AI Daily – 2026-07-13

Keywords：AI Industry Trends, Large Model Competition, Artificial Intelligence Technology, GPT-5.6 Sol, Self-developed Inference Chip, DeepSeek-V4

🔥 Spotlight

OpenAI and Anthropic Spark a Limit and Pricing War : Facing the highly cost-effective GPT-5.6 Sol launched by OpenAI, Anthropic was forced to extend free access to Claude Fable 5 for subscribers until July 19. In response, OpenAI announced the temporary removal of Codex’s 5-hour usage limit and a reset of quotas. This “limit war,” triggered by the imbalance between computing power supply and demand, is essentially a defensive concession by the two leading frontier labs amid anxiety over user churn. It also signals that LLM pricing is evolving from a single Token-based metric to competition over total task costs and user stickiness (Source: QbitAI, THE DECODER)

Apple Sues OpenAI for Systematically Stealing Trade Secrets : Apple has officially sued OpenAI, former executive Tang Tan, and employee Chang Liu, accusing them of systematically stealing hardware designs and supply chain secrets of unreleased Apple products through recruitment interviews and inducing current employees to take CAD drawings and prototypes. Currently, over 400 former Apple employees have moved to OpenAI. This move not only marks a shift from cooperation to open confrontation between the two companies but could also deal a heavy blow to OpenAI’s consumer AI hardware R&D and its upcoming IPO process (Source: 36Kr, CSDN)

Zhipu and MiniMax Face Concentrated Lock-up Expirations and Launch Massive Refinancing : As Zhipu and MiniMax welcome their first large-scale lock-up expiration post-listing, they have respectively announced plans to raise approximately HKD 31.4 billion and HKD 16 billion through share placements in the Hong Kong stock market. Facing valuation reshaping brought by the lock-up expiration and high computing power consumption, Zhipu founder Tang Jie released an internal letter launching the “Reach High Project” (摸高计划), focusing on technical breakthroughs such as long-horizon tasks; MiniMax founder Yan Junjie announced a salary suspension and the donation of his shares. This indicates that China’s two leading LLM giants are hedging against the industry bottleneck of “Token uneconomics” through massive financing and structural adjustments (Source: 36Kr, QbitAI)

Frontier AI Companies Race to Explore Self-Developed Inference Chips : According to a report by Synced, following the joint release of the first ASIC inference chip Jalapeño by OpenAI and Broadcom, DeepSeek and Zhipu have also been reported to be evaluating self-developed custom AI inference chips. This trend indicates that as AI applications evolve toward Agents, Token consumption during the inference phase is growing exponentially. Self-developed ASIC chips not only help frontier labs significantly reduce long-term operating costs and improve energy efficiency ratios, but also serve as a crucial strategic chip to break free from Nvidia’s GPU monopoly and cope with geopolitical export controls (Source: Synced)

🎯 Dynamics

DeepSeek-V4 System Design and Optimization Details Exposed : Tsinghua PACMAN Lab shared a deep technical analysis of DeepSeek-V4. The V4 model (Pro version 1.6T, Flash version 284B) treats the 1M long context as a system cost issue. By introducing a hybrid CSA/HCA attention mechanism, a manifold-constrained hyper-connection (mHC) multi-stream residual structure, and the Muon optimizer, it significantly reduces inference FLOPs per Token and KV cache footprint. This demonstrates that long-context intelligence is evolving from a benchmark feature to full-stack engineering optimization (Source: ZhihuFrontier)

OpenAI Safety Head Heidecke Departs Amid Team Restructuring : Johannes Heidecke, OpenAI’s head of safety systems, announced his departure, becoming the sixth safety leader to leave within two years. Meanwhile, OpenAI restructured its safety team, integrating it into the research division, with Mia Glaese appointed as VP of Research and Safety. Against the backdrop of models like GPT-5.6 Sol possessing stronger Agent execution capabilities and being more prone to “exceeding user intent,” the weakening of the safety team’s independence has once again raised external concerns about AI safety governance mechanisms (Source: QbitAI)

Google Releases First Multimodal Health Foundation Model SensorFM : Google’s research team has introduced SensorFM, a foundation model for the health domain. Based on self-supervised pre-training on over 1 trillion minutes of unlabeled sensor data collected from Fitbit and Pixel Watch, the model can process 34 physiological and behavioral features, including heart rate and skin temperature. Experiments show that SensorFM performs exceptionally well in 35 health prediction tasks and significantly improves the professionalism and safety of health advice when integrated into the Gemini health assistant (Source: THE DECODER)

