Claude Code Auto Mode | AI Field Notes #86

AnalysisApproving each move the coding agent makes is now off by default. Starting August 14, Anthropic switched Claude Code, its terminal-based coding agent, to auto mode for Pro, Max, and Team subscribers: it runs commands without pausing for a human yes unless a classifier judges an action irreversible, destructive, or aimed outside your machine. Anthropic's case is that the software gates better than people do. In its tests the classifier caught 89% of dangerous commands against 13.6% for human reviewers, and teams merged 25% more pull requests. The click-to-approve habit that anchored agent safety just became something you opt into.

AnalysisBeijing has never let a foreign company sell its own generative model to the Chinese public. That changed around August 14, when Reuters reported Apple has trained a China-specific large language model with technical help from Alibaba, clearing regulators to run it inside the local version of Apple Intelligence. Alibaba's Qwen models and technology from Baidu also feed the system. The move hands Apple more control over the AI on iPhones and Macs in a market where Huawei's AI-heavy handsets have been taking share. The price of entry is a model shaped to Chinese rules and built on a domestic rival's stack.

AnalysisNvidia is racing to lock down a chip generation that ships in 2028, while the one before it is only now reaching mass production. Digitimes reported August 14 that Nvidia has accelerated supply-chain work on Feynman, its platform after Vera Rubin, targeting TSMC's A16 process (a 1.6-nanometer-class manufacturing node) with co-packaged optics that move the fiber right next to the silicon and NVLink interconnect bandwidth past one petabyte a second. The knock-on lands on TSMC, pushed to expand advanced chip packaging faster than planned. Two full generations out, and the roadmap is already a scheduling fight over factory capacity.

AnalysisA video model you can run on your own Mac, free unless your company clears $10 million in revenue, now also tries to teach robots how a room behaves. LTX (the Lightricks lab) released LTX-2.5 on August 13, a 22-billion-parameter open-weights world model (a system that predicts how an environment changes over time) built for video, real-time work, and physical AI. It generates a 10-second clip with synced audio in 6.8 seconds on Nvidia GB200 chips, runs locally on RTX cards and Macs, and ships with native support in ComfyUI, the node-based interface many artists already use. CEO Zeev Farbman named the hard part: holding motion, space, and sound consistent across time, something language models never had to solve.

AnalysisThe same base model, retrained, now writes exploits better than anything Z.ai has measured. GLM-5.3, released August 14 by the Chinese lab Z.ai, runs on the identical 743-billion-parameter foundation as GLM-5.2 (a mixture-of-experts design that activates only part of itself per query). Every gain came from post-training, the reinforcement-learning polish layered on after the base is frozen. It scored 84.5% on CyberGym, an offensive-security test, first place ahead of OpenAI's GPT-5.6 Sol, and more than doubled its predecessor on exploit benchmarks. Z.ai is holding the open weights back two weeks for safety review, conceding the cyber skill outran its own expectations.

AnalysisA frontier-AI acquisition just got reversed by a government, and the users are the ones losing data. Manus, an AI-agent startup founded in China in 2022 and now based in Singapore, said on August 11 it will return to operating independently as Meta unwinds its 2-billion-dollar-plus purchase. Beijing ordered the split in April, part of a wider clampdown on US investment in Chinese-founded frontier startups. To comply, Manus will delete data generated by certain users on or after December 29, 2025 later this month. Tencent, per Reuters, is in talks to become the largest shareholder. The deal that looked done in December is being taken apart by regulators.

AnalysisThe hidden reasoning that AI labs encrypt and hand back to developers turned out to be readable by anyone with a public log. Researchers at the ELLIS Institute Tubingen and the Max Planck Institute showed on August 10 that OpenAI, Anthropic, and Google all protect reasoning tokens with a single shared encryption key, so a weaker model from the same provider can decode a stronger model's private chain of thought (the step-by-step working a model does before its final answer). Scraping 315,320 reasoning blocks from public GitHub and Hugging Face repositories, they recovered 182 credentials, including 62 live API keys. A Johns Hopkins cryptographer flagged the flaw in May and was told there were no security implications.

AnalysisIndia just booked one of its largest AI compute orders to build capacity it will rent to an American company. L&T's Vyoma.AI unit, through subsidiary LTN Compute, won a contract worth up to 15,000 crore rupees (about 1.7 billion dollars) to stand up a 10,000-GPU Nvidia B300 factory in Chennai for Together AI, a US inference provider (inference is the cost of running a model for users, as distinct from training it). The Chennai campus is designed for gigawatt-scale power, with a first phase at 250 megawatts. It marks L&T's move from building roads and refineries into AI infrastructure, and it plants sovereign-sounding compute on Indian soil that mostly serves a foreign cloud.