Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

AI Trending Briefing · October 12, 2026 · afternoon edition

The bill for agent-written code is landing on reviewers, with a 718-firm study showing more code and longer reviews but no measurable gain in finished work, and QGIS closing its main repo to agent-assisted code because review could not keep up.

Artificial Intelligence in the Firm (Chen and Stratton)QGIS QEP-428Muse WonderAndrej Karpathywheresryan22/anatomyagent-harnessclaude-codeai-skillsfrontier-modelsagent-security

Trending AI Briefing: Monday, October 12, 2026 (afternoon ET)

A quiet vendor Monday, and a loud argument about who pays for agent output. The sharpest piece of evidence is a Harvard paper on 718 software firms: agents pushed up lines of code, commits and pull requests, while resolved issues showed no measurable gain and review time rose by about half. The same pattern shows up in a GIS project that closed its main repository to agent-assisted code, and in a Karpathy post about a widening gap in how people understand what LLMs can do. Writing code got cheap. Reading it did not.

What's hottest in AI news right now

A Harvard job-market paper on 718 firms found that coding agents raise output without raising finished work, and Ars Technica's write-up of it was posted to Hacker News on October 12. Fiona Chen and James Stratton's "Artificial Intelligence in the Firm: Bottlenecks in Software Production" (current version dated August 4, 2026) uses Jellyfish engineering data on 725,938 workers from January 2021 to March 2026. After a firm adopts agents, lines of code rise about 30%, commits about 20% and pull requests about 23% per worker-month. Resolved Jira issues show no significant change, and the authors' interval rules out a gain above 12%. Review time climbs 3.45 days on a 7.03-day baseline (about 49%), and the share of pull requests with changes requested nearly doubles. Short-run employment barely moves: the paper rules out an overall decline above 2.9%, though its engineering-headcount bound is much looser (13.8%). The catch: adoption is measured at the firm level, which the authors say likely understates what individual heavy users get, and individually licensed tools are not captured. Ars Technica's page was blocked from this workspace, so figures here come from the paper itself. Paper (PDF) · Slashdot summary · HN

QGIS merged QEP-428, "No ai contributions permitted," on October 7, and the pull request was posted to Hacker News on October 12. The new policy is one line long: no LLM, AI or agent-assisted code contributions in the main qgis/qgis repository. Contributors later found to have used one face reverted code and sanctions on their contributor rights. Maintainer Nyall Dawson describes it as correcting an earlier, softer policy that did not work in practice, and Even Rouault calls it a stop gap. Contributor uclaros notes that about 50 pull requests were tagged "AI Assisted" and roughly 40 of them closed without merging, plus others closed untagged. The honest catch: the wording also appears to ban LLM use for review and security auditing, which other maintainers objected to in the PR discussion, and it covers only the main code repository. PR #387 · HN

Meta announced Muse Wonder and Muse Wonder Full on October 12. Meta calls them its most capable models for personal and professional agent work, built around two features: Active Memory, which carries context about a user across conversations, and Active Focus, which keeps working toward a goal, routes around blockers and asks for clarification before important actions. Muse Wonder is available through the Meta Model API and Muse Code; Wonder Full is available to Muse Code subscribers, and Meta says Wonder will power its Muse app soon. The announcement names Harvey's Legal Agent Benchmark, FocusBench and SWE-Marathon but prints no scores on the page, and the full safety report "will be available soon." Treat the state-of-the-art claim as unverified until that report lands. Meta AI Research · HN

A Karpathy post titled "The gap in shared understanding of LLM capability is widening" was posted to Hacker News on October 11 and became one of the weekend's AI argument threads. The X post itself was not fetched this run. Commenters quote it as describing a small share of people who now see large projects finished far faster, and as saying "human review and comprehension are starting to fall behind." If that wording holds, it is the Chen and Stratton paper in one sentence. Read the original before quoting it. HN thread · X post

"Why aren't AI companies being charged for hacking?", published October 12 on Musing Mortoray, argues that agent break-ins are a foreseeable result of design choices, not accidents. The author's case is that training data, tool access and configuration are deliberate, that foreseeability can establish intent, and that "I believe these companies are committing crimes." He links news coverage of incidents involving Anthropic, OpenAI, Google and Meta systems; this briefing did not re-verify those incidents, and the legal argument is his opinion, not settled law. It matters for builders because it is the first widely read version of the question your legal team will eventually ask about your own agents. Essay · HN

New tools and features worth actually trying

anatomy. A Claude Code skill (it also runs in OpenCode and Codex) that explains a technical idea by inventing a working isometric machine for it, output as standalone HTML, SVG or React. Install with npx skills add wheresryan22/anatomy -g -a claude-code, then try /anatomy explain how a token bucket rate limiter works, interactive. Honest tradeoff: the README says "It is slow on purpose," a full 3D orbit audit takes 15 to 20 minutes for 30 parts, and the 3D mode is beta.

Insomnia. A macOS menu bar app that keeps a MacBook working with the lid closed for a timed session, so a long agent run survives your commute, and it journals every change so a crash can undo them. It can even type continue into tmux panes you mark after a network drop. Honest tradeoff: macOS 26 on Apple Silicon only, it calls itself experimental, and the installer adds a sudoers rule that lets any program running as you call four pmset commands without a password.

