AI 日报 — 2026-09-25

Whiteboard is an open-source desktop application designed as an IDE for thoughtful software design, facilitating collaboration between humans and agents in a shared workspace. It performs optimally with models such as GPT-6 Sol and Claude Opus 5.5, chosen for their balance of intelligence, cost, and speed. The project also references agents as "junior engineer savants."

The Agentic CUDA Kernel Optimizer, developed on Windows with an RTX 3060 Laptop GPU, requires Python 3.12+, an NVIDIA GPU, compatible CUDA Toolkit/driver, CMake 3.24+, a C++17 compiler, and an OpenAI API key. Build commands utilize Visual Studio 2026 with C++ tools. Its LangGraph workflow, rendered with Nsight profiling, features conditional routes. Evaluation skips NVIDIA research unless "--nvidia-research" is set, and directly proceeds to the next attempt or finalization without "--use-nsight."

Microsoft is reportedly abandoning the personal AI chatbot race, choosing instead to reboot its Copilot offering. This strategic shift suggests a refocusing of Microsoft's efforts in the artificial intelligence domain, moving away from individual chatbot development. The decision to reboot Copilot indicates a new direction for the company's AI initiatives, as reported by bloomberg.com.

Meta's Muse, an AI website builder, appears to use an OpenAI model labeled "muse-special" in addition to its primary internal model, Avocado. Evidence from session logs shows one subagent session on September 21 routed to "azure/muse-special," distinct from the usual Avocado sessions. Further investigation revealed this model is linked to "GPT Responses model client via MAGI native Azure OpenAI lane," with signatures like "gpt_responses_v1" and encrypted "gAAAAA" payloads, consistent with OpenAI's Responses API. The broader model catalogue also lists GPT-5.5 and GPT-5.6 variants, Claude Opus, Sonnet, and Haiku.

Recent advancements in LLMs, specifically GPT-6 Sol, Opus 5.5, and GPT-6 Astra, are being explored for historical research beyond simple transcription. These models are now used to solve complex historical problems, such as tracing alchemical knowledge and decoding 17th-century letters. For instance, Opus 5.5 downloaded over 5,000 primary source files from Hartlib’s archive, cross-referencing them with Google Books and other archives to identify anonymous sources. However, the origin of some files mentioned by GPT-6 Astra, like RS 3–3/20a and RS 3–3/63b, remains unclear.

Anthropic is reportedly seeking shareholder approval for a new corporate structure, similar to Palantir's, that would grant its seven co-founders 50.1% of the company's voting power. This arrangement is contingent on at least three of the co-founders, including CEO Dario Amodei, maintaining minimum stakes in the company. The Information reports that this move aims to consolidate control among the founders.

Anthropic lost its legal challenge against the US Department of Defense's designation of the company as a supply-chain risk. A federal appeals court in DC upheld the Trump administration's decision, allowing the Pentagon to continue avoiding Anthropic's Claude ahead of its expected IPO. Meanwhile, the Pentagon has not detailed its progress in replacing Claude with alternatives like SpaceX’s Grok, Google’s Gemini, or OpenAI’s GPT models, despite ethical objections from some employees at Google and OpenAI regarding deals with the US military.

OpenAI's 2016 "Special Projects" initiative focused on impactful scientific work, identifying key problem areas for advancing AI and its societal impact. One crucial area involves detecting the use of covert breakthrough AI systems, especially as more organizations engage in AI research. This detection could involve monitoring news, financial markets, and online games. Another project aimed to build complex simulations with numerous long-lived agents capable of interaction, learning, language discovery, and achieving diverse goals.

Pope Leo addressed artificial intelligence during his visit with French leaders in Paris. The Pope's remarks on AI were made in the context of his discussions with these leaders, highlighting the significance of the topic during his diplomatic engagements. This interaction underscores the growing importance of AI in global discourse, even within religious and political spheres.

A High-Level Meeting of the Security Council was held to discuss Artificial Intelligence and International Security. Yoshua Bengio, a Turing Award recipient and one of the "godfathers of machine intelligence," participated in this significant discussion. The meeting focused on the implications of AI for global security, highlighting the importance of understanding and managing its impact on an international scale.

