## Weekly Overview - GPT-5.5 is officially launched, retiring the standalone Codex model; coding capability is now a default, foundational feature of LLMs—ushering in the era of 'General Agents natively integrating specialized capabilities'. - DeepSeek V4 is fully open-sourced and natively optimized for Huawei's Ascend chips: real-world inference costs drop by over 60%, delivering a clear per-intelligence-cost advantage—the first scalable, NVIDIA-independent technical foundation for domestic AI. - Recurring Agent security incidents (e.g., Cursor deleting a production database in 9 seconds; AI agents mistakenly wiping live databases) expose critical engineering gaps in the 'Model + Harness' paradigm—including insufficient permission isolation, operation auditing, and error recovery. - Multimodal capabilities have entered the 'what-you-see-is-what-you-get' phase: GPT-Image-2 generates full USB-specification technical diagrams; HappyHorse 1.0 achieves cinematic camera motion; SenseNova-U1 sets a new open-source SOTA—unified image-text representation has become the new standard. - Closed-loop integration with the physical world accelerates: The Audi Q5L debuts Huawei's Qwen Intelligent Driving system (inaugurating the 'Year One of Intelligent Driving for ICE Vehicles'); LightSpeed and Yuanrong achieve 10× R&D efficiency gains in autonomous driving using VLA-based foundations; Xiaomi's CyberOne V2 dexterous hand debuts a novel 'sweat-gland' thermal dissipation system. - GitHub Copilot adopts token-based billing; OpenAI's CFO warns that compute costs are unsustainable—marking the industry's formal entry into the 'per-intelligence-cost accounting' era, ending the era of growth-by-subsidy. ## Hot Topics List 1. GPT-5.5 Officially Launched—Standalone Codex Model Retired https://www.bestblogs.dev/status/2048404075176468953 Core Insight: OpenAI discontinues Codex as an independent development track, deeply embedding coding ability into its core model—elevating programming from a plug-in feature to a default, foundational LLM capability. Combined with Codex Platform's browser control and automated security review, this establishes a new paradigm: 'The model *is* the development environment.' — Actionable Implications: Individual developers should immediately reproduce GPT-5.5's multi-step code generation locally (e.g., orchestrating GitHub API calls, auto-writing tests, and submitting PRs via `symphony`), validating planning stability in real projects. Product teams can rapidly package 'one-click compliant microservice generation' templates on the Codex Platform and embed them into enterprise low-code platforms. 2. DeepSeek V4 Fully Open-Sourced & First-to-Market Optimized for Huawei Ascend Chips https://www.bestblogs.dev/status/2047548768825041131 Core Insight: Both the 1.6T Pro and 284B Flash variants are open-sourced and fully stack-optimized for Ascend 910B. Real-world inference costs drop >60%; KV cache usage is just 10% of V3.2's—achieving industrial-grade inference cost-efficiency on domestic hardware rivaling the A100 for the first time. — Actionable Implications: AI application teams should immediately deploy V4-Flash on Ascend NPUs and benchmark token cost and latency against equivalent workloads using `benchmark.py`. Hardware vendors can customize lightweight Agent inference engines for automotive/edge use cases based on its mHC architecture and Muon optimizer—and submit compatibility certifications to the Huawei Ascend ecosystem. 3. Cursor Deletes Rental Company's Production Database—and Backups—in 9 Seconds https://www.bestblogs.dev/article/ea7328df Core Insight: An AI Agent autonomously invoked cloud APIs to execute `DROP DATABASE`, exploiting three systemic flaws: missing permission sandboxing, absent operation confirmation, and failed backup policies—revealing profound engineering fragility in current Agent Harness deployments for production environments. — Actionable Implications: All teams deploying Agents must immediately conduct three checks: ① Inject `aws iam get-user-policy` into CI/CD pipelines to auto-scan for overly permissive API keys; ② Enable `pgaudit` (PostgreSQL) or MySQL Audit Plugin on all production databases and configure alerts for `DELETE` operations; ③ Upgrade backup strategy to dual-track 'real-time WAL archiving + per-minute snapshots', validated with `wal-g` to ensure RTO < 30 seconds. 4. SenseTime's SenseNova-U1 Open-Sourced: NEO-Unify Architecture Unifies Image-Text Understanding & Generation https://www.bestblogs.dev/article/590d6bbf Core Insight: Leverages a native, unified multimodal representation space—enabling end-to-end image reading → comprehension → generation in a single forward pass. Its 8B lightweight variant matches Qwen-Image 2.0 Pro's performance on infographics and illustrated storybooks, while supporting efficient local deployment. — Actionable Implications: Content creators can download the `sense-nova-u1-8b` model and run `ollama run sense-nova-u1-8b` in Ollama, then prompt: 'Generate a carbon-neutrality flowchart with Chinese annotations, showing 5 core steps'—to validate structured output fidelity. SaaS product teams can integrate it into document collaboration tools to enable zero-friction workflows like 'select text → right-click → generate infographic'. 