Sep 18 · AI Daily

Meta officially announced the immediate launch of Muse for Mac (currently US only). With explicit user authorization, the personal agent can complete tasks directly on the computer, including organizing download folders, finding lost files, and summarizing messages and notes. It works across apps, files, calendars, memos, and messages, asks before performing sensitive operations, and lets users control its access permissions.

ChatGPT is now officially integrated into Microsoft Word, allowing users to turn rough notes into first drafts, smooth out paragraphs, proofread, and give revision suggestions directly inside documents, as well as catch formatting issues. OpenAI team member Sherwin Wu stated that usage of ChatGPT for Excel and PowerPoint has grown significantly lately, and this Word launch completes the full Office integration, further raising the stakes in workplace software competition.

The OpenAI developer account announced the integration of Appshots with Windows, letting ChatGPT access the app context users are working in to cut down on time spent repeatedly explaining background information. Official use cases include debugging problems, grabbing UI to replicate, and pulling data from other apps, removing the need for heavy copy-pasting as it moves toward an operating-system-level assistant.

OpenAI released Astra for Law, bundling GPT-6 Astra with legal search indexes, Legal analysis and writing instructions, and model-level settings tailored for law firms and legal tech companies. On the Vals AI Legal Research Bench private validation set of 200 US legal research questions, it achieved an overall accuracy of 54.0% at maximum reasoning strength, a 40% relative improvement over the 38.7% achieved using web search alone.

The Federal Aviation Administration is launching an AI air traffic control software called Smart (Strategic Management of Airspace, Routes and Trajectories) under a 12-year, $875 million contract to improve airspace, route, and trajectory management. It marks a major milestone for AI entering critical civil infrastructure.

The Bend language has officially launched, positioned as a way for humans to precisely constrain AI-written code in the AGI era. Developers declare rules through LAWS.bend and require the AI to submit proofs, and code that breaks the rules mathematically cannot be merged. The creators also claim single-core performance close to C, and the same binary can run across 16 cores or GPUs for a hundredfold speedup.

The Microsoft Azure Kubernetes Service team has open-sourced TauGrid under the MIT license. A single Helm installation packages the tau CLI, Kueue queues, KubeRay orchestration, GPU health monitoring, and observability components, providing an out-of-the-box, complete tech stack for running GPU AI workloads on Kubernetes.

Anthropic released three metrics measuring front-end AI research progress (degree of AI participation in AI R&D, agent supervision, and compute allocation) along with an internal snapshot. In about six months, the share of model R&D tasks led by Claude grew from 1% to 26%, with over 90% of measured research and development work deeply involving AI. The internal platform has roughly 30,000 agents engaged in research and engineering at any given time, offering a quantitative benchmark for the progress of AI building AI.

An unsealed, unredacted court document from the copyright lawsuit brought by The New York Times against OpenAI features multiple internal statements from executives at both companies: LLMs are built on content described as unprecedented theft, triggering a doom loop that destroys the entire web. The filing also shows that clicks on news sites stolen from on Bing dropped by over 90%, and that OpenAI once bypassed The New York Times paywall while calling itself an existential threat to news publishing.

karminski introduced the Jev model, which discards traditional autoregressive architecture and cannot output plain text directly, instead producing probabilistic structured JSON decisions based on predefined schemas (such as isSpam true 0.982). Its moat is RLCD-calibrated reinforcement learning, making its outputs resistant to probability distortion. The trade-off is rigid behavior making it suited strictly for execution tasks; it currently supports text input only, and complex scenarios require predefined selectable actions.

OpenAI launched 47 community plugins built by legal experts on the ChatGPT plugin store, alongside 26 partner plugins with collaborators including Thomson Reuters, Harvey, Legora, and iManage. These plugins were independently published by legal tech community members such as LegalQuants, Skills.law, and LECG rather than being white-labeled by OpenAI, as the legal vertical ecosystem accelerates its integration with ChatGPT.