DigUp. A free Mac app that searches your photos, PDFs, documents and recordings by meaning, on device, using Google's EmbeddingGemma 2. v0.6.1 shipped October 12. Honest tradeoff: Apple Silicon and macOS 14 or later, an 865 MB model download, and the language evaluations are the author's own and cover only English, Bengali and Arabic.

ppt-master v6.7.0. A skill for Claude Code, Cursor or Codex that turns a source document into a natively editable PowerPoint file instead of slide images; v6.7.0 shipped October 8. Honest tradeoff: you pay for your own model usage, many README links carry affiliate parameters, and the optional PDF converter depends on AGPL-licensed PyMuPDF, so check before you redistribute a bundle.

Trending AI repos on GitHub today

Trendshift read at about 15:08 ET on October 12; Trendshift is a live momentum board and its ranks are momentum scores, not star totals. Stars below come from cache-busted shields.io reads and licenses from each repo's LICENSE file.

  • wheresryan22/anatomy (#2): a Claude skill that builds interactive isometric figures with many small parts to explain one idea. Why now: v1.1.0 shipped October 12. MIT, about 5.5k to 5.7k stars depending on the counter, Node.js 20.10+ with nothing to npm install (the skills installer itself wants Node 22.20+). Caveat: 3D mode is beta, and its frame-rate figures are self-measured.
  • storytold/pdfcraft (#8): an alpha PDF workbench in Rust for desktop and browser, from the same ArtCraft team as this morning's Word and Excel clones. Why now: v0.7.1 shipped October 12. MIT or Apache-2.0, about 11k stars. Caveat: its README claims "0 crashes" on the pdf.js corpus and also says fuzzing "still turns up crashes and hangs on hostile files."
  • hugohe3/ppt-master (#11): the editable-PowerPoint skill above. Why now: v6.7.0 on October 8. MIT, about 61k stars. Caveat: affiliate-tagged links throughout the README.
  • ARahim3/DigUp (#15): the on-device Mac file search above. Why now: v0.6.1 on October 12. MIT, about 1.8k stars. Caveat: "nothing you index leaves your Mac," but it goes online to fetch the model and check for updates daily.
  • FeiZhuLulu/DeepSeek-Bot (#19): a plugin for DeepSeek Harness (DSH) that lets you build bots and run group chats of several bots. Why now: v0.1.2 on October 11. Apache-2.0 (template with no copyright holder filled in), about 1.3k stars. Caveat: a v0.1 preview on a DSH alpha, not an official DeepSeek product, and bots can run commands.
  • krishhgg/Insomnia (#25): the lid-closed session app above. Why now: v0.1.0 stable on October 10, nightlies daily since. MIT, about 1.1k stars. Caveat: the passwordless pmset sudoers rule.
  • shaongitbd/doory (HN, not on Trendshift): a self-hosted Canva alternative. About 410 stars, no releases. Caveat: the license is the Attribution Assurance License, not MIT, and it requires the running program to display the author's attribution at every launch.

What actually matters from today's signal

The trend to track is review capacity. The Chen and Stratton numbers give it a shape builders can plan around: nearly a quarter more pull requests per person, about half again as long to get each one merged, and no measurable gain in finished work at the end. If you run a team, the places to look this week are pull request count per reviewer (the paper finds pull requests got more numerous, not bigger), the changes-requested rate as an early warning, reviewer load (the share of engineers doing review rises about 14% in the paper), and whether your AI reviewer is catching things or just adding comments; the paper counts AI as writing 23.3% of review comments by March 2026.

The counter-signal is what QGIS chose. A ban is the bluntest possible fix, it throws out autocomplete and translation along with agent-written patches, and the maintainers already plan to reopen it case by case. The risk most teams are missing runs the other way: Muse Wonder and every launch like it sells longer autonomous runs, and none of those launches ships a better way for a human to read what came back. Until someone builds that, more agent output means a longer queue in front of the same people.

Source access notes: Vendor scan at about 15:07 ET. OpenAI (latest October 8), Anthropic (October 8), Cloudflare (October 9), Hugging Face (October 9), LangChain (October 9), Mistral (October 6), Microsoft Foundry (October 7), x.ai (September 28) and the GitHub changelog (October 10) had nothing new since the morning run. Google's AI blog listing showed no dates. Claude Code's npm latest is 2.1.296, published October 9 (packument time map), so nothing new there either. Ars Technica returned SITE_BLOCKED; the study figures come from the paper PDF. After the articles' Codex fact checks, the anatomy star figure was widened to the 5.5k to 5.7k range GitHub and shields showed, the skills-installer Node requirement was added, "stayed flat" was softened to "no measurable gain" (the paper's issue estimates are positive but not significant), and a "pull request size" recommendation was corrected because the paper finds PR size unchanged. HN point and comment counts disagreed between Algolia's search and item endpoints, so this briefing prints none. The Karpathy post on X was not fetched; its wording here is as quoted by HN commenters. z.ai's GLM-Legal page rendered empty and was left out. The Hugging Face papers page served a stale listing and was not used. Codex changelog skipped (JS-rendered). Product Hunt not checked this run. Repo facts come from a Sonnet 5.5 verification subagent (cache-busted shields, raