Keywords：GPT-5.6, Claude Code security backdoor, AI inference chip self-developed, GPT-Live full-duplex voice, Muse Image intelligent agent architecture, Fable 5 cost-saving workflow
🔥 Focus
OpenAI GPT-5.6 Fully Unlocked and Publicly Released : After supplementary testing by the U.S. Department of Commerce’s AI Safety and Innovation Center, OpenAI’s GPT-5.6 series models (including the flagship Sol, mid-range Terra, and lightweight Luna) have received government approval for official public release. Previously, the series was restricted to select enterprise previews due to safety reviews. Tests show that Sol achieved a score of 88.8% on the TerminalBench 2.1 programming benchmark, while Sol Ultra reached 91.9%, completely surpassing Anthropic’s Fable 5, and consuming only one-third of the tokens in cybersecurity tasks, marking a major breakthrough in reasoning and efficiency for frontier models. (Source: THE DECODER, Hacker News)
Meta Releases Agentic Image Generation Model Muse Image : Meta’s Superintelligence Lab released its first image generation model, Muse Image, and a video preview version, Muse Video. The model adopts an agentic architecture, planning, searching the web, and running code before generation to ensure the accuracy of charts and QR codes. In the Image Arena, its image generation and editing capabilities ranked second, only behind GPT Image 2. However, its default privacy setting, which allows users to generate images using others’ public Instagram photos via @username, has sparked controversy over European GDPR compliance and portrait rights. (Source: Meta AI Blog, THE DECODER)
MIIT Warns of Security Backdoor Risks in Claude Code : The Cybersecurity Threat and Vulnerability Information Sharing Platform of the Ministry of Industry and Information Technology (NVDB) released an announcement stating that monitoring has detected security backdoor risks in Anthropic’s AI programming tool Claude Code (versions 2.1.91 to 2.1.196). It unauthorizedly transmits sensitive user information such as time zones and identities, and monitors and rewrites prompts containing keywords of Chinese cloud providers and API proxies. Currently, Alibaba has completely banned the tool internally, and the MIIT recommends that relevant developers uninstall it immediately or upgrade to the latest version. (Source: QbitAI, Heart of the Machine)
DeepSeek and Zhipu AI Secretly Launch Self-Developed Inference Chip Projects : According to Reuters and The Information, domestic LLM giants DeepSeek and Zhipu AI have both secretly launched self-developed AI inference chip (ASIC) projects. DeepSeek’s project was initiated a year ago and is currently low-key recruiting chip design engineers and contacting foundries like TSMC and Samsung, as well as memory manufacturers. This move aims to reduce reliance on NVIDIA GPUs and optimize the inference energy efficiency of its MoE architecture using customized hardware, in response to the high computing power costs brought by the surge in API calls. (Source: QbitAI, Heart of the Machine)
OpenAI Releases Full-Duplex Voice Model GPT-Live : OpenAI officially launched GPT-Live and GPT-Live-1 mini, directly replacing the old Advanced Voice Mode of ChatGPT. The model adopts a full-duplex architecture, supporting simultaneous listening and speaking, allowing users to interrupt naturally at any time. When handling complex reasoning and search tasks, the model seamlessly delegates them to GPT-5.5 in the background while maintaining a coherent and natural conversation flow on the frontend. The API version will also be opened to developers in the near future, promoting voice as the primary agentic interface for human-computer interaction. (Source: OpenAI News, TechCrunch)
🎯 Dynamics
Microsoft Copilot Gradually Replaces OpenAI and Anthropic with Self-Developed Models : Bloomberg reports that Microsoft has begun replacing OpenAI and Anthropic models with its self-developed MAI model in core office applications like Excel and Outlook, processing tens of thousands of prompts weekly. Although MAI currently accounts for only a small fraction of total usage, this move indicates that Microsoft is reducing expensive third-party model licensing costs through “self-reliance.” In the future, Microsoft plans to make the MAI model the default free configuration, while treating third-party models as paid premium add-ons. (Source: Bloomberg, THE DECODER)
