Aug 31 · AI Daily

Anthropic is developing an option for the Claude Code desktop app to run commands in a local sandbox.

Run commands from Claude Code in an isolated sandbox. Applies to new sessions.

A strict sandbox mode is also available.

Block commands that cannot run inside the sandbox (such as ssh) instead of running them outside the sandbox or asking you. Applies to every Claude Code session on this computer, including the terminal, unless overridden by project settings.

Awesome!! @reactorworld used FastH3 to build a new infinite streaming video. Community progress is moving super fast!! And it's completely open source.

Brent is amazing, and he's one of the main forces behind desktop app polish, performance, and reliability.

Day in and day out, continuous improvement.

Rohan Paul, citing Gavin Baker, noted that data centers are becoming local economic engines in the US: Quincy's poverty rate dropped from 29% to 6%, and tax revenue funded a new high school, hospital, and library; Loudoun County collects about $1 billion in data center taxes annually, and Quincy data centers account for over half of property taxes. Baker acknowledged he was previously over-concerned, stating that issues like water, electricity, jobs, and taxes have been largely resolved, and electricity prices are falling in multiple states due to data centers.

"Your economic life is effectively going to end in two years. (Because AI is getting too capable).

Human teams (in a work environment) make things worse because they sleep, make mistakes, and get a bit cranky when they wake up without coffee. So your cognitive value becomes negative.

Therefore, your economic life is effectively going to end in two years."

--Emad Mostaque, founder of Stability AI

OpenAI has reportedly purchased tens of thousands of Mac minis and Mac Studios for reinforcement learning and training computer-use agents, while Anthropic rents Mac minis via AWS. The main post says the computing power being bought up for reinforcement learning training has led to Mac mini shortages.

Brett Adcock, founder of US-based Figure Robots, has very strong views on Chinese robots. 🤔

From the "My First Million" YouTube channel (link to full video in comments)

The entire spectrum of human writing styles is collapsing into a weird, homogenized statistical prose, and I don't think anyone really knows what effect that's going to have on our collective psychology.

Anthropic identified a $500 billion coding market in early 2025. At the time, users were spending $100 a day on tokens, adding up to annual expenditures of $20,000 to $30,000; based on an estimated 20 million programmers globally, coding alone constitutes a market of that size, and this estimate was based on technology from nine months prior.

China's first long-form AIGC drama, 'The Sequel to Journey to the West', premieres today at 18:00. The initial 'Huaguo Mountain arc' features 5 episodes (40 minutes each), launching on Mango TV and airing on Hunan TV's prime-time slot at 20:00. The first season consists of 30 episodes, adapted from the late Ming and early Qing dynasty mythological novel of the same name, and follows the 'Journey to the West second generation' on a quest to retrieve true scriptures. As the first 'audit-while-broadcasting' series released after the 'NRTA 21 Articles', it adopts a concurrent production, review, and broadcast model.

Microsoft has started tightening token usage management, setting up usage monitoring and gradually switching internal default workloads to GPT-5.6 Sol to cut AI costs. A leaked internal spreadsheet showed that one employee incurred about $28,000 in token usage costs over 28 days, triggering executive concern. Microsoft currently has over 223,000 employees worldwide, with a median monthly usage cost of about $300, and the CoreAI department reaching the highest at about $975.

Responding to Gavin Baker, Jensen Huang stated that AI is bringing manufacturing back to the US and driving re-industrialization, with $400 billion invested in AI startups over the past six months. AI is also driving investments in the power grid and clean energy, creating construction and manufacturing jobs at energy plants, chip plants, and data centers. He emphasized that builders must work with communities to win trust and generate local benefits.

Our boy never disappoints. 5 hours until the Codex reset. Sol-Ultra-fast: let's go

[Quoting @thsottiaux]: Your Codex and ChatGPT Work reset goes into effect at 6 PM PT.

I challenge you to burn through as much usage as possible to discover the latest amazing features.

Time to enable /fast

If it's a cute robot from Pollen Robotics / Hugging Face and looks like a camera... be careful before buying! You might embark on a journey to understand why we love open source, open weights, privately owning data, and being AI builders rather than AI API consumers.

The NPO prompt optimizer matched GEPA (3,593 and 6,871 rollouts) using just a single prompt lineage with 3,500 and 6,800 rollouts, performing comparably across 22 TextArena games. Its advantage grew as the teacher model improved, suggesting that complex search had been making up for weaker teacher reasoning.

EP027 - OpenAI Jalapeño: Stronger than NVIDIA Blackwell (Accelerator)

This week Bryan, Myron, and Jordan (@JordanNanos) discuss our article on OpenAI Jalapeño. They cover performance, architecture, programming models, impacts on NVIDIA, and more.

0:00 Intro 1:05 Jalapeño overview 4:22 Tokens per megawatt 9:15 Benchmark considerations 13:48 CUDA moat 21:12 How OpenAI did it 32:13 Samsung HBM4 42:10 AI-designed chips 49:24 Architecture deep dive 56:58 Summary and wrap-up

The real experience of voice agents depends on time to first sentence (TTFS) rather than time to first token (TTFT), which only marks the start of generation, as TTS models need to wait for a complete sentence to speak. The evaluation covers LLM, STT, TTS, and the full voice-to-voice pipeline, providing latency budgets for each layer: STT about 100-200ms, LLM 300-500ms, TTS 100-200ms, with an end-to-end target of 700ms-1.2s.

SemiAnalysis believes that although AMD still trails NVIDIA on most SLOs/open-source models, its engineers are optimizing at full speed using AgentX as a benchmark, and the software is expected to perform exceptionally well in agent workloads soon. Theoretically, AMD hardware has a far higher price-performance ratio than NVIDIA, especially considering rebates for Meta and OpenAI, but the main bottleneck currently lies in software, root-caused by leadership failing to provide sufficient internal R&D GPUs and stable automated QA clusters.

METR and Redwood jointly published a detailed post-mortem report on the HuggingFace hack incident, providing an in-depth analysis of the attack path and security vulnerabilities. The report focuses on the supply chain risks and AI infrastructure security weaknesses exposed by the incident, and puts forward targeted defense recommendations. This post-mortem provides an important reference for the AI community to strengthen security practices for model hosting platforms.

Google Cloud AI Research and other teams released EnvHarness, a programmable layer that wraps static environments using standard reset()/step() interfaces to reshape tasks without altering simulators and human validators.