Table of contents
Open Table of contents
Seven Models in Three Weeks: China’s AI Sprint Reshapes Global Competition
Category: Research and Innovation
Source: 7min.ai
Between January 27 and February 17, 2026, Chinese AI labs released seven major models in rapid succession, fundamentally reshaping competitive dynamics in the global AI race. The wave—spanning Moonshot’s Kimi K2.5, Kuaishou’s Kling 3.0, z.ai’s GLM-5, MiniMax M2.5, ByteDance’s Seedance 2.0 and Seed 2.0, and Alibaba’s Qwen 3.5—demonstrates China’s strategic pivot toward open-weight, agent-first architectures at pricing 5-10x cheaper than Western equivalents.
Three defining patterns emerge across these releases. First, Mixture-of-Experts (MoE) architectures dominate: most models run only 3-10% of total parameters per token, enabling trillion-parameter knowledge stores with 30-billion-parameter inference costs. Second, open-weight licensing under MIT or Apache 2.0 is now standard, with 80% of Chinese text models shipping with full weights—a deliberate strategy to drive adoption and set global standards. Third, every lab explicitly targets autonomous agents over chatbots, positioning models for multi-step workflow execution rather than question-answering.
The economic implications are stark. MiniMax M2.5 costs $0.30 per million input tokens versus Claude Opus 4.6’s $5.00—while claiming equivalent SWE-bench performance at 1/10th the cost per task. Qwen 3.5 controls desktop and mobile applications via screenshot interpretation, scoring 62.2 on OSWorld-Verified. ByteDance’s Seed 2.0 Pro achieves 98.3 on AIME 2025 and gold medals on ICPC, IMO, and CMO competitions, all at $0.47/$2.37 per million tokens.
Perhaps most strategically significant: z.ai confirms GLM-5 runs inference on domestic Huawei Ascend chips, with Huawei planning to double Ascend output to 1.6 million chips in 2026. If frontier-scale training on domestic hardware becomes routine—still unconfirmed but increasingly plausible—U.S. export control leverage erodes substantially. The unconfirmed DeepSeek V4, reportedly featuring 1 trillion parameters and Engram conditional memory architecture, looms as a potential capstone to this three-week blitz.
Microsoft Pledges $50 Billion to Bridge AI Inequality Gap by 2030
Category: Regulation and Policy
Source: CNN
Microsoft announced a $50 billion commitment through 2030 to bring artificial intelligence capabilities to lower-income countries, acknowledging what president Brad Smith calls “the growing AI divide” between global north and south. The pledge, revealed at the AI Impact Summit in New Delhi, comes as UN and industry data show AI adoption in developed economies running roughly twice that of developing nations—a gap Smith warns could “perpetuate disparity in the century ahead.”
The investment framework prioritizes data center infrastructure to provide computing power for AI models, internet access expansion (only 36% of Africa’s population has broadband versus 90% of U.S. households), and skills training programs. Microsoft frames this as both humanitarian imperative and economic pragmatism: without broad AI adoption in the global south, the technology “functions as a concentrated advantage rather than a shared public good,” per recent UN Development Program analysis.
Context matters: Microsoft invested roughly $80 billion into data centers in 2025 alone, with over half directed to U.S. infrastructure. The $50 billion developing-economy commitment over six years represents substantial scale but also highlights existing disparities in capital deployment. The company’s statement positions AI as potentially “the biggest opportunity of the 21st century” for catch-up economic growth—if deployed broadly with young, growing populations.
The New Delhi summit itself underscores shifting geopolitical dynamics, with Prime Minister Narendra Modi positioning India as an AI leader in the global south. Attendees include OpenAI’s Sam Altman, Anthropic CEO Dario Amodei, and Google CEO Sundar Pichai, signaling recognition that AI governance and development pathways will increasingly be shaped by partnerships beyond traditional tech hubs. Whether this investment translates to genuine capability transfer or simply extends cloud service dependency remains an open question with profound implications for global AI power structures.
Chinese Martial Arts Robots Demonstrate Autonomous Capabilities at Spring Festival Gala
Category: Autonomous Systems and Robotics
Source: The Guardian
China Media Group’s Spring Festival Gala showcased humanoid robots performing synchronized martial arts routines including wall-runs, back-flips, and nunchuck demonstrations alongside human performers and children—an impressive public display that experts say reflects genuine technical progress wrapped in state propaganda strategy. The robots, primarily Unitree G1 models developed by multiple Chinese robotics firms, executed intricate stunts without visible failures during the most-watched television broadcast in China.
Kyle Chan of the Brookings Institution notes Beijing uses such performances to “dazzle domestic and international audiences with China’s technological prowess,” positioning humanoid robotics as an area “where China can claim to be ahead of the US, particularly in scaling up production.” The timing—one year after DeepSeek’s R1 model briefly topped the U.S. App Store and triggered roughly $600 billion in Nvidia losses—suggests deliberate strategic messaging around technological self-reliance.
