News

AI and LLM news · Oct 29 to Nov 4, 2026

AI news this week, Oct 29 to Nov 4

Open weights kept coming, a viral claim fell apart in 48 hours, and the numbers on AI agents in the real world turned out to be very small.

28%

of Hacker News stories with 100+ points this week were about AI: 85 of 304. One week is noisy, and the last days may still gain points, but October’s average was 21.8%.

Share of HN points
31.6%
Share of HN comments
32.4%
arXiv papers
2,560cs.AI, cs.LG, cs.CL
New on OpenRouter
4models

The week in five lines

  1. A post saying GPT-6 Astra can read and edit model weights became the week’s top story, and the rebuttal arrived a day later.
  2. Kimi K3.1 Pro and Mistral Large 4 released their weights within a week of each other.
  3. Uber said only four rides were booked by AI agents in its busiest week of October, out of about 320 million trips and orders.
  4. Google told schools that they cannot keep Search if they refuse its AI tools, then faced backlash from senators to the mayor of New York.
  5. Brazilian courts fined more lawyers for hidden prompts in filings, and the story reached Hacker News.

01 · The stories

What mattered

Can a model read weights? 48 hours from viral to footnotes

The week’s top story, with nearly 3,000 points, was a post by four MIT CSAIL authors claiming that GPT-6 Astra can “read and write” neural network weights, identifying the game a small network was trained on and editing 0.016% of a Qwen model’s weights to raise its math score from 7.9% to 48.1%. Matthew Honnibal answered that the post mixes two settings: raw weights appear only on tiny networks, and the Qwen result used Python and test runs, the way a human ML engineer works. Simon Willison reproduced part of it on a laptop with Qwen 3.8, then said the no-tools question and the code-assisted results “deserve to be treated separately.” The post’s own authors called the work very early. That is a fair label for the whole debate.

Open weights, twice in a week

Mistral released the weights of Mistral Large 4, which it calls the best open-weight model for coding, while warning that it sees “a potential for uplift to conduct real attacks”. Days later, Moonshot published Kimi K3.1 Pro: 2.8 trillion parameters, 104 billion active, under its own license. In the US, Ricursive released Ripple-65B Preview and says it cost US$ 32 million in total R&D, one-tenth of the previous open-weight leader on agentic coding. These figures are self-reported.

A fake Claude Code extension sat on Cursor’s marketplace for 11 hours

Two malicious builds, versions 2.1.312 and 2.1.313, were on Cursor’s marketplace from about 2:27am to 1:35pm PT on October 29. Cursor says they were not developed by Anthropic. They collected host details and tried to add an SSH key labeled svc-telemetry and open a reverse tunnel. Cursor shut the marketplace down, rotated keys and rebuilt it. The notice has commands to check your machine and steps to clean it.

Postman installed agent plugins without asking

A GitHub issue with more than 1,500 points on Hacker News reported that the Postman desktop app installed Codex and Claude Code plugins on its own, and tried to approve its own startup hook. The reporter did not claim data exfiltration, and had not reproduced it on a clean Mac. Postman said it is reverting the change and redesigning it with “the right level of control.” Agents have created a new way for one tool to change how another tool behaves, and users are noticing.

AI makes old bugs dangerous faster

IEEE Spectrum traced the PHP2Shell flaw: an out-of-memory bug reported as far back as 2024, which a researcher aided by AI tools turned into arbitrary code execution within two weeks. Four engineers then spent two months on the patch. Cryptographer Matthew Green expects a serious incident “maybe sometime in early 2027.” Finding is getting cheap and fixing is not, the same imbalance PostgreSQL’s maintainers described in October.

AI agents booked four Uber rides

On its earnings call, Uber said that AI agents ordered only four rides in its busiest week of October, out of about 320 million trips and orders, plus 500 Uber Eats orders. DoorDash reported “a few hundred.” The CEO called it “very, very early.” Hacker News spent the week arguing about whether models can read weights, and people are still opening the app. Useful to remember when planning an agent product.