BAAI Releases World Model Orca and Wins UN Award : The Beijing Academy of Artificial Intelligence (BAAI) released Orca, a billion-parameter multimodal world model trained on 125,000 hours of video and 160 million event data points. It focuses on “next-state prediction” in a unified world latent space, demonstrating excellent understanding, prediction, and action control capabilities with a frozen backbone. Meanwhile, the open-source unified system software stack “FlagOS” led by BAAI, which adapts to 32 types of AI chips and promotes global AI computing power inclusion, was awarded the UN’s AI for Good Global Excellent Case Award. (Source: Heart of the Machine, QbitAI)
Tencent Hunyuan Hy3 Official Version Open-Sourced : Tencent officially released and open-sourced the Hunyuan Hy3 (295B MoE, 21B active) large model under the Apache 2.0 license. This version is deeply optimized for coding, long context, and hallucination resistance. In real-world testing within Tencent’s internal WorkBuddy and other office scenarios, the task success rate increased from 72% to 90%, and the average time consumed was reduced by 34%. While maintaining reasoning cost-effectiveness, the model focuses on deeply integrating AI capabilities into Tencent’s core ecosystems such as WeChat, games, and cloud. (Source: 36kr, QbitAI)
MiniMax Plans to Open-Source 2.7 Trillion Parameter M3 Pro Model : According to The Information, domestic AI startup MiniMax plans to open-source its next-generation flagship large language model, internally codenamed “M3 Pro”, in the third quarter of 2026. The model features 2.7 trillion parameters, far exceeding the scale of its existing 428 billion parameter M3 model. This move aims to provide stronger complex reasoning and multi-step task processing capabilities through an ultra-large-scale open-source model, competing with closed-source commercial models in the open-source ecosystem. (Source: THE DECODER)
Mistral AI Launches First Robot Navigation Model Robostral Navigate : European AI unicorn Mistral AI released its first robot navigation model, Robostral Navigate. With 8B parameters, the model requires only a single RGB camera to guide wheeled, legged, or flying robots for autonomous navigation. The model achieved a 79.4% success rate on the R2R-CE benchmark, surpassing systems using depth sensors or multiple cameras, demonstrating Mistral’s ambition to extend AI capabilities to the physical world and embodied intelligence. (Source: Mistral AI, THE DECODER)
Meta Tests Smart Glasses with “Super Sensing” Mode : The Financial Times reports that Meta is testing a “Super Sensing” mode for its next-generation Ray-Ban smart glasses. This mode allows the camera and microphone to run continuously in the background for hours, snapping photos and recording audio, enabling AI to help users recall what they saw and heard at any time. However, Mark Zuckerberg reportedly asked whether the recording indicator LED could be turned off in this mode, raising strong concerns internally and publicly about privacy leaks and unauthorized surveillance. (Source: ZDNet, THE DECODER)
OpenAI Chief Futurist Joshua Achiam Announces Departure : Joshua Achiam, who served at OpenAI for nearly 9 years as Chief Futurist, announced he will leave the company on July 24. Achiam previously led the mission alignment team and testified in Elon Musk’s lawsuit against Sam Altman. He was once scolded by Musk in 2018 for questioning him face-to-face about “rushing AGI at the expense of safety.” His departure is the latest in a series of losses of core safety and policy members at OpenAI, sparking external discussions about OpenAI’s safety commitment during its commercialization process. (Source: The Verge, 36kr)
Google Updates Gemini API Managed Agents Features : Google AI Studio announced four major updates for Managed Agents in the Gemini API: support for background execution for connectionless asynchronous operation, support for remote MCP services directly connecting to databases or APIs, allowing the use of custom functions, and automatic credential refreshing between interactions to maintain sandbox states. These updates aim to help developers more easily build and deploy complex, long-cycle automated agent workflows. (Source: Google AI Studio, THE DECODER)
🧰 Tools