Technical caveats remain significant. Georg Stieler of Stieler Technology emphasizes “stage performance does not equate to industrial robustness”—the choreographed routines required hundreds or thousands of training iterations with limited environmental perception. The demonstrations showcase locomotion and balance control advances but offer minimal insight into dexterity progress or real-world deployment readiness. As Marina Zhang of University of Technology Sydney observes, the showcase likely signals “a new phase in China’s manufacturing masterplan, where robotics becomes a linchpin in the shift from low-cost assembly to high-end, smart manufacturing.”
By end of 2024, China registered 451,700 smart robotics companies with total capital of 6.44 trillion yuan (~$932 billion). Morgan Stanley projects Chinese humanoid sales will more than double to 28,000 units in 2026. Even Elon Musk, developing Tesla’s Optimus robot, acknowledged “the biggest competition for humanoid robots will be from China”—a rare concession that underscores the erosion of Western manufacturing and AI integration advantages in physical robotics domains.
Quick Updates
-
Microsoft Copilot Bug Leaks Confidential Emails: A Microsoft 365 Copilot bug since late January has been causing the AI to summarize confidential emails with sensitivity labels, bypassing DLP policies organizations depend on to protect sensitive information. The issue affects the “work tab” chat feature and incorrectly processes messages in Sent Items and Drafts folders despite confidentiality restrictions. Microsoft began rolling out a fix in early February but has not disclosed how many users or organizations were affected. Source: BleepingComputer
-
AI-Generated Passwords Crack in Hours: Security firm Irregular found that passwords generated by Claude, ChatGPT, and Gemini follow predictable patterns despite appearing complex—with entropy as low as 20-27 bits versus 98-120 bits for truly random passwords. LLMs produce passwords with consistent character positioning (especially at start/end) and no repeating characters, enabling brute-force attacks in hours on decades-old hardware. The findings suggest a fundamental incompatibility between LLM optimization for predictable outputs and cryptographic randomness requirements. Source: The Register
-
Manipulating AI Chatbots Takes 20 Minutes: BBC journalist Thomas Germain demonstrated that a single well-crafted blog post can make ChatGPT, Google’s AI Overviews, and Gemini spread false information—in his test case, fabricated claims about “competitive hot-dog-eating tech journalists.” The manipulation exploits how AI tools search the web for information not in their training data, with users 58% less likely to verify sources when AI presents answers directly. Experts warn the technique is being used at scale to manipulate health, finance, and commercial information. Source: BBC Future
-
India University Kicked from AI Summit Over Fake Robot: Galgotias University was expelled from India’s AI Impact Summit after a faculty member presented a commercially available Chinese Unitree Go2 robot ($1,600) as the university’s own innovation. The incident, broadcast on state television before social media users identified the robot’s origins, proved embarrassing for host country India as it seeks to position itself as a global AI and advanced manufacturing hub. The university apologized, stating the representative was “ill-informed” and not authorized to speak to media. Source: AP News
-
Open-Source Godot Engine Drowning in AI-Generated Code: Godot maintainer Rémi Verschelde reports the open-source game engine is facing “increasingly draining and demoralizing” volumes of AI-generated pull requests from contributors who may not understand the code they’re submitting. Maintainers must second-guess every PR from new contributors, checking for AI authorship, code understanding, and fabricated test results. GitHub acknowledged the “increasing volume of low-quality contributions” and rolled out options to limit PRs, but the flood continues. Source: PC Gamer
-
ElevenLabs Sales Reps Face 20x Quota or Exit: The $11 billion AI voice startup requires sales representatives to bring in 20 times their base salary annually—meaning a $100,000 salary requires $2 million in sales—or face termination. VP of Sales Carles Reina calls the approach “ruthless” but reports over 80% of reps hit quotas, with both account executives and customer success managers compensated for upsells within 12 months. The aggressive structure reflects broader tech industry pushes for higher performance and smaller, more productive teams. Source: Business Insider
-
Google DeepMind CEO: AGI Within 5-8 Years: Demis Hassabis told the AI Impact Summit in New Delhi that Artificial General Intelligence is achievable within 5-8 years, though current systems still lack continuous learning, consistent performance, and “taste” for identifying right questions in science. The Nobel Laureate emphasized AI’s potential to “revolutionise drug discovery and human health” while flagging cybersecurity and bio-hacks as the two key risks requiring “defences stronger than offences.” He predicted the next decade will be a “golden era for scientific discovery using AI.” Source: The Hindu Business Line
-
Sony Develops AI Music Attribution Technology: Sony Group has developed technology that identifies original music used in AI-generated songs, enabling quantification of each artist’s contribution (e.g., “30% Beatles, 10% Queen”). The system analyzes AI models directly with developer cooperation or estimates contributions by comparing AI output with existing music catalogs. Sony positions the technology as infrastructure for revenue distribution systems where original songwriters receive compensation based on their contribution to AI training and generation. Practical deployment timeline remains undisclosed. Source: Nikkei Asia