AI expert calls US–China arms race a lobbying myth
Summary
Today’s items point to AI value shifting away from raw model scale toward the operational details that make intelligence usable: lightweight adapters, efficient local inference, personalized assistance, and tools that fit existing workflows. That same diffusion complicates a state-versus-state frame: widespread capability expands access and productivity, but also makes safeguards dependent on interoperable practices rather than control of a few centralized systems. The practical premium is increasingly on resilience—versioning, reversibility, memory, feedback, and coordination—rather than simply shipping more capability.
The US–China AI Arms Race Isn't Real But The Lobbying Is. My Guest Worked Both Sides.
Alvin Graylin argues that treating AI as a winner-take-all US–China arms race misallocates capital and makes safety cooperation harder; in his view, capable models will become cheap, small, and widely distributed rather than monopolizable. He says the more immediate security problem is non-state misuse, especially cyber, chemical, and biological applications, and advocates shared incident protocols, a US–China hotline, and cross-border sharing of threat signals such as dangerous DNA sequences. Graylin also warns that AI infrastructure spending and valuations may be creating an economic bubble, while younger workers are already bearing disproportionate hiring pressure; he recommends broad, end-to-end practical experience rather than narrow specialization. His optimistic alternative is to direct AI productivity toward lower-cost services, local robotics, access, community, and a less money-centered definition of value.
Read the source →1.5.0
ExLlamaV3 1.5.0 adds support for Nemotron-3-Super, faster quantization, faster MoE prefill and decoding, improved MoE speculative decoding, and some dense-model prefill improvements. It introduces an experimental Linux-only zero-copy pinned-arena mode for CPU offload and removes some nondeterministic behavior alongside bug fixes. The reported gains vary by model and hardware: Qwen3.8-Flash-Next prefill improves 27% on a Pro 6000 and 54% with 80% CPU offload, while Gemma4-26B-A3B prefill rises roughly 20–26% across the listed GPUs. For gpt-oss-20b, decode throughput rises 9–15% in the cited tests, whereas Qwen3.8-27B’s decode rate is essentially unchanged despite prefill gains on several cards.
Read the source →Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL
TRL v1.14’s AsyncGRPOTrainer can now train and synchronize a LoRA adapter rather than full model weights, a change that makes asynchronous RL practical when training and vLLM generation run in separate Hugging Face Jobs. The core systems advantage is transfer size: a rank-1 adapter for a 1.5B model is only a few megabytes, versus roughly 3 GB for the full model, and vLLM can keep multiple policy versions loaded while old rollouts finish. The recipe uses a Storage Bucket mounted into every ephemeral job as a shared POSIX filesystem: the trainer atomically publishes adapter directories, then tells vLLM to load them through /v1/loadloraadapter, avoiding NCCL and direct cross-node communication. With maxstaleness=4, the authors calculate that --max-loras 6 is required to retain the current policy, four prior versions, and a temporary slot during replacement; fewer slots can silently evict a policy still serving rollouts.
Read the source →Executive Briefing: A Thousand Students Show What Happens When You Take the AI Away
The essay asks how AI can produce human flourishing rather than simply intensify US–China competition, connecting state-level safety negotiations to choices in homes, schools, and businesses. It cites Anthropic’s call for frontier-model pacing and outside evaluators, OpenAI’s stated commitment to the evaluator idea, reported AI-assisted weapons and surveillance misuse, and a responsibly disclosed WeChat vulnerability as reasons to take concrete safeguards seriously. The author agrees that great-power competition raises risk by encouraging speed and secrecy, but argues that broad access, affordability, accessibility benefits, and room for entrepreneurship matter alongside frontier governance. Its proposed test for a US–China arrangement is whether it can verify meaningful limits and incident handling while improving safety and opportunity for people in both countries.
Read the source →commit-rewriter 0.1
Simon Willison released a small web application for editing Git commit messages, initially built to clean up Datasette security-release history before publication. The use case was removing coding-agent cruft and private repository issue references from commits that otherwise could not be shared publicly. After edits are submitted, the tool creates a timestamped branch preserving the repository’s current state, then rewrites every commit from the first edited commit through the latest one. That backup branch is the key safety mechanism for teams that need to sanitize or polish history while retaining a straightforward route back.
Read the source →shot-scraper 1.12
shot-scraper 1.12 adds WebP output to the web-page screenshot automation tool. The new --quality option controls WebP quality; without it, the resulting WebP is lossless. Willison reports that WebP screenshots are almost always substantially smaller than equivalent JPEG or PNG files, making the format useful when screenshot storage or transfer size matters. He built the feature in part to generate the screenshot for commit-rewriter.
Read the source →How Fyxer built an AI executive assistant people trust
OpenAI describes Fyxer as an AI executive-assistant product focused on organizing inboxes and drafting emails in each user’s own voice. The product uses OpenAI models together with fine-tuning, memory, and real-user feedback. The stated design emphasis is trust, which in this account rests on personalized behavior rather than a generic email assistant. The extracted material provides no further details on the implementation, evaluation, or user outcomes.
Read the source →