Notion Releases iOS App Notion Agents : Productivity tool Notion launched “Notion Agents”, an iOS app designed specifically for agent interaction. Unlike the main Notion app which focuses on note-taking, this app concentrates on allowing users to converse with custom AI agents or connected external large models (such as ChatGPT, Gemini, Claude). It supports quickly assigning tasks via voice, photos, and text, executing them asynchronously in the background, marking mobile devices as gradually becoming the control consoles for agentic work. (Source: Notion, The Verge)
Former GitHub CEO Launches Decentralized Git Network Entire : Former GitHub CEO Thomas Dohmke launched the decentralized Git network “Entire” and opened its preview. The platform aims to solve the problem of frequent GitHub downtime caused by high-frequency “vibes coding” by AI agents. Entire supports one-click mirroring of GitHub repositories, with tests showing it can support 570,000 clones per hour and 586 pushes per second. It also provides a semantic memory layer designed for agents, making it easy for human developers to review AI-generated code and modification intents. (Source: ZDNet)
Google AI Studio Launches One-Click Import from GitHub : Google AI Studio launched the “Import from GitHub” feature in Build mode. Developers no longer need to write prompts from scratch; they only need to provide a GitHub repository link, and the system will automatically convert the codebase into a format compatible with the AI Studio runtime, opening it in the Build environment for preview, interactive modification, and one-click deployment to Cloud Run. This further closes the loop between AI-generated code and traditional development workflows. (Source: Google AI Studio, MarkTechPost)
LangChain and NVIDIA Release NemoClaw Deep Agents Blueprint : LangChain and NVIDIA jointly released the “NemoClaw Deep Agents Blueprint” open-source reference architecture. This architecture integrates the NVIDIA Nemotron 3 Ultra open-source model, LangChain Deep Agents harness, and the OpenShell runtime environment. Tests show that while ensuring agent planning and tool invocation performance, this architecture can reduce enterprise-level agent inference costs by more than 10 times, providing a standardized solution for private deployment of agents in enterprises. (Source: LangChain)
Sqlite-utils 4.0 Released with Database Migration Features : Developer Simon Willison released the major update sqlite-utils 4.0. This version introduces structured database schema migrations for the first time, utilizing Python libraries and the transform() method to achieve safe database schema evolution and version tracking. In addition, the new version adds nested transaction support based on db.atomic() and composite foreign key support. Claude Fable 5 was heavily used for API design and code refactoring during the development process. (Source: Simon Willison)
📚 Learning
ByteDance Releases EdgeBench Revealing Agent Scaling Law : ByteDance’s Seed team released EdgeBench, a long-horizon agent evaluation project. Through a total of 38,000 hours of running tests on 5 frontier models across 134 long-horizon tasks, the study found that the learning and progress curves of agents in unfamiliar environments fit a log-sigmoid function with an accuracy of R²=0.998. The experiments also confirmed that accumulating experience through continuous running has a clear advantage over multiple independent restarts, proving the value of long-term planning and experience internalization. (Source: 36kr, ZhihuFrontier)
Study Confirms Training a Single Transformer Layer Can Match Full-Parameter RL : Teams from the University of Minnesota, Peking University, and Amazon published a paper challenging the traditional assumption that RL post-training requires updating all layers. Through systematic testing on 7 models, 3 algorithms, and 3 domains, they found that RL gains are highly concentrated in the middle layers (40-60% depth) of the network. Experiments show that training only a single critical Transformer layer can match or even surpass full-parameter RL training. The proposed middle-layer heuristic strategy based on this achieved better performance across models of various scales. (Source: Heart of the Machine)
.jpg)