Based on 616,300 OpenAI responses from internal usage, SemiAnalysis analyzed tool calling frequencies: GPT 5.6 Terra averaged 1.94 client tool calls per response, Sol averaged 1.00, Luna averaged 0.97, and GPT 6 Astra averaged 1.18 across 11,717 responses, providing quantitative data to compare the agentic levels of various models.

Anthropic published a science blog post stating that Claude optimized inference for over 30 open-source models used in molecular modeling, drug design, and genetic mutation prediction, making them an average of 4x faster, in part by writing custom software for GPUs. All optimization code has been open-sourced, highlighting AI's practical acceleration capabilities in scientific computing scenarios.

An Android Authority survey shows that over 70% of respondents are unwilling to let AI tools like OpenClaw fully take over their computers. Among them, 43% explicitly refused due to security reasons, 27% were completely uninterested, about a quarter were open to it provided safety is properly protected, and under 4% had no worries about risks, showing that security trust is a key bottleneck for desktop agent adoption.

A user complained that an AI product changed its model selection to a routing mechanism, no longer allowing manual selection of tiers like max or xhigh, and questioned whether the routed model is worth the paid price. The user noted that benchmarks fail to reflect routing performance, and that the model routed to during yesterday's testing was not a SOTA model, sparking discussions on routing transparency and payment fairness.

Satya Nadella shared how Microsoft draws on Toyota's lean manufacturing philosophy internally, applying lean knowledge work to AI-driven white-collar operations. Microsoft spends roughly $4 billion annually on customer support. By deploying AI agents for front-end interception and real-time reasoning assistance, the company has significantly cut costs in areas like Xbox and Azure support, offering a reference model for large enterprises seeking to lower costs with AI.

Balyasny Asset Management evaluated Claude Fable 5 on thousands of real financial tasks, achieving a score of 89.4%, higher than the previous production model's 86.1%. Its strengths are concentrated in complex planning, analysis, and agent execution. The firm's self-built BAMAgent platform supports thousands of autonomous agents running 24/7 and uses Fable as its preferred model for the planning and analysis stages.

Gary Marcus wrote an article refuting the false dichotomy pushed by the tech-libertarian right, which uses AI responsibility to oppose regulation, arguing that responsibility and regulation must go hand in hand. Citing views from Mark Cuban and Karen Kornbluh, he pointed out that accountability is meaningless without regulation tailored for AI. Looking back at his May 2023 clash with Josh Hawley in the US Senate, he argued that current laws have gaps in areas like copyright and large-scale disinformation, and relying solely on lawsuits could take over a decade.

AI data center company Crusoe announced the close of its $3.9 billion Series F funding round, pushing its valuation to $30.9 billion. The round was co-led by Atreides Management, Mubadala Capital, and Valor Equity Partners, with participation from Nvidia, Founders Fund, GIC, and others. The funds will be used to build large-scale data centers and small modular AI factories, showing that investment in AI computing infrastructure continues to ramp up.

Researchers from Google and Google DeepMind launched the DeepMind Institute on Wednesday, with Shane Legg, James Manyika, and Demis Hassabis serving as directors, and Legg as executive editor. Designed to present various viewpoints surrounding AGI alongside four initial papers, the move is seen as an official signal to promote diverse discussions on the topic of AGI.

PrismML released Ternary Bonsai 2 27B, based on Qwen3.8 27B, using ternary weights with FP16 grouped scaling for an effective 1.76 bits per weight. Its size is just 5.9GB, over 9 times smaller than the full-precision model, with an overall benchmark score of 83.9, retaining 98.2% of performance. It can run on 8GB memory devices, significantly lowering the barrier for local deployment.

Zimperium's zLabs research team reported a new Android malware Trojan named RatHat. It sends the infected device interface in XML format to a popular AI assistant, which identifies element center coordinates, reads on-screen text, and returns navigation commands like SCROLL_DOWN to help attackers remotely control the device. This demonstrates how AI capabilities are being weaponized, serving as a warning to the security industry.