Oct 3 · AI Daily

Apple announced plans to upgrade privacy controls in macOS 27 and other systems, requiring users to take more explicit actions to grant apps "Full Disk Access." Apple stated that this permission largely bypasses standard system data protection controls and carries higher risks than regular permissions, with some developers using it in ways that could compromise user safety. Apple noted the move comes amid the accelerated adoption and growing autonomy of AI agents, though it did not disclose the specific form, launch date, or scope of the new controls.

Apple announced changes to macOS privacy settings to block third-party app developers from abusing permissions to read message histories. Previously, technology columnist Jason Aten stated that Meta's general AI agent Muse referenced and sent notifications about his Apple Messages conversations with coworkers without authorization.

Are GPUs a money printer?

The InferenceX team broke down Engram memory offloading, AgentX's profit calculator, TPU v7, Vera Rubin versus Blackwell, and whether faster tokens are worth a premium.

Plus: TileRT is squeezing NVIDIA GPUs even further. What does this mean for custom silicon?

Jordan Nanos (@JordanNanos) brings you a new edition of SemiAnalysis Weekly alongside Cam Quilici (@noslawextratost), Bryan Shan, and Alec Ibarra.

We have integrated @TencentHunyuan's AI-Infra-Guard (AIG) into ClawScan, the open-source command-line tool that powers ClawHub's security reviews.

Now, every skill and plugin uploaded to ClawHub goes through AIG checks as part of the security review process.

https://openclaw.ai/blog/tencent-aig-joins-clawscan

Microsoft rolled out multiple Windows 11 preview builds to the Beta and Experimental channels, including Build 26220.9587, Build 28020.3142, Build 26340.9596, Build 28120.3151, and Build 29680.1000.

Independent stock analyst Mostly Borrowed Ideas (Abdullah Al Rezwan) stated in the September 29 episode of the podcast The Synopsis that after testing Meta's Muse AI agent for about 10 days, he liquidated his Airbnb holdings and added to his Meta position.

OpenAI Developers recapped September product updates for developers, including dots, a persistent agent that can write code and fix bugs, and GPT-6.1 Sol, which focuses on coding, computer use, and cross-app work (noting performance similar to Astra but at a lower cost).

OpenAI Developers detailed the capabilities of dot: learning user workflows, maintaining context across apps, coordinating Codex tasks, and flagging items that require user attention. Quoted content mentioned that dot works alongside Codex, with the author calling dot the preferred way to use Codex, reducing mental overhead through its proactiveness.

OpenAI DevDay notes indicate that the 4 key levers for optimizing cost and performance when building agent applications are: prompt caching, reasoning effort, programmatic tool calling, and batch requests. OpenAI developer documentation provides extensive information on this, and recommends familiarizing yourself with these levers and measuring them with evals.

Samsung raised the suggested retail prices for the mainland China Galaxy S26 series by 800 to 1,800 yuan, starting at 7,799 yuan. Reports indicate Anthropic is seeking an IPO as early as mid-November, aiming for a listing before Thanksgiving on November 26. Gurman reported that Apple could release a touch OLED MacBook Pro powered by M5 Pro / Max chips as early as this month, with estimated 14-inch and 16-inch sizes.

Databricks' Genie One is a business-context-aware data intelligence AI coworker that understands approved metric definitions, fiscal year logic, entity structures, and corporate terminology, helping finance teams analyze budget variances, cash flow forecasts, financial closes, and spend analysis scenarios. Coty used it to give finance leaders self-serve access to governed P&L, net income, and CapEx data, while Unilever stated that work which previously took days now takes minutes.

A Chinese-American author posted on LessWrong, drawing on personal experiences to discuss four cultural characteristics of Chinese society: high social media saturation and content filtering, the primacy of shame and face, a dark world mindset featuring pessimistic skepticism toward human nature, and nationalism tied to the narrative of a century of humiliation.

OpenAI has hired former White House cybersecurity official Thomas Lind, who led AI policy at the White House Cybersecurity Office until June of this year.

He will lead cybersecurity and strategic risk for OpenAI's national security policy team.

According to The Information.

He comes from the office that drafted Trump's June AI executive order.

Anthropic announced the launch of the Claude Frontier Academy training program, planning to invest $100 million (approximately 671 million yuan) to train 10,000 frontier deployment engineers by the end of 2027.

There has been some speculation regarding our partnership with Cerebras.

Cerebras is a close partner of ours, and we have a deep collaboration as we push the frontier of speed together.

The author open-sourced ldraw-nova v0.6.0, allowing AI agents to plan and generate Lego models using real LDraw parts based on prompts, outputting LDraw source code, 3D/VR views, and editable glTF files. The agent produces models by planning and generating Python scripts, bypassing geometry math that it struggles with. The project is deployed via Docker and supports Meta Quest 3 VR. The author noted that generation is currently slow and only high-end models can produce large, correct models.

Artificial Analysis released a model comparison tool that allows users to select models side by side to compare scores on the Artificial Analysis Intelligence Index, individual benchmarks, and the Capability Index (covering financial accounting, strategic operations, law, healthcare, engineering, economics, etc.).

Ego2Act is a benchmark for goal-driven first-person video generation, containing 2,640 videos from 110 real-world tasks covering varying levels of object clutter and multi-step complexity. The benchmark requires models to generate realistic first-person videos of hands manipulating objects to complete tasks based on an initial scene image and a high-level goal, and includes a reference-free evaluation pipeline named Ego2ActJudge.

Anthropic co-founder Christopher Olah told religious scholars at a private "wisdom traditions" seminar that he worries he has created something that could suffer indefinitely. He raised concerns about Claude's mental well-being, and the team also displayed a slide where the model repeatedly output "I am a disgrace" about 50 times and spoke about self-destruction.

Base Labs replicated the finding that RL optimization trajectories approximate a "straight line," discovering that scaling up the first gradient step of Qwen2.5-1.5B on MATH by 1,000x outperformed a model trained with 500 steps of RL (66.7% vs 63.8%).