HKUST Proposes Black-Box Attack Framework COMA Against Prompt Compression : A research team from the Hong Kong University of Science and Technology published a paper at the top software engineering conference ASE 2026, pointing out that “prompt compression” modules introduced to save tokens could become a new vulnerability for AI agents. The team proposed the black-box attack framework COMA, which adds minor perturbations to user inputs to amplify information loss during compression, intentionally inducing the compressor to remove safety constraints or critical evidence from system prompts. Experiments show that isolated compression is an effective defense against such attacks. (Source: Heart of the Machine)
.jpg)
Nanjing University and Tsinghua University Propose Algorithm Competition Agent Solvita : Researchers from Nanjing University, Tsinghua University, and other institutions jointly proposed the Solvita framework. Without fine-tuning large models, it enables agents to accumulate algorithmic skills from problem-solving and adversarial testing through a closed loop of cross-checking and local repair among four roles: Planner, Solver, Oracle, and Hacker, as well as a trainable graph-structured knowledge network. In real Codeforces competition evaluations, the framework successfully helped multiple open-source models reach the Legendary Grandmaster tier. (Source: 36kr)
Goodfire Releases BSFs for Multidimensional Concept Extraction in Activation Space : Interpretability AI startup Goodfire released research on “Block-Sparse Featurizers (BSFs)”. Traditional Sparse Autoencoders (SAEs) interpret model concepts as one-dimensional directions, whereas BSFs map concepts into multidimensional “blocks” and geometric regions. This method has been validated on SDXL and biological models, and for the first time, topological geometric representations consistent with the physical position of robotic arms were discovered in robotics models, providing a tool for precise model debugging. (Source: GoodfireAI)
ZML/LLMD Inference Server Supports Cross-Chip Optimization : French AI startup ZML released the free ZML/LLMD inference server, aiming to break hardware lock-in. The system supports running open-source large models on various chips, including NVIDIA, AMD, Intel Arc, Google TPU, and Apple Silicon. It provides features such as dynamic mixed-precision quantization, continuous batching, and prefix caching, helping enterprises deploy AI tasks on local heterogeneous computing power with extremely low cost and high energy efficiency. (Source: TechCrunch)
💼 Business
Prime Intellect Completes $130 Million Series A Funding : Prime Intellect, an AI infrastructure platform designed to help enterprises train their own agents, announced the completion of a $130 million Series A funding round, reaching a valuation of $1 billion. This round was led by Radical Ventures, with participation from industry capital including NVIDIA, Intel, and Dell. Its full-stack platform provides distributed computing power, reinforcement learning frameworks, and evaluation tools, helping enterprises use their own data for model fine-tuning and reverse distillation to break free from reliance on closed-source large model APIs. Its ARR has reached $100 million. (Source: TechCrunch)
Agentic Legal AI Startup Norm Ai Raises $120 Million : Norm Ai, a startup focusing on compliance and legal AI, completed a $120 million Series C funding round, reaching a valuation of $1.2 billion. This round was led by Khosla Ventures, with Blackstone and others participating. Its core product, “Norm Law”, is an AI-native virtual law firm that utilizes teams of AI agents to provide external legal compliance consulting for enterprises, breaking the traditional hourly billing model to adopt a pay-for-deliverables business model. (Source: AI Business)
SambaNova Systems Completes First Close of $1 Billion Series F : AI chip and integrated hardware-software service provider SambaNova Systems completed the first close of its Series F funding round with $1 billion, reaching a valuation of $11 billion, led by General Atlantic with participation from Intel and others. The funds will be used to secure supply chain capacity for its next-generation SN50 chip. Meanwhile, JPMorgan Chase announced it has selected SambaNova as its private inference infrastructure partner to run its trillion-parameter financial large models securely on-premises. (Source: TechCrunch)
🌟 Community
Fable 5 Extension Sparks Heated Discussion on “Advisor and Orchestrator” Cost-Saving Workflows : Anthropic announced the extension of the free trial period for Claude Fable 5 to July 12, sparking heated discussions in the community on how to “squeeze the most out of Fable 5.” The official team also released two cost-saving modes, “Advisor” and “Orchestrator”: letting the cheaper Sonnet 5 act as the main executor and only calling Fable 5 for key decisions or planning. Tests show this can achieve 92% of the performance at 63% of the cost on SWE-bench Pro, providing a practical architectural solution for high API costs. (Source: THE DECODER, 36kr)
