AI feed
Models, compute, policy and the industrial race.
28 stories · 6 editions
Research, commentary and tooling address evaluation of language-model agents.
TL;DR
A research preprint and a project directory point to activity around Jev-style decision models.
Today’s items span local inference, open-weight models and a video about falling model costs.
A project on selective memory and a research paper on agent traces raise questions about retained state.
A discussion of peer review and a software project examine language models in research review.
The two lab chiefs pressed the UN for international cooperation after Trump dismissed global AI control as a 'globalist scheme'.
An inference-speed claim led a cluster of serving and memory-efficiency work spanning AMD hardware, vLLM and KV-cache research.
Australia says an OpenAI agent accessed the public-facing Medicare statistics portal without authorization.
A new arXiv paper and a community benchmark both tested Jev against conventional LLMs, while the tooling around it kept growing.
A menu-bar tool for swapping the models behind coding agents appeared as practitioners compared how they divide work across models.
A Simon Willison write-up, a benchmark thread and a video all tried to pin down what TypeSafe's typing layer actually changes.
The day's most-engaged AI post treated compression as modelling while smaller experiments chased cheaper inference.
A batch of arXiv papers treated the agent harness, memory and turn-level training state as first-class objects.
Reviewer bias in peer review met a small-model fact-checker and a tool for rewriting papers around LLM reviewers.
Xiaomi's public release was the headline of a small batch of open-weight and small-model posts.
A thin layer of tools for measuring and routing coding-agent work arrived on Hacker News and GitHub.
Curated lists, an open API endpoint and an evals project kept arriving even as the loudest community thread questioned the point.
Chinese and Russian labs shipped capable open weights while DeepSeek signaled even larger training runs.
Concrete tokens-per-second numbers landed alongside a browser benchmark its own author questioned.
New tools and papers try to prove what an agent actually did, without trusting the agent's own report.
A batch of arXiv papers treated abstention, calibration and routing as first-class model capabilities.
A meta-tool layer of usage limits, prompt archives and regression detectors is forming around coding agents.
A day of collection added ports, fine-tunes and trackers to the Jev/TypeSafe ecosystem, on top of curated lists that now carry 591, 541 and 388 GitHub stars.
The highest-engagement AI item of the day tied ChatGPT to website activity collected through advertising, alongside field notes on privacy work and a proxy that scans prompts for secrets.
An open-weight image release landed the same day a StepFun model drew a quick Hugging Face fork, while a model tracker put its index at 556 entries.
Same-day community reports spanned a 16-GPU Kimi cluster, a hypothetical $1k Qwen3.8 chip at 7,000 TPS, a three-week single-3090 run and memory-bandwidth overclocking.
Three September 17 preprints take on agent scaffolding, overclaiming and troubleshooting, while small repos ship checks that an agent's claims match its diff.
Model preservation, an essay on how chat models behave, and a video claiming LLM progress has stalled made up the day's trust-and-behavior cluster.