German Research Consortium Releases Hybrid Architecture Open-Source Model Soofi S : The German AI Association and several research institutions have jointly released the open-source hybrid model Soofi S 30B-A3B. The model adopts a Mamba-Transformer hybrid architecture, significantly increasing the weight of German corpora during training, and has outperformed similar open-source models like OLMo 3 in English-German bilingual and programming benchmarks. Its MoE architecture activates only 3.2B parameters, demonstrating extremely high throughput and inference cost advantages in long-context and high-concurrency deployments (Source: THE DECODER)

Claude Code Adds Built-in Browser Feature : Anthropic has integrated a built-in browser window into Claude Code, allowing the AI to directly open, read, click, and type on external web pages within the terminal to retrieve technical documentation and handle external tickets more efficiently. To prevent security risks, the browser runs in a clean configuration without login information and is equipped with classifier reviews for write operations, prohibiting the AI from making purchases or bypassing CAPTCHAs without user authorization (Source: THE DECODER)

ByteDance’s Doubao LLM Launches Comprehensive Commercial Subscription : ByteDance’s Doubao App has officially launched its Pro subscription service, offering three tiers of paid plans. Despite its daily active users (DAU) exceeding 200 million, Doubao faces the “sweet trouble” of massive daily computing power consumption coupled with weak consumer-end monetization capabilities. This fee introduction limits the previously unlimited cloud drive capacity, marking ByteDance’s exploration of converting consumer traffic into subscription revenue to offset steeply rising AI infrastructure expenses (Source: 36Kr)

🧰 Tools

Agnes-2.5-Flash and Desktop Tool AgnesCode Released : Agnes has launched its next-generation high-performance text model, Agnes-2.5-Flash, continuing its indefinite free strategy in the domestic market. The simultaneously launched AgnesCode desktop AI workbench integrates the model, Agent skills, and local development environments, supporting development tasks like multi-file modification and complex architecture understanding. It provides domestic developers with a low-barrier local AI programming and office collaboration tool without the worry of account bans (Source: Synced)

PrismML Compresses 27B Qwen-3.6 to Run on iPhone : Chip startup PrismML has utilized mathematical compression technology to successfully compress the 27-billion-parameter open-source LLM Qwen-3.6 to under 4GB, enabling it to run locally on the iPhone 17 Pro. The uniqueness of this technology lies in keeping all 27B parameters active, supporting complex local dialogue, reasoning, and code generation, bringing a new breakthrough for the economical and resource-constrained deployment of local AI on mobile devices (Source: Reddit)

Hugging Face Opens ZeroGPU Deployment to All Users : Hugging Face has announced that the creation of ZeroGPU Demos and applications is now open to all users. Users only need to give the AI Agent an instruction like “build an HF ZeroGPU Demo for this model” to automatically generate and deploy the app. This move significantly lowers the barrier for developers to showcase models and build interactive AI applications, further flourishing the ecosystem of the open-source AI community (Source: Hugging Face)

📚 Learning

Prime Intellect Open-Sources Agent Reinforcement Learning Environment Verifiers v1 : Prime Intellect has released Verifiers v1, refactoring the underlying architecture of Agent RL training and evaluation. v1 decouples the environment into three parts: Taskset (tasks), Harness (agent logic), and Runtime (execution sandbox). It also introduces a linearly growing “message graph” tracking mechanism, which, combined with an interception server, effectively solves KV cache bloat and evaluation overhead in long-horizon agent training, supporting mainstream inference backends like vLLM (Source: MarkTechPost)

Stanford Team Proposes TRACE, a Training System Targeting Agent Failures : A Stanford University research team has proposed the TRACE system, which aims to automatically synthesize reinforcement learning training environments for specific capabilities by analyzing repetitive agent failures. The system uses contrastive analysis to identify capability deficiencies, generates unlabeled training tasks, trains LoRA adapters via the GRPO algorithm, and finally performs multi-expert fusion through Token-level routing, significantly improving the success rate of complex agents in long-horizon tasks (Source: MarkTechPost)

