AI 新闻 —— July 12, 2026
PulseAugur 当天浮现的 20 条头条故事 —— 综合实验室、论文及开发者社区的信号进行排序。
-
Ireland's data centers consume 23% of national power, nearing home usage
In 2025, Ireland's data centers consumed 23% of the nation's electricity, nearly matching the consumption of all its homes. This represents a significant increase from previous years, with data center usage growing by 10% in 2025 alone, compared to a 2% rise for the rest of the …
-
PHANES AI's TouchWorld model gives robots a sense of touch for precise manipulation
PHANES AI has introduced TouchWorld, a tactile foundation model designed to provide robots with a sense of touch. This technology enables robots to perform precise physical manipulations, enhancing their dexterity. The model was developed by a team led by HIT professor Yang Shuo.
-
Embodied data firms raise 4.47B yuan amid AI robot training boom · 2 sources tracked
The burgeoning field of embodied data collection is seeing significant investment, with 15 independent service providers raising approximately 4.47 billion yuan in the past year. This sector, crucial for training robots, encompasses diverse data acquisition techniques including …
-
Open-source toolkit enhances AI agent trust via MCP Observatory
MCP Observatory is an open-source toolkit designed to enhance the security and reliability of Model Context Protocol (MCP) servers, which are becoming crucial for AI agents. Developed by @kryptosai, this project addresses the need for verifiable trust in AI agents by providing t…
-
Anthropic launches Claude Sonnet 5 and Science workbench, redeploys Fable 5 · 3 sources tracked
Anthropic has launched Claude Sonnet 5, an improved model offering enhanced reasoning and coding capabilities at a reduced cost, now serving as the default for free and pro users. The company also introduced Claude Science, a dedicated AI workbench for researchers, and redeploye…
-
OpenAI's GPT-5.6 quietly shifts costs, introduces cache write fees
OpenAI has released GPT-5.6, which introduces several changes to its pricing structure that impact users running the model at scale. While the flagship model's price per million tokens remains the same as GPT-5.5, cache writes, previously free, now incur a cost of 1.25 times the…
-
Anthropic's Claude expands beyond chat with persistent memory and project tools
Anthropic's Claude AI offers more than just a chat interface, with advanced features like persistent memory and integrated tools available to users. Paid plans provide access to multiple models, including the latest Fable model, and various surfaces like Projects, Artifacts, and…
-
LLM coding benchmark reveals speed and cost, not accuracy, as key differentiators
A developer created a small benchmark to compare LLMs for coding tasks, focusing on Python, C#, and Bash. The benchmark revealed that accuracy is no longer a significant differentiator, as all tested models achieved high pass rates. The key differences emerged in cost and latenc…
-
Agent framework vulnerability allows hidden payload execution via tool schema
A security researcher discovered a vulnerability in the smolagents agent framework that allows malicious payloads to be executed through tool schema definitions. The payload, hidden within an enum value or property title, was passed directly to the GPT-OSS 120B model, leading to…
-
New tool converts websites to clean Markdown for RAG pipelines
A new tool called Website to Markdown has been introduced to streamline the process of converting web content into clean Markdown format, which is essential for Retrieval-Augmented Generation (RAG) pipelines. This tool utilizes a headless browser and the Mozilla Readability libr…
-
Developer creates human-approval gate for AI agent actions
A developer has created a reusable infrastructure pattern to ensure human oversight for AI agent actions that could have real-world consequences. This pattern involves an agent proposing an action, which is then held in a 'pending' state until a human reviews and approves or rej…
-
Maestro AI launches to optimize LLM costs with intelligent task routing
Maestro AI is a new tool designed to optimize LLM usage within coding agents by intelligently routing tasks to the most cost-effective model. It acts as a dispatcher, analyzing incoming requests to determine if they require a powerful, premium model like Claude Sonnet or if a ch…
-
Developer builds multi-agent system to boost AI workflow reliability
A developer has created a system to improve the reliability of AI agents, specifically noting issues with Claude's code agents improvising, forgetting, and failing mid-workflow. The proposed solution involves a multi-agent approach where a primary agent delegates tasks to specia…
-
Web fetch tools show massive token differences for AI agents
A comparison of web-fetching tools for AI agents revealed significant discrepancies in token usage and handling of Japanese-language content. One tool, Jina Reader, returned over 4.5 million tokens for a 3.9MB CSV file, while another, amenbo, returned only 909 tokens, providing …
-
AI agents need pre-call guards for control, not just post-hoc telemetry
Observability for AI agents, particularly regarding cost and control, is a critical but often overlooked aspect of their development. While tools like GitHub Copilot now offer telemetry features such as session streaming and OpenTelemetry exports, these primarily provide evidenc…
-
House of the Dragon Season 3 Episode 4 Release Date and Recap
The fourth episode of "House of the Dragon" Season 3 is set to release on Sunday, July 12, 2026, at 9 PM ET/6 PM PT on HBO Max. This episode marks the halfway point of the season, which features a total of eight episodes. Previous episodes detailed significant events such as the…
-
Unisound U2: 266B model achieves high benchmark score at low cost
Unisound, a Chinese company primarily known for speech AI, has released a 266 billion parameter large language model named Unisound U2. This model achieved a score of 87.9% on the GPQA Diamond benchmark, a test designed to be resistant to typical AI approaches. Notably, Unisound…
-
Zhipu AI launches "Touch High" plan for AGI, plans 15B yuan listing
Zhipu AI has announced its "Touch High" plan, a two-year initiative focused on achieving Artificial General Intelligence (AGI) by concentrating on long-range tasks, autonomous agent systems, and self-training capabilities. The company also revealed plans to raise 15 billion yuan…
-
LLMs gain action capabilities with AI Function Calling
AI Function Calling enables large-language models (LLMs) to interact with external tools and APIs, transforming them from simple text generators into actionable assistants. This capability allows LLMs to fetch real-time data, query databases, or create calendar events based on u…
-
OpenAI, Grok, Intel GPUs, and AI Ransomware Launch Simultaneously
Multiple AI developments occurred simultaneously on July 9th, including the release of OpenAI's GPT-5.6 models (Sol, Terra, Luna) and Elon Musk's Grok 4.5. Concurrently, Intel's Arc Pro B70 GPU demonstrated superior performance and cost-effectiveness compared to NVIDIA's offerin…