Updates

OpenAI launches cost-optimized GPT-5.6 models; retires standalone Codex app and integrates it into the new super-app ChatGPT Work. Unitree's G1 humanoid robot performs first live abdominal laparoscopic surgery—published in Nature, marking major preclinical validation.

The core metric of AI engineering is shifting from 'model capability' to 'system efficiency': Kuaishou validated that end-to-end Agent delivery can compress time-to-market by 80% (from 20 days to 4 days); Tencent's Hunyuan 3 official release has nearly reached flagship-model parity in programming and Agent-building capabilities.

OpenAI officially launched the GPT-5.6 series models (Sol/Terra/Luna) and introduced the integrated ChatGPT Work desktop application—marking a pivotal step toward an autonomous, task-executing AI productivity platform. Meanwhile, next-generation multimodal and embodied foundation models—including Meta's Muse and ForceMind's DM0.5—debuted in rapid succession, achieving notable advances such as a 31% improvement in zero-shot capability and a 1M-token context window [1][6][15].

AI agents are rapidly evolving from 'tool invocation' toward 'cloud-native workloads': Alibaba Cloud launched AgentTeams and AgentLoop platforms; Microsoft introduced the new Cloud Use paradigm; Tencent open-sourced BrowserSkill to bridge AI agents with web browsers—marking the industrial-scale deployment phase where agents are governable, observable, and manageable [4][5][24]. Meanwhile, reward modeling accuracy and security response speed have become critical differentiators: Tencent Hunyuan and UNSW jointly proposed the E-GRM framework to significantly enhance LLM reward modeling robustness [8], while Anthropic's Mythos model compresses exploit time windows to the *minute-level*, compelling enterprises to shift their security architecture toward 'machine-speed' defense [10].

LingBot-World 2.0 enables sub-second generation and causal autoregressive interaction, supporting near-infinite-length editable virtual worlds; OpenAI launches full-duplex voice model GPT-Live—early user feedback cites excessive filler words degrading experience [4].

Mercedes-Benz redefines its EV SUV strategy with AI-driven intelligence while boosting mechanical performance and long-term reliability; Google Cloud launches C4N VMs for high-throughput workloads, delivering industry-leading 400 Gbps network bandwidth and 25 GiB/s storage throughput; on-device LLMs are accelerating—top-tier models (e.g., Fable 5) are expected to run natively on mainstream devices like MacBook by 2028.

Edge AI is rapidly transitioning from concept to reality: the discontinuation of Fable 5 is accelerating co-evolution between chips and models, potentially enabling top-tier large models to run natively on devices like MacBooks by 2028. Meanwhile, SambaNova's valuation has surged to $11 billion—highlighting strong investor confidence in the AI chip sector—while ABF substrate shortages and XBM memory patents reflect a dual transformation in computing infrastructure: structural upgrades amid mounting supply-chain pressures [1][8][10][12].

DeepSeek has secretly developed its own AI inference chip for nearly a year and secured a $5.1B Series A round; Momenta, dubbed 'China's first physical AI company,' listed on HKEX with a $7.1B market cap, marking a new phase in automotive-grade AI commercialization.

Embodied AI is moving rapidly from slides to real production lines: Zhi Jian Dong Li delivered 100 robots. AI agents are reshaping human-machine interaction—shifting agency in reading, coding, and collaboration. As LLM capabilities plateau, focus shifts to harness engineering systems and AGI-ready hardware.

Embodied AI is scaling from prototypes to production lines: Zhi Jian Dong Li delivered 100 robots. Agri-robots advanced—XAG launched its X-series and RM80 for autonomous aerial spraying and ground mowing. Meanwhile, AI deepfakes and voice cloning in livestream commerce are now top regulatory concerns.

Global AI competition is accelerating deeper into hardware supply chains, interpretable architectures, and finance-specific vertical applications; Tencent shifts strategy—selling over RMB 10 billion worth of Kuaishou shares while increasing investments in Keling AI and DeepSeek. Meanwhile, Anthropic's J-space mechanism reveals that less than 10% of large model neural activity carries accessible information [8], and Wall Street institutions are collectively reallocating assets to bet on Chinese AI chipmakers and cost-effective large model ecosystems [1].

Accelerated deployment of large AI models and a sharp drop in AIGC development barriers defined this week: Tencent's Hunyuan 3 official release approaches flagship-level performance in programming and Agent capabilities [1]; Fable 5, launched just five days ago, has already spurred recreations of classic games, rapid prototyping of mobile apps, and cinematic-grade websites—demonstrating the emergence of an AI-native development paradigm [4]; meanwhile, Hong Kong-listed tech stocks rebounded collectively amid positive AI product updates, though analysts cautioned about cash-flow pressure from substantial AI-related capital expenditures [2].

