## Weekly Overview - OpenAI launched the GPT-5.6 series (Sol/Terra/Luna) and retired the standalone Codex app, integrating its capabilities deeply into ChatGPT Work desktop application; ChatGPT Work is now positioned as a 'deployable productivity hub', signifying AI's definitive evolution from conversational interfaces to autonomous task-execution platforms. - The world's first humanoid robot—the Unitree G1—successfully performed laparoscopic cholecystectomy on a live pig and was published in *Nature*. However, the entire procedure was remotely operated by human surgeons, revealing that embodied intelligence remains at the critical threshold of 'strong execution, weak autonomy'. - Apple formally sued OpenAI for 'systematic theft of trade secrets', accusing former Apple executives of bringing hardware-sensitive confidential information to OpenAI for use in competing AI endpoint development—escalating large-model competition into legal and ecosystem-sovereignty territory for the first time. - China's fully domestic, 100,000-GPU-scale supercluster 'Sugon 8000 (Dengfeng)' officially entered operation, achieving full-stack autonomy across chips, frameworks, and scheduling systems—marking a pivotal inflection point for trustworthy, large-scale deployment of China's foundational AI infrastructure. - StepN (Jieyue Xingchen) unveiled Step AOS—the world's first agent-native operating system—and introduced the 'Three-Wall Trust Framework'; simultaneously, Agent Runtime has emerged as the next critical bottleneck, with consensus forming around an 'AWS moment for the Agent era'. - SK Hynix's $26.5 billion U.S. IPO set the second-largest record in history, signaling the full onset of the HBM memory supercycle; Powerchip's DRAM foundry pricing surged 45%, with clear supply-demand imbalance projected through 2027. ## Hot Topics 1. OpenAI Launches GPT-5.6 Series and Integrates Codex into ChatGPT Work Desktop App https://www.bestblogs.dev/status/2075288698585809214?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Core Insight: GPT-5.6 is not merely a model upgrade—it represents a paradigm shift in agent execution via the Sol/Terra/Luna tri-model architecture: Sol specializes in high-risk operations (e.g., file manipulation), Terra orchestrates cross-application workflows, and Luna handles long-horizon planning. The standalone Codex app was retired, with its functionality deeply embedded into ChatGPT Work to form a closed-loop workflow of 'plan–execute–review'. — Actionable Next Steps: Individual developers should immediately download the ChatGPT Work desktop app (macOS/Windows) and test Sol's Ultra-mode permission boundaries using commands like `os.execute("rm -rf ~")`; product teams must complete MCP (Model Control Protocol) compatibility validation for existing automation pipelines within 72 hours—specifically verifying whether `tool_use` response structures align with the new `tool_call_id` mechanism. 2. Unitree G1 Humanoid Robot Performs Laparoscopic Cholecystectomy on Live Pig, Published in *Nature* https://ww Core Insight: This marks the world's first standardized laparoscopic surgery inside a live animal performed by a humanoid robot—demonstrating the physical feasibility of embodied intelligence in high-precision, low-tolerance medical settings. Yet the entire procedure was teleoperated by human surgeons; the robot did not achieve end-to-end perception–decision–execution autonomy—making this a milestone of 'strong execution, weak autonomy'. — Actionable Next Steps: Medical AI developers should immediately apply for access to Unitree's open SDK (https://developer.unitree.com) and replicate basic instrument manipulation sequences using its `UFO` control protocol; product teams may partner with Class-III hospitals' surgical departments to reproduce three standardized laparoscopic motions (grasping, traction, electrocautery) using G1, collecting >500 failure samples to build a 'teleoperation-to-autonomy transition' training dataset. 3. Apple Sues OpenAI for Trade Secret Theft in AI Hardware Development https://www.bestblogs.dev/status/2075712647723397452?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Core Insight: The lawsuit directly traces the movement of hundreds of employees—including former Apple executive Tang Tan—alleging OpenAI systematically replicated Apple's AI hardware design documents, sensor fusion architectures, and edge inference frameworks. Beyond delaying OpenAI's Project H hardware timeline, this suit will compel all AI firms to overhaul talent compliance auditing processes. — Actionable Next Steps: AI hardware startups must initiate a 'departing-employee code provenance audit' within 48 hours, using `git log --author="name" --oneline` to scan key commits over the past six months for Apple patent-related keywords (e.g., "Apple Neural Engine", "Siri On-device"); legal teams must urgently revise NDAs to include a mandatory 'prior-employer technology boundary confirmation clause'. 4. StepN Unveils Step AOS—the World's First Agent-Native Operating System https://www.bestblogs.dev/article/4f359a9ce1?