Ant's inclusionAI open-sources Ling-3.1-flash-1019: 560B parameters, MIT license, 1M context
Models
Ant Group's inclusionAI open-sourced Ling-3.1-flash-1019 on the evening of Oct 27 (SGT) under the MIT license. It has 560B total parameters with 25B active, a 1M-token context, and a base model released alongside. Per the official model card, its AA Intelligence Index v4.3.2 score rose from 41 to 44 and AutomationBench-AA from 62% to 68%. The Artificial Analysis page lists API pricing of $0.30 input and $0.90 output per million tokens. Note that the SGLang example command in the model card still points to the previous model path, so copying it as-is loads the older version.
Why it matters to you
It is another MIT-licensed large model you can use commercially. If you run automations on a budget, compare its API against your current model on the same tasks before switching.
Sources
DeepMind unveils Genie 4, its first world model with native audio and natural-language scene edits
Models
Google DeepMind released Genie 4, a world model that lets you change the weather, move the camera and add characters in real time with natural language, and it is the first to include native audio. The Verge reports each session lasts about 2 minutes, versus 1 minute for Genie 3. It is a research preview for trusted testers only, with Google AI Ultra subscribers and early Gemini API partners next. DeepMind gave no benchmark numbers or pricing and lists known limits such as object permanence and text rendering.
Why it matters to you
You cannot use it yet, but the official demo videos are good material for explaining world models. If you work on agents or robotics, watch its progress as a training environment.
Sources
FrontierCode 2.0 grades coding agents on whether a PR would be merged, and the top score is 45.5%
Agents
Cognition released FrontierCode 2.0, a new coding benchmark. More than 20 open-source maintainers wrote tasks from their own repos, spending over 40 hours per task, and each result is graded on whether it would be merged, with a failed blocking check scored as zero. The leader, Opus 5.5, reaches only 45.5%, GPT-6 Astra gets 42.8%, and Kimi K3 drops from 44.2% on version 1.1 to 17.7%. Note that Cognition both wrote and graded the tasks, its own SWE-2 model is on the board at 23.9%, and no third-party rerun has been verified yet.
Why it matters to you
Code that runs is still far from code you would merge. If you build with AI, write your own review standards as a checklist and go through it before every merge.
Sources
Claude Code adds an even more dangerous switch that no longer stops rm -rf ~
Agents
Starting with v2.1.311, Claude Code supports the environment variable CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1. Set in bypassPermissions mode, it stops prompting even for rm and rmdir on critical paths such as your home directory. The official docs say plainly that with it set, rm -rf ~ can run without asking you. The Register reported the change on Oct 28, and the 2.1.311 changelog does not mention the variable.
Why it matters to you
If you really want an AI to run commands fully on its own, do not enable this on your main computer. Use a disposable container or VM, and back up important files first.
Sources
FDEs need not be engineers: ex-bankers embed for weeks and turn senior judgment into 54 grading criteria
Forward-deployed engineering
An OpenAI customer story profiles financial AI startup Terminal X, whose FDEs are former JPMorgan and Blackstone investment professionals. They embed with clients for weeks, reverse-engineer the investment memos the client trusts most, and encode unwritten processes into hundreds of agents tailored by asset class. An eval agent checks 54 criteria one by one, attaching a reason and correction to every failure, which becomes a new workflow instruction. The page says one PE client saves $2.4M a year and the eval agent agrees with senior human reviewers 88.6% of the time. All figures are self-reported.
Why it matters to you
Domain experts can deliver AI too: first write the expert's judgment down as gradable criteria, then have the AI work and check itself against them. The method works for small client projects as well.
Sources
Bayer Crop Science: with embedded AWS experts, 22 specialist agents now resolve about 80% of issues without engineers
Forward-deployed engineering
An AWS blog post co-authored by two Bayer staff describes how Bayer Crop Science's breeding R&D team worked with embedded AWS generative AI specialists from design to production, holding three technical stand-ups a week and clearing 35 blockers before launch. They shipped 22 specialist agents and about 20 API tools. In the first phase about 80% of user questions needed no engineer, troubleshooting dropped from over an hour to under 10 minutes, and caching stable prompt prefixes cut input token costs by 20% to 30%. The figures are self-reported.
Why it matters to you
The value is in the engineering decisions, not the prompts: monitor whether tasks actually succeed rather than whether the API call returned, and do not assume a newer model is a drop-in replacement. Use these as a checklist when you build agents for clients.
Sources
hallmark: an anti-generic design skill for AI coding tools, adding 2,830 stars in a day
Trending on GitHub
hallmark is an open-source design skill from developer Nutlope. It gives Claude Code, Cursor and Codex design guidelines for 50 themes plus a checker called sloplint, with redesign and audit commands, aimed at AI-generated pages that all look the same. As of the Oct 28 snapshot it had about 39,500 stars, with 2,830 added that day, under the MIT license. It is a plain-text skill, so the risk is low.
Why it matters to you
If you use AI to build landing pages or a portfolio, browse its 50 theme screenshots, pick two or three that fit, then have the AI run an audit on your current page to see what looks most generic.
Sources
Anthropic's knowledge-work-plugins packages Claude plugins by job role
Trending on GitHub
knowledge-work-plugins is an official Anthropic repo that packages Claude plugins for 11 roles such as marketing, sales, finance and support, including skills, commands and connector configs for Claude Cowork and Claude Code. As of the Oct 28 snapshot it had about 33,500 stars, with about 2,800 added this week, under Apache-2.0. It is more structured than scattered community skills but leans on the Claude ecosystem, and some plugins need third-party connector authorization.
Why it matters to you
Solo builders can borrow its role breakdowns, such as how the marketing plugin splits topic planning, writing and review, and apply them to their own workflow. Grant connectors only the minimum permissions.
Sources
Skillry sells tasteful skills to AI coding tool users and reached about $4K verified MRR in two months
AI money case
Skillry is a library of visual agent skills built by indie developer Yihui (@yihui_indie) for Claude Code, Codex and Cursor users, covering web pages, slides, images and video. Each skill shows real output before you install it, and pricing is free plus $9.99 a month. Revenue is Stripe-verified on TrustMRR: $3,988 MRR, $12,445 in the last 30 days and 404 subscriptions, after launching in August. Risks: acquisition leans heavily on the founder's X account with about 64,000 followers, and competitors such as TypeUI already exist.
Why it matters to you
The lesson: people pay even for lightweight skills, as long as they see real results first. Pick a visual task you are good at, build one skill with three real output examples, and test installs with a free tier.
Sources
Do not start Claude Code or Codex in Downloads, and audit unfamiliar AGENTS.md files
Creator playbook
The most-discussed safety play among Claude creators this week: an attacker can drop a malicious AGENTS.md or CLAUDE.md into your Downloads folder, and starting an agent there injects its instructions. The fix has three steps: start agents only inside project folders; have the agent list, read-only, every AGENTS.md, CLAUDE.md and .claude folder under your home directory plus the MCP servers and base_url in its config, without changing anything; delete anything unfamiliar and rotate the related credentials. The success rates quoted in posts come from a report by security firm Forever, which has not been read directly, and there is no CVE or patch yet.
Why it matters to you
It costs almost nothing: have your AI coding tool run the read-only check today. Browser downloads land in Downloads too, so never use it as a working folder.