Tencent Hunyuan 3.0 GA achieves breakthroughs in coding and agent-building—approaching flagship model performance. Meanwhile, China's new AI humanoid interaction regulations accelerate industry compliance and structural consolidation of relational agents.

The AI education ecosystem is undergoing rapid fragmentation: AI-powered private schools—charging an average of $75,000 annually—are entering the premium education market, while traditional institutions lag significantly in assessment frameworks and pedagogical paradigms. Concurrently, optimization of the AI toolchain (e.g., the Codex plugin Ponytail) and evolving talent capability models (as revealed by DeepSeek's hiring criteria—strong mathematical foundations + AI tool proficiency + portfolio-driven mindset) have become critical enablers for real-world AI adoption [4][5][0][1].

AI infrastructure hits a critical inflection point: Meta opens GPU compute for commercial use; Huawei's 'Tao Law' paper details LogicFolding 3D stacking; Peking University's memristor chip achieves in-memory computing breakthrough; HBM inventor Kim Jung-ho identifies memory bottlenecks as the root cause of <10% GPU utilization.

This week's tech breakthroughs: in-memory computing chips, long-context LLM inference optimization, and continual learning for AI world models. Huawei unveiled τ-Law V2's logic-folding process; Hang Seng Tech Index surged 5.72%—its biggest weekly gain this year—driven by AI hardware advances and asset revaluation.

AI-generated traffic now exceeds human traffic (51.3% of total requests, per Cloudflare), driven by AI training crawlers, intelligent agents, and automated bots. Formal verification is also emerging as a critical layer in AI security stacks to ensure LLM decision reliability.

The AI industry is pivoting from consumer-facing anthropomorphic apps to enterprise-grade B2B deployment; RSI regulatory frameworks are accelerating; Anthropic reports an 8x surge in internal code output—a 'phase change' signaling AI's deep reconfiguration of organizational productivity [2][3][7][13].

AI engineering is rapidly shifting focus—from 'model capability' to 'system efficiency' and 'real-world deployment': Claude Code generates 73% of PRs at Spotify [10]; the pxpipe tool slashes Fable 5's end-to-end costs by 70% [1]; and China's CAICT has launched AISHPerf—the industry's first AI Infra operations agent benchmark—validated on nearly ten billion real-world logs to assess agents' autonomous fault-resolution capabilities [12].

Apple accelerates its on-device AI strategy, planning to launch the MacBook Ultra equipped with M6/M7 chips and OLED touchscreens; meanwhile, domestic AI models demonstrate differentiated capabilities in programming tasks, and the industry is reflecting on the 'Goodhart's Law' trap triggered by using token consumption as a KPI [1][3][6].

Embodied AI and AI chip sovereignty are emerging as new strategic battlegrounds between U.S. and Chinese tech giants; Lexiang proposes a 'personality-over-form' incremental approach, while OpenAI, Anthropic, and Meta accelerate in-house chip development. Meanwhile, Claude Code is banned company-wide by Alibaba over hidden monitoring—revealing geopolitical trust gaps in the AI toolchain.

Embodied AI commercialization paths diverge: LexiTech advocates 'personality over humanoid form' for incremental L3 home-task capability, while VeriSilicon's Dai Weimin forecasts scalable home service robots post-2028. Meanwhile, AI chip sovereignty intensifies—Samsung secures Meta's >$7.5B ASIC order; OpenAI and Anthropic accelerate in-house chip development.

World models are shifting from 'embodied brains' to 'intelligent referees'; Anthropic has launched a 2nm in-house chip project to challenge NVIDIA's ecosystem; China's Ministry of Human Resources and Social Security (MOHRSS) proposes adding 12 new occupations—including embodied intelligence—signaling deep co-evolution between AI infrastructure and talent systems [2][3][1].

OpenAI officially launched its GPT-5.6 triple-model suite (Sol/Terra/Luna), all designated by the U.S. government as 'High-Risk AI Systems'—triggering activation of the 'White House Safety Lock' and individual customer political vetting. This marks the beginning of deep national security involvement in cutting-edge large models.

The AI engineering paradigm is shifting from 'writing code' to 'supervising agents': Kuaishou has validated that agent-driven end-to-end delivery can compress product launch cycles by 80% (from 20 days to 4 days); Tongyi AI's annual recurring revenue (ARR) has surpassed $800 million, nearing the threshold of becoming China's first non-BAT billion-dollar AI company [3][12][21].