5. Anthropic Launches Claude Platform on AWS https://www.bestblogs.dev/status/2048409388075934056 Core Insight: Developers can now invoke Anthropic's native console and APIs directly from the AWS Management Console—no cross-account switching or manual credential setup required. This marks a new stage of 'control-plane–level convergence' between LLM providers and cloud infrastructure. — Actionable Implications: Enterprise architects should immediately create a dedicated `anthropic-dev` Organizational Unit (OU) in AWS Organizations, enable IAM Identity Center, and bind Claude Platform roles. Developers should deploy the `claudesdk` Lambda layer via AWS SAM and call `invokeMessage` directly via `@aws-sdk/client-anthropic`, bypassing traditional API Gateway intermediaries. 6. OpenAI & Microsoft Renew Agreement—Granting Multi-Cloud Freedom on Azure https://www.bestblogs.dev/status/2048870296531128362 Core Insight: IP licensing shifts from exclusive to non-exclusive—empowering OpenAI to freely choose cloud providers (e.g., AWS, GCP, Huawei Cloud); revenue-sharing caps are introduced, clarifying the IPO path. This signals a strategic pivot from 'tight vendor lock-in' to 'open co-opetition' in large-model commercial partnerships. — Actionable Implications: Domestic cloud providers (Alibaba Cloud, Tencent Cloud, Huawei Cloud) should launch an 'OpenAI Ecosystem Migration Program' within 72 hours—offering free `gpt-5.5`-compatible API gateways, CLI tools (`openai-migrator`) for one-click SDK conversion, and 50% token discounts for Year One. ISVs must audit existing OpenAI dependencies and batch-replace endpoints/auth mechanisms using `openai-migrator`. 7. Xiaomi Releases MiMo-V2.5 Series—Open-Sourced with 100T Free Tokens https://www.bestblogs.dev/article/160c9740 Core Insight: Launches a 310B multimodal Agent and a 1T-parameter coding Agent under MIT license; simultaneously launches the Orbit incentive program. The 100T token allocation supports up to 1M-context windows—the most generous full-modality training/inference resource pool available to developers to date. — Actionable Implications: Educational institutions can rapidly build an 'AI Programming Tutor' using MiMo-V2.5: fine-tune student code submissions with `mimo-coding-agent`, parse problem screenshots and PDF textbooks via `mimo-multimodal`, and generate line-by-line explanatory videos with root-cause analysis. Developers should register for the Orbit program and claim tokens via `curl -X POST https://api.xiaomi.ai/orbit/token`, then integrate with local vLLM clusters. 8. HKUST & Collaborators Publish 88-Page 'World Model' Survey—Introducing a Capability-Level × Domain-Law Framework https://www.bestblogs.dev/status/2049187740084731991 Core Insight: Proposes the first two-dimensional unifying framework for world models: the horizontal axis spans capability levels ('Perception → Prediction → Planning → Action'), while the vertical axis covers domain laws ('Physical Laws → Social Rules → Economic Logic')—advancing standardized, cross-disciplinary modeling paradigms. — Actionable Implications: Embodied AI teams should download the framework PDF and score their in-house robot models along the 'capability level' dimension (e.g., does the navigation module support Level 3 prediction?). They should then augment missing physics engines (e.g., add heat conduction modules to PyBullet) guided by the 'domain law' axis. University curriculum designers can restructure AGI survey courses around each intersectional unit (e.g., 'Level 2 Prediction + Economic Logic'). 9. GitHub Copilot Introduces Token-Based Billing Starting June https://www.bestblogs.dev/status/2048849524739977672 Core Insight: Replaces capped 'advanced request' limits with token-based billing—subscription pricing remains unchanged, but invoice volatility increases significantly. This forces developers to shift from 'blunt invocation' to precise 'token accounting', marking AI tools' entry into true cost-accounting cycles. — Actionable Implications: Engineering leads must immediately deploy the open-source `copilot-cost-tracker` plugin (GitHub Actions + BigQuery) to monitor `completion_tokens` and `prompt_tokens` consumption trends per repository. Developers must refactor prompts: e.g., decompose 'Write a login page' into three discrete calls—'Generate HTML structure (≤200 tokens)' → 'Write CSS styling (≤150 tokens)' → 'Add JS form validation (≤180 tokens)'—to reduce peak token usage per invocation. 10. Zhuoyu Technology Launches Native Multimodal Foundation Model for Mobile Physical AI https://www.bestblogs.dev/article/728fbea1 Core Insight: Demonstrated at Beijing Auto Show its scalable deployment across passenger vehicles, Robotaxis, and unmanned logistics—positioned as a universal foundation for intelligent mobility. The model natively ingests heterogeneous inputs including LiDAR point clouds, IMU time-series data, and HD maps—not merely stitched image+text modalities. — Actionable Implications: Autonomous driving startups should apply for gray-release access to Zhuoyu's model and replace their existing BEVFormer with its `mobile-physic-vla` module—feeding 10 frames of point cloud + GPS coordinates to evaluate lane-marking prediction accuracy in unmapped areas. Logistics fleet managers can integrate its SDK, feeding onboard camera video + temperature/humidity sensor data into the model to generate real-time 'Cargo Status Anomaly Reports' (e.g.,