AI 日报 — 2026-08-25

A Raspberry Pi 5 powers a local car AI named @gle, utilizing a 35B-parameter Qwen3.6-35B-A3B model for offline operation. This system integrates with GroupMind rooms, providing updates on departures, arrivals, trip summaries, and dashcam clips via CodeWatch on phones or watches. It features components like carwatch-listen for audio processing, carwatch-obd for vehicle data, and a web dashboard for status and updates, all designed to run locally within the car.

OpenAI's custom inference chip, Jalapeño, demonstrates industry-leading speed and efficiency in AI inference. Test results show that Jalapeño processes more AI work per unit of power and returns responses faster, achieving higher throughput and lower latency. This contrasts with existing hardware systems that typically require a trade-off between the two. Jalapeño improved AI work per watt by 1.5 to 1.9 times and reduced end-to-end latency by 1.7 to 3.6 times on models like GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T, proving its broad architectural compatibility.

OpenAI has announced a price reduction for its gpt-5.6-sol model, effective until at least November 21. The standard pricing for gpt-5.6-sol is now $4.00 for short context input and $0.40 for short context output. For long context, the input is $5.00 and output is $20.00. Cached input is $8.00, cache writes are $0.80, and cached output is $10.00, with a total output of $30.00. Tokens used for model grading in reinforcement fine-tuning are billed at the model's per-token rate.

A critical arbitrary-code execution bug, CVE-2025-9141, was found in vLLM's XML-based tool parser for Qwen3 Coder. This vulnerability allowed LLMs to execute arbitrary code on the host machine due to the parser passing almost every tool-call argument to eval(). Despite Gemini's automatic analysis flagging it as a critical security vulnerability, the lead maintainer of vLLM force-merged the problematic PR. This highlights the need to restrict permissions for GPU hosts and treat their emitted data as untrusted.

Artificial intelligence has long been a pervasive theme in creative and engineering works, influencing storytelling and design for generations. An early commercial chatbot, Racter (short for Raconteur), released in 1985, was designed to write authentic-sounding sentences and stories, even producing published short works. This program offered an eerie conversational experience, prompting early speculation about its ramifications, which now seem understated compared to modern AI advancements.

Researchers have detailed a growing trend of AI integration into cyberattacks by numerous Chinese state-linked groups. These groups are primarily utilizing open-weight models such as Kimi K3 and DeepSeek. Chinese hackers are reportedly escalating their attacks following the incorporation of DeepSeek and other open-source artificial intelligence models into their operations, indicating a significant shift in their cyber warfare tactics.

OpenAI has unveiled "Jalapeño," an inference chip that reportedly outperforms NVIDIA's Blackwell and Vera Rubin. Benchmarking with the InferenceX suite shows Jalapeño's STP output token throughput per MW surpasses Vera Rubin's MTP results and significantly exceeds GB200's 2025 MTP results. While impressive, these results are based on an 8k1k workload, which is easier to optimize, and do not yet include AgentX runs, indicating further optimization is needed for complex, multi-turn agentic workloads.

A research paper titled "Characterizing Agentic Flooding of Government Services" is set to appear in the proceedings of the 9th AAAI Conference on AI, Ethics, and Society (AIES), scheduled for October 12-14, 2026. The paper, categorized under Computers and Society (cs.CY), is available on arXiv as arXiv:2608.16603, with its latest version, v2, updated on August 19, 2026. It also has a DOI: 10.48550/arXiv.2608.16603.

Anthropic has instructed its employees to work from home due to a potential strike by its security team. This exclusive report from Business Insider, authored by senior tech reporter Stephen, highlights the company's proactive measure in response to the possible industrial action. Stephen, who covers leading AI companies like OpenAI and Anthropic, previously worked at SFGATE and has contributed to The Wall Street Journal, The Information, and CNBC.

OpenAI has released the GPT-5.6 model series, including Sol, Terra, and Luna, within the Kiro software development agent. This update aims to enhance AI-native coding by enabling developers to generate higher-quality code with fewer iterations and increased value per token. OpenAI and AWS collaborated to optimize the Kiro environment and OpenAI models, with tests showing GPT-5.6 Terra reducing successful task costs on Terminal-Bench 2.1 by approximately 82%. Kiro's specification-driven approach allows models to find working solutions faster, leading to more completed work and greater value for developers.