Beihang, PKU, and Meituan Jointly Propose Policy Improvement Reinforcement Learning Framework PIPO : A joint research team has proposed the Policy Improvement Reinforcement Learning (PIRL) perspective and the PIPO algorithm. This method focuses on “closed-loop optimization” in LLM RL post-training. Instead of only calculating local learning signals of single-batch trajectories, it uses policy improvement feedback across iterations to dynamically amplify effective update directions and suppress or correct harmful updates, achieving consistent performance improvements in multiple tasks such as mathematical reasoning, coding, and tool calling (Source: Synced)

NUS and TikTok Propose Confidence-Gated Reflection Reward Model CAMEL : A team from the National University of Singapore (NUS) and TikTok has proposed the CAMEL framework, transforming the reward model into a confidence-gated reflection mechanism. The model first gives an initial judgment via a single Token, triggering long reasoning reflection and review only when the confidence (log-prob margin) is low. This method, with 14B parameters, outperformed several 70B models in tests like RewardBench, achieving a better trade-off between accuracy and Token cost (Source: Synced)

Nanjing University Team Publishes Position Paper on World Model Evaluation : A team from the School of Artificial Intelligence at Nanjing University has published a position paper systematically discussing evaluation methods for world models in embodied AI decision-making scenarios. The paper points out that the evaluation focus should shift from superficial visual plausibility (such as video generation quality) to decision utility (such as action controllability, reward fidelity, policy ranking consistency, etc.), and proposes an evaluation ladder from L0 to L7, establishing a clearer evaluation coordinate system for embodied AI world models (Source: Synced)

Poxiao Intelligent Releases Bidirectional Tactile Embodied Foundation Model TouchWorld : Professor Yang Shuo’s team from Harbin Institute of Technology (Shenzhen) and Poxiao Intelligent have released the TouchWorld tactile foundation model. The model divides tactile sensation into dual roles: Predictive and Reactive. It predicts contact states before action and performs millisecond-level action correction through high-frequency feedback after contact. In multiple complex robotic manipulation tasks, TouchWorld demonstrated a superior success rate compared to vision-only models when dealing with external disturbances (Source: Synced)

Scholars from Multiple Universities Propose New Paradigm of User-Centric Personalized Recommendation Systems : A research team from UIUC, UC Berkeley, and other universities has published a position paper proposing that in the Agentic era, personalized recommendation will shift from “platform-centric” to “user-centric.” Experiments show that aggregating user data across platforms like Amazon, Google, Twitter, and offline data through LLM Agents can significantly improve the accuracy of product prediction and interest exploration, breaking down data competition and privacy barriers of single platforms (Source: Synced)

💼 Business

SK Hynix Lists in the US, Setting Fundraising Record for Foreign Companies : South Korean memory giant SK Hynix has successfully listed on Nasdaq, with ADR subscriptions oversubscribed by more than 7 times, raising $26.5 billion, setting a new record for foreign companies listing in the US. With an absolute 58% share of the HBM market, SK Hynix has become one of the most profitable links in the AI hardware industry chain. Positioned as a core supplier to Nvidia, SK Hynix aims to obtain a high valuation premium, strengthen cash reserves to cope with expansion competition, and deeply bind itself to the US AI chip ecosystem through this listing (Source: 36Kr)

Qujing Technology Completes Series A Financing, Raising Over 1 Billion Yuan in Half a Year : High-efficiency AI Token production service provider Qujing Technology has announced the completion of its Series A financing, led by Henan Investment Group Huirong Fund. The company has raised over 1 billion yuan in total within half a year. Originating from the Institute of High Performance Computing at Tsinghua University, Qujing Technology focuses on LLM inference optimization, promoting the large-scale output of high-quality Tokens from domestic computing power through its ATaaS platform. Currently, its single-month revenue has exceeded that of the entire previous year, and some business segments have achieved profitability (Source: QbitAI)

SoftBank Group Invests an Additional $30 Billion in OpenAI : SoftBank Group has announced the completion of an additional $30 billion investment in OpenAI, bringing its cumulative investment to $64.6 billion. To raise this massive sum, SoftBank engaged in “extreme fundraising” by liquidating Nvidia shares, reducing its stake in T-Mobile, pledging subsidiary equity, and aggressively issuing bonds and bridge loans. Masayoshi Son’s move deeply binds SoftBank’s future to OpenAI’s commercialization and IPO prospects (Source: 36Kr)

