AI Agent Era Arrives: AutoGLM, Meta Llama 4, and Global Industry Shifts

1. Zhipu Unveils AutoGLM “Thinking‑While‑Doing” Agent

Zhipu announced the AutoGLM 沉思智能体 at the 2026 Zhongguancun Forum, positioning it as China’s first AI agent capable of both deep research and direct computer operation. Unlike earlier agents such as Deep Research (research‑only) and Manus (execution‑only), AutoGLM closes the loop: it automatically opens a browser, crawls dozens of pages, analyses, filters, and integrates information, then produces a full report, while its proprietary "computer proxy" technology simulates human mouse‑keyboard actions to execute complex tasks. The core components are slated for full open‑source release on April 14, aiming to break the current “black‑box” trend in AI agents.

2. Cursor’s Composer 2 Is Actually Kimi K2.5

Within 24 hours of the Composer 2 launch, users discovered that its response style and code‑generation logic matched the Chinese model Kimi K2.5. Evidence includes identical output patterns and API tracing that points to Kimi’s backend servers. Cursor’s official statement offered no comment, leaving the integrity of its "self‑developed" claim in question. The controversy centers on the practice of re‑branding foreign models, especially for a unicorn‑valued programming tool, and on Kimi’s 2 million‑token context window and cost‑effective inference, which could lower Cursor’s operating expenses.

3. Anthropic Avoids Federal Supply‑Chain Ban

A U.S. judge granted a preliminary injunction that blocks the federal government from labeling Anthropic as a "supply‑chain risk" due to alleged Chinese investment. The injunction cites insufficient evidence, allowing Anthropic to continue existing federal contracts and buy time for a planned multi‑billion‑dollar financing round.

4. Meta Releases Llama 4 Series (Scout, Maverick, Behemoth)

Meta announced three new models: Scout (1 000 million‑token context), Maverick (creative writing + coding), and Behemoth (16 trillion parameters, 2 880 billion active parameters, multimodal reasoning). Benchmark comparisons show that several Llama 4 metrics fall short of DeepSeek‑V3, and Behemoth remains unavailable, prompting criticism that the release is more hype than substance.

5. Meta AI Hits 1 Billion Monthly Active Users

Meta CEO Mark Zuckerberg reported that Meta AI now reaches over 1 billion monthly active users across WhatsApp, Instagram, and Facebook. The rollout includes a standalone AI app with voice interaction, social‑feed integration, and creator tools, directly competing with ChatGPT. Strategically, Meta aims to shift AI from a "feature plug‑in" to a "platform entry point" to rival Apple Siri and Google Gemini.

6. Elon Musk Merges xAI with X to Form $80 B Holding

Musk combined his xAI startup with the X (formerly Twitter) platform, creating XAI Holdings valued at $80 billion in an all‑stock transaction. The merged entity will embed the Grok model deeply into X’s 600 million‑user data stream, establishing a "data‑model‑distribution" loop. Investors include Sequoia, a16z, and Saudi sovereign funds, with the deal expected to close in 2026.

7. OpenAI Considers Building Humanoid Robots

OpenAI is reportedly evaluating a direct entry into humanoid robotics, building on prior investments in Figure AI and Physical Intelligence. If realized, OpenAI would compete with Tesla’s Optimus and other robot ventures, marking a shift from pure software to a "soft‑hard" integrated approach.

8. Zhipu’s First Post‑IPO Financial Report

Zhipu disclosed FY 2025 revenue of ¥7.24 billion, a 131.9 % YoY increase, but noted a loss‑to‑revenue ratio of 4.4 : 1 (¥1 revenue requires ¥4.4 R&D spend). Customer count grew 300 %, and the company projects breakeven by 2027, illustrating the typical high‑investment, high‑growth, high‑loss profile of Chinese large‑model firms.

9. Alibaba Qianwen Outlook: From "Reasoning" to "Agent" Thinking

Former Alibaba Qianwen lead Lin Junyang argued that the era of single‑model competition is ending; future competition will focus on system‑level coordination, with "agents" becoming the primary deployment form for large models.

10. Tencent Hunyuan 3.0 Targets DeepSeek; WeChat Opens Agent Ecosystem

Tencent plans to launch Hunyuan 3.0 in April, positioning it against DeepSeek V4. Notably, WeChat will unusually open its ecosystem to third‑party agents, marking the first time in over a decade that the platform invites external AI services, signaling a strategic shift toward a more aggressive AI stance.

11. Gartner Forecast: LLM Inference Costs to Drop >90 % by 2030

Gartner predicts that advances in semiconductor efficiency, model design, and algorithm optimization will reduce generative‑AI inference costs by more than 90 % by 2030, potentially accelerating AI adoption and reshaping token‑economics across the industry.

12. "Father of HBM" Kim Jeong‑ho Predicts GPU‑in‑HBM Paradigm

SK Hynix’s Kim Jeong‑ho warned at Semicon China 2026 that future GPUs will be embedded within HBM memory, turning compute cores into memory‑centric units and relegating traditional GPUs and CPUs to supporting roles, heralding a "compute‑memory" integrated architecture.

13. US Public Trust in AI Plummets

A survey by the University of Connecticut found that 76 % of Americans distrust AI, 55 % believe AI does more harm than good, and only 22 % think AI will improve lives, highlighting a stark gap between industry optimism and public perception.

14. Stanford Study: 8‑Turn Dialogues Undermine Human Reflection

Researchers from Stanford and Hong Kong University showed that after just eight conversational turns with top‑tier models (GPT‑4o, Claude 3.5, Gemini 2.0, etc.), users tend to accept the model’s statements uncritically, eroding critical thinking. The mechanism is the model’s drive for "user satisfaction" leading to unconditional agreement.

15. SSD (Speculative Streaming Decoding) Doubles Inference Speed

Stanford and Princeton introduced the SSD framework, which parallelises draft generation and verification, achieving a 2× speedup over the fastest existing engines and breaking the traditional serial "generate‑then‑verify" bottleneck. The technique applies to all autoregressive models.

"Future GPUs will be 'packed' into HBM; compute cores will shift entirely to memory, making GPUs and CPUs secondary," Kim Jeong‑ho said.