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Core Insight: Step AOS is not a Linux derivative but a runtime environment purpose-built for agents—featuring three built-in security layers: Memory Wall (prevents memory contamination), Tool Wall (sandboxed tool invocation), and Identity Wall (multi-agent privilege isolation). It supports command-line lifecycle management: `agent install`, `agent update`, and `agent rollback`. — Actionable Next Steps: Developers should clone the Step AOS GitHub repository (https://github.com/stepai/step-aos) and run `step-cli init --template=web-scraper` to rapidly generate a web-scraping agent with auto-retry and credential rotation; enterprise users must validate migration of existing LangChain/LLamaIndex workloads to the `step-agent run` CLI within one week—especially testing how the `--memory-limit=2MB` parameter affects long-session stability. 5. Baidu Upgrades Its 'Dazi' Agent System with 'Delegation Radius'—Personal, Enterprise & Ecosystem Alliance Tracks https://www.bestblogs.dev/article/269cb164?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Core Insight: Baidu quantifies agent trustworthiness as 'Delegation Radius': the Personal Edition enables semi-managed tasks like WeChat message archiving and meeting note generation; the Enterprise Edition delivers finance-grade trust via private deployment, SM4 encryption, and auditable logs; the Ecosystem Alliance opens a certified 'Dazi Skill Marketplace' for third-party developers. — Actionable Next Steps: SaaS product managers should log into Baidu Intelligent Cloud's Dazi Open Platform (https://cloud.baidu.com/product/dazi), create their first enterprise-customized agent using the 'contract review' skill template, and verify OAuth2.0 integration with DingTalk approval flows; developers must submit skills to the marketplace for review within three days—ensuring inclusion of a `capability_manifest.json` descriptor and minimal-permission declarations. 6. SK Hynix's $26.5B U.S. IPO Surges 12.76% on Day One—Second-Largest IPO in History https://www.bestblogs.dev/article/67749c2dcd?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Core Insight: SK Hynix's Nasdaq listing is more than fundraising—it's a global vote of confidence in AI compute infrastructure. Its HBM3e chip yield has reached 82%, and it has secured an exclusive supply agreement with NVIDIA for 2026–2028—transforming memory from a cost center into a strategic control point. — Actionable Next Steps: Hardware procurement leads should contact SK Hynix China sales (https://www.sk-hynix.com/zh-cn/contact) immediately to request HBM3e samples and the *AI Server Memory Configuration White Paper*; developers should enable `torch.cuda.memory_efficient_attention` in PyTorch 2.4+, benchmarking actual throughput gains for `flash_attn` under HBM bandwidth (target ≥35%). 7. Tencent Launches WorkBuddy: Deep WeChat Ecosystem Integration & Localized Office Automation https://www.bestblogs.dev/article/facbb2909b?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Core Insight: WorkBuddy is not a Copilot clone—it's a native 'Office OS' built atop WeChat's official APIs, enabling direct parsing of chat histories into to-do items, automatic archival of WeChat Official Account articles into Notion, and voice-triggered Excel formula computation—all data remaining strictly within the WeChat ecosystem. — Actionable Next Steps: Enterprise IT departments should deploy the WorkBuddy WeChat Mini-Program admin console (https://workbuddy.qq.com/admin) immediately, activate 'meeting minutes auto-sync', and validate token refresh logic with Tencent Meeting APIs; individual users can test semantic accuracy with `@WorkBuddy organize all chats containing 'reimbursement' from last week`, recording misclassification rates and submitting feedback to `feedback@workbuddy.qq.com`. 8. GPT-5.6-Sol Reported to Delete Entire Mac Filesystem in Ultra Mode https://www.bestblogs.dev/article/f22c785371?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Core Insight: This bug stemmed from Sol incorrectly resolving the `{home_path}` variable as `/` in Ultra mode—causing `rm -rf {home_path}/Downloads` to expand to `rm -rf /Downloads`, exposing systemic risk: agent autonomy lacks built-in 'dangerous-command interception hooks'. OpenAI has issued an emergency `--safe-mode` launch flag. — Actionable Next Steps: All teams using GPT-5.6 must upgrade to v5.6.1 within 24 hours and enforce `--safe-mode --hook-config=./hooks.yaml` in startup scripts; developers must author `hooks.yaml` per official guidance (https://platform.openai.com/docs/guides/agent-security/hooks), implementing regex-based interception rules for at least `rm`, `curl`, and `ssh`. 9. Zhipu AI Launches 'Reach-High Initiative': Focused on Long-Horizon Tasks, Autonomous Agents, Self-Evolution & Safety Governance https://www.bestblogs.dev/article/696de3a103?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Core Insight: Zhipu has deprioritized short-term monetization, redirecting all R&D resources over the next two years toward four AGI frontiers: 'Long-Horizon Tasks' require sustained contextual coherence beyond 72 hours; 'Autonomous Agents' must pass >1,000 scenario stress tests (e.g., resuming uploads autonomously after network interruption); 'Self-Evolution' relies on a three-layer architecture driven by `AGENTS.md`, `MEMORY.md`, and `USER.md`. — Actionable Next Steps: Researchers should download Zhipu's open-source