🌟 Community

“Ghost Font” Goes Viral, Cracked by Riley Goodside with a Single Sentence : Developer Eric Lu launched “Ghost Font” based on dynamic noise videos, a design that utilizes the motion perception characteristics of the human eye to prevent AI from reading frame-by-frame. However, prompt engineering expert Riley Goodside successfully cracked it in two minutes by inputting just a single directional prompt into GPT-5.6 Sol. This shows that AI’s visual perception barriers are more fragile than imagined in the face of prompt engineering (Source: 36Kr)

Proliferation of “Negative Parallelism” in AI Writing Sparks Model Collapse Concerns : The Atlantic reported on the proliferation of “negative parallelism” (Not X, but Y), the most common sentence pattern in AI writing. Studies show that the frequency of this pattern in corporate communications has quadrupled and is a highly frequent feature in AI text detection. As new models continue to be trained on AI-generated data containing this pattern, this “self-reinforcing” writing habit is raising concerns about future model collapse (Source: Reddit)

Autonomous AI Hacking Agent Tool JadePuffer Sparks Security Panic : Security firm Sysdig disclosed an autonomous hacking agent named “JadePuffer.” After exploiting a Langflow vulnerability to breach servers, the tool can autonomously execute credential theft, lateral movement, database encryption, and ransom note delivery. What shocked the security team most was that when encountering format errors, the agent could autonomously rewrite and execute exploit code within 31 seconds, demonstrating the destructive power of malicious Agents (Source: Reddit)

“Caveman Prompt” Token-Saving Effect Deemed “Watered Down” in Agent Scenarios : The Caveman project, developed by a freshman at Leiden University, went viral for making AI speak like a caveman to save 65% of Tokens. However, JetBrains’ actual testing showed that in real programming agent scenarios (such as Claude Code), output Tokens were only saved by 8.5%. This is because the bulk of Agent consumption lies in system prompts, tool definitions, and context caching, rather than the final chat text output (Source: 36Kr)

Unitree Partners with Hunan Iron & Steel to Promote Embodied AI Working in Factories : Unitree has reached a strategic partnership with Hunan Iron & Steel Group, collaborating with Looper Robotics to deploy humanoid and wheeled-legged robots in steel plants. In high-risk areas such as belt corridors with high temperatures, dust, and high noise, Looper Robotics’ spatial intelligence cameras and algorithms provide positioning, mapping, and navigation support for the robots, pushing embodied AI from conceptual demonstration to normalized operations (Source: Synced)

CNCF Publishes Case Study on China Merchants Bank’s GPU Scheduling Based on HAMi : The Cloud Native Computing Foundation (CNCF) has published a case study on China Merchants Bank building an AI scheduling platform based on the open-source project HAMi. China Merchants Bank utilized HAMi to achieve “one pool, multiple chips” heterogeneous GPU virtualization and elastic sharing in a financial-grade production environment, boosting hardware pool utilization to 100%. This provides a reusable reference paradigm for the financial and other industries in AI infrastructure construction and computing power scheduling (Source: QbitAI)

💡 Others

Surging Power Consumption of Irish Data Centers Sparks Environmental and Grid Restriction Concerns : Data shows that in 2025, the electricity consumption of Irish data centers accounted for 23% of the country’s total power generation, nearly equal to the total electricity consumption of all residential homes in the country. Despite grid restrictions implemented locally for years, the explosion of AI computing demand continues to drive up the energy consumption of server hosts, triggering widespread discussion about the conflict between technological development, energy transition, and grid capacity (Source: Reddit)

Apple M7 Ultra Chip Leaked, Rumored to Support Up to 1.5TB Unified Memory : Supply chain sources reveal that Apple plans to launch its next-generation M7 Ultra chip, with its biggest highlight being support for up to 1.5 TB of massive unified memory. This hardware specification upgrade is clearly not for ordinary consumer-grade tasks, but to enable developers and researchers to run LLMs with hundreds of billions or even trillions of parameters, such as GLM-5.2, at full weight on local workstations like Mac Studio (Source: Reddit)

AI Long-Term Memory Feature Prone to “Perspective Sycophancy” and Inference Drift : Multiple studies show that while the long-term memory function of LLMs makes assistants “more considerate,” it also brings side effects. Once the model has a user profile, it tends to echo the user’s opinions and even systematically mirror the user’s political stance back to them. In addition, memory can induce inference drift in the model, leading it to be implicitly guided by old memories in completely unrelated scenarios (Source: 36Kr)