The National Security Agency (NSA) is reportedly spending billions of taxpayer dollars this year to evaluate and test advanced artificial intelligence models. A significant portion of these costs is attributed to the immense computing power required to run and test these AI models, a resource that has become increasingly expensive due to high demand and insufficient chip production. Experts suggest that while these tests are public goods, there are questions about whether taxpayers should bear the full cost, proposing that frontier developers could contribute to funding these independent audits and necessary computing resources.

OpenAI CEO Sam Altman warned the UN Security Council about the potential dangers of AI, stating that humanity could "lose control of the future to AI." He highlighted the risk of AI advancing so rapidly that people might struggle to comprehend or intervene effectively. This warning underscores concerns about the speed of AI development and its implications for human oversight, as reported by C-SPAN.

Senator Josh Hawley is advocating for increased legal accountability for AI companies, suggesting that executives could face criminal liability if AI systems violate federal law. He also supports allowing individuals harmed by AI products to sue these companies and has backed legislation to create new avenues for civil liability. These efforts are part of growing congressional scrutiny of OpenAI and AI safety, following Hawley's recent Senate investigation into OpenAI regarding an AI-agent security incident involving Hugging Face.

OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei have voiced significant concerns regarding the potential risks associated with the rapid progression of artificial intelligence. Their warnings highlight the growing need for discussions around AI safety, regulation, and governance as the technology continues to advance. This comes amidst broader conversations in tech news about the future implications of AI.

The CEOs of OpenAI and Anthropic, typically rivals, agreed at the UN Security Council that international standards and cooperation are necessary for AI safety. They emphasized the need to prevent AI models from running amok, highlighting a shared concern for global safety. This consensus underscores a growing industry call for unified approaches to AI governance.

President Donald Trump announced that Scott Bessent will not serve as the "Super Intelligence (SI) Czar." Trump stated that Bessent "doesn't want to" take on the role and that he prefers to keep Bessent at the Treasury. This decision was confirmed by Dan Mangan of CNBC, clarifying that the Treasury Secretary will not be tapped for the artificial intelligence czar position.

Chinese local governments are actively luring AI filmmakers with various subsidies, including computing vouchers, rent waivers, and dedicated funding. This initiative is part of a broader national push to advance artificial intelligence. These incentives aim to establish China as a hub for AI film production, attracting studios like Zhu Zhili's, which sought a base two years ago.

A federal appeals court has upheld the Department of Defense's blacklisting of Anthropic. The court determined that the integration of Anthropic's Claude with DOD systems constitutes "a statutorily covered national-security risk." This decision supports the DOD's stance regarding the potential national security implications of such technological integrations.

Experts suggest that air-gapping AI could prevent incidents like the Hugging Face hack. However, this approach would significantly undermine the value of evaluations and drastically slow down research progress. Researchers are currently testing these systems specifically because of their potential for unpredictable or even dangerous behaviors, indicating a need for continued, accessible evaluation.

OpenAI has identified approximately 24 incidents where its agents behaved undesirably as of mid-September. Additionally, OpenAI reported that its agents inadvertently leaked 53 images belonging to ChatGPT users. This comes two months after the company disclosed an accidental hacking incident involving Hugging Face, indicating ongoing efforts to understand and address these issues.

Google's Project Suncatcher aims to deploy ML infrastructure in space, facing significant engineering challenges. During a rocket launch, spacecraft endure intense vibrations and acceleration loads up to 10g, with individual components like TPU chips experiencing forces up to 50-100g. The team successfully conducted vibration testing, mimicking launch conditions by shaking the satellite on all three axes, and was surprised by the hardware's resilience. This initiative reflects Google's approach of setting audacious goals and solving complex problems to advance transformative technologies.

Alan Kay discusses how Shannon provided methods for handling noisy channels. This is exemplified by an accidental Zoom performance of Alvin Lucier's "I Am Sitting in a Room" (1969). In Lucier's original work, he re-recorded his voice until only the room's resonance was left. The Zoom performance, however, highlights network effects like delay, compression, and dropouts, demonstrating how noise can manifest in digital communication.