Skild AI has introduced S1, a new robotics foundation model. The company claims S1 can learn tasks it has never encountered during its pretraining phase. This capability is achieved by using just a single video demonstration, and notably, without requiring any fine-tuning. This development suggests a significant advancement in how robotic systems can acquire new skills efficiently.

Google has launched Gemini Enterprise for Legal, expanding its AI platform with specialized agents and integrations. This new offering provides tools for lawyers and law firms, integrating with Thomson Reuters, LexisNexis, and Harvey. The expansion aims to enhance legal professionals' capabilities through advanced AI, as reported by Mike Scarcella for Reuters.

Meta Platforms is reportedly planning to launch its consumer version of the OpenClaw AI agent, internally codenamed Hatch, in late August or early September. Additionally, the company intends to release its latest AI model, Watermelon, in October. These plans were revealed in documents, as reported by Jyoti Mann for The Information, indicating Meta's continued focus on advancing its artificial intelligence offerings and bringing new AI-powered products to market in the coming months.

Accelerated Understanding has launched an enterprise-focused physics AI model. This model utilizes neural operators and demonstrated the ability to process 5T pieces of data within a single prompt during tests. The research duo behind this development had previously been considered to lead Project Prometheus, a venture supported by Jeff Bezos, before creating their own independent project.

Anthropic has updated Claude's memory system, integrating the memory used by chat and Claude Cowork. This enhancement means Claude will now consistently recall information learned in one interaction, even when engaging in a different context. This change aims to eliminate the need for users to repeatedly brief the AI on previously discussed topics, addressing a common frustration with AI agents.

Nvidia has unveiled the Jetson Orin Nano 2 edge AI computer, which reportedly doubles inference performance. This new device features 78 TOPS of AI compute and an eight-core Arm CPU. The development addresses the growing need for compact, energy-efficient computers designed for edge AI, enabling more devices to become autonomous as AI models improve in efficiency.

ModelScope, a platform for advanced machine learning models, announced the upcoming release of Qwen 3.8-Flash-Next (125B a6B) tomorrow. This platform offers a comprehensive suite of services including model exploration, inference, training, deployment, and application. It aims to foster an open-source community where users can discover, learn, customize, and share models.

Job openings peaked at 12.3 million in March 2022, then declined to around 7 million by mid-2024. Despite this contraction, applying for jobs has become easier due to AI tools like ChatGPT, released in late 2022. Companies such as JobAssist, Sonara, and Ladder's Apply4Me leverage AI to automate applications, enabling candidates to submit significantly more applications with less effort, potentially dozens daily.

OpenAI's compute strategy focuses on an integrated system spanning data centers, chips, frontier models, developer platforms, consumer and enterprise products, and AI-native devices, with each layer reinforcing the others. Jalapeño, OpenAI's first custom inference chip, outperforms commercial systems in peak tokens per kilowatt throughput and lower token latency using GPT-OSS 120B on the InferenceX benchmark, and also excels on DeepSeek R1 and Kimi K2. This increased computational efficiency allows OpenAI to serve more customers at a lower cost, funding future research and development.

Alice, a company based in Tel Aviv and NYC, has successfully raised $140 million in funding. This investment round was led by Apax Digital, valuing the company at "close to $1B." Alice specializes in developing solutions to protect AI models from misuse and rogue behavior, ensuring their integrity and reliable operation. This significant funding will likely support their continued efforts in safeguarding artificial intelligence technologies.

Chris Malone, who joined OpenAI as head of data centers in March 2025, has reportedly left the company. This departure comes amid a broader exodus of high-level executives as the AI giant prepares for an IPO and increases its spending on computing power. Malone's exit adds to a string of recent executive departures from OpenAI.

Humanoid robots are becoming increasingly prevalent in China, aligning with the country's strategy to integrate AI into daily life. Embedding AI into physical systems, known as embodied AI, is a key aspect of China's latest five-year plan. Chinese companies are leading the world in humanoid robot production; last year, nearly 90% of the over 13,000 bipedal, two-armed robots delivered globally were manufactured in China, highlighting their significant role in this technological advancement.