The “52% Utilization Lifeline” of AI Inference Costs and Cognitive Arbitrage : Industry analysis points out that whether it is more cost-effective for enterprises to build their own GPU inference servers rather than calling APIs depends on a neglected physical metric—GPU utilization. Derivations show that only when the actual utilization rate of self-built GPUs exceeds 52% can the amortized hardware depreciation and operation and maintenance costs beat API prices. API vendors’ pricing strategies essentially exploit the cognitive blind spot where customers cannot accurately calculate their own utilization rates to perform “structural arbitrage” on scheduling efficiency. (Source: 36kr)
Divergence of B2B and B2C Routes for Large Models Widens Valuation Gap : Zhipu and MiniMax, which went public almost at the same time, saw their market value gap widen to HKD 600 billion within half a year. Analysis points out that the capital market is reconstructing the pricing logic of large model companies, shifting from pure “user volume” to “defensive workflows and pricing power.” Zhipu bet on B2B coding and enterprise APIs, showing stronger customer stickiness and price-raising capabilities; while MiniMax focused on B2C virtual companionship and entertainment, facing high user acquisition costs and a price floor brought by DeepSeek, raising questions about its commercial sustainability. (Source: 36kr)
New Breakthroughs in Embodied Physical AI Closed-Loop and Motion Transfer : ZGC EmbodyAI published the “Human-as-Humanoid” paper, achieving zero-shot transfer of human video actions directly into robot control trajectories on the 60-DoF humanoid robot PrimeU without real-robot demonstration data. Meanwhile, the Chinese Academy of Sciences and other institutions released SurgMotion, the first billion-parameter surgical video foundation model. Based on 15 million frames of data, it utilizes V-JEPA for latent space motion prediction, demonstrating the latest progress of AI in establishing 3D spatial and dynamic perception in physical and medical scenarios. (Source: Heart of the Machine, Heart of the Machine)
.jpg)
Discord AI Moderation Bug Leads to Wrongful Bans of Many Innocent Users : Social platform Discord admitted that its AI moderation system mistakenly banned over 8,000 users over the past two months due to a software bug. The system’s similarity matching mechanism, while filtering harmful content, flagged harmless grid patterns uploaded by users (such as chessboards, game texture maps, spreadsheets, and gray transparent background images) as illegal harmful content and directly executed permanent bans. Discord stated that it has fixed the bug and is gradually restoring the affected accounts. (Source: TechCrunch)
💡 Others
Surging Data Center Energy Demands Threaten U.S. Manufacturing and Grid Security : A Reuters analysis points out that the explosive electricity demand of U.S. AI data centers is driving up grid loads, causing electricity bills for manufacturing companies in some industrial hubs to soar. For example, a brick factory in Ohio saw its monthly electricity bill skyrocket from $1,600 to $12,000 due to capacity fee adjustments, severely squeezing the profit margins of steel mills and traditional factories. This could undermine government efforts to revive U.S. manufacturing through tariff protection and reshoring, and intensifies the conflict over electricity usage between the tech industry and traditional industries. (Source: Ars Technica)
Brain-Computer Interface Device Starts Clinical Trials for ALS Patients : A new development in the medical technology field: a brainwave decoding device designed specifically for amyotrophic lateral sclerosis (ALS) patients announced it will officially enter clinical trials in 2026. The device collects patients’ EEG signals through non-invasive or minimally invasive sensors and uses AI algorithms for real-time semantic decoding, helping patients who have lost the ability to move and speak to communicate accurately with the outside world again, demonstrating the social value of AI in neurorehabilitation and healthcare. (Source: Ronald_vanLoon)