence market is projected to reach RMB 367.5 billion by 2030. Proceeds will fund R&D and global expansion through production bases in Huzhou, Hefei, and Shenzhen.

On April 3, 2026, the popular multi-agent Python framework PraisonAI was found to have seven critical security vulnerabilities (CVEs), allowing full compromise within 48 hours. The most severe flaws include bypassing sandboxing by overriding the startswith method of the str class (CVSS 10.0); OAuth token validation logic errors allowing arbitrary tokens to pass verification; two unauthorized endpoints exposing agent topology and WebSocket control interfaces; and vulnerabilities in SQL injection, CLI injection, subprocess escape, and SSRF. Attackers can chain these flaws to escalate from zero permissions to full system control within one minute. The article urges all agent framework developers to immediately review sandbox validation, authentication logic, and endpoint authorization. Vulnerabilities have been patched in versions 1.5.90 to 4.5.97.