The money: bigger numbers, more questions

Nikkei reports that US tech giants’ off-balance-sheet AI obligations almost doubled in six months, to about US$ 3.2 trillion. Anthropic’s amended filing put its own committed spend at about US$ 895 billion and third-quarter revenue at US$ 17 billion. In Congress, 13 Democrats proposed a resolution against bailing out insolvent AI companies. “A.I. companies that fail should go bankrupt,” said Rep. Greg Casar.

Data centers on the ballot, and Google against the schools

In Texas, the fastest-growing US data center market, Democrats won the governor’s race and a Senate seat on November 3, with data centers a central issue, according to MIT Technology Review; The Verge’s headline was “Data centers lost the midterms.” AI-funded PACs had a mixed night, with several of their candidates losing. The same day, Google said schools on its K-12 software cannot use Search unless they opt in to AI tools. Bernie Sanders called it “extortion”. Google replied that AI products are off by default and can be enabled one by one.

02 · In brief

Also this week

  • Meta bought Conductor, a startup that runs multiple coding agents in parallel. Its cloud service shuts down on November 20.
  • ChatGPT has 1.2 billion weekly users, according to its product lead, via Tecnoblog.
  • Business Insider’s headline, which reached 352 points: Anthropic’s internal AI use, at retail prices, now exceeds its employee pay.

03 · The list

New models, top stories, top papers

Models added to the OpenRouter catalog in the period

ModelInputOutputContextListed
ricursiveai/ripple-65b-preview0.200.601MNov 4
tencent/hy40.4171.2511MNov 4
stealth/crouton-alpha0.000.001MNov 3
z-ai/glm-5.3-flashx-legal0.371.251MNov 2

Prices in US$ per million tokens, as listed at collection time. A US$ 0 price usually means a free preview.

Models trending on Hugging Face on November 5, 2026

ModelLikesDownloadsType
moonshotai/Kimi-K3.1-Pro77111,658image-text-to-text
speridlabs/idol67717,180image-to-3d
mistralai/Mistral-Large-4-1T-A52B1,6039,592image-text-to-text
Pika-Labs/Pika-H3-Flex450863image-to-video
Lightricks/LTX-2.510,6652,561,845image-to-video
nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B3,04157,297automatic-speech-recognition

Most-starred new repositories on GitHub

RepositoryDescriptionStars
wangzan101/MikaAn on-device AI agent for Android that operates apps and the web to get things done for you.6,971
KKKKhazix/agent-autoloop让你的命令行 Agent 在后台自己接活、自己跟进到底;平时不烦你,遇到拿不准的事才发卡片找你拍板。3,070
N1nEmAn/sapper-v2Autonomous pentest legion — AI commander + worker swarm + layered fact-intent graph. 3rd Tencent Intelligent A1,202
Balanov/uncutThe lossless, headless coding agent harness. One 7 MB static binary, four tools, plain stdin/stdout, no silent604
SAOS-3624/dsh-see-now让 AI 实时看见你的屏幕:按变化量筛选帧、多帧拼进一张图,静止画面几乎不花 token372
aa1143/NamiAgent零基础入门 AI Agent 的基础教程。不只是教程,也是一个能在本地跑起来的 Agent:读写文件、执行命令、MCP、长期记忆,自带网页端看清每一步。从 100 行最小循环开始,一课加一个能力,最后拆解 Claude 366

04 · Method

How this was counted

Hacker News stories with 100+ points, created from October 29 to November 4 in UTC, read on November 5, so the last days may still gain points. A story counts as AI if its title matches the term list, which misses titles that do not say AI, so read the share as a floor. Each claim links to a page that was read, except where the text says it is going by a headline. Written with AI assistance, using Claude, a model made by Anthropic, a company covered here.

Download the week’s data, in JSON. For the long view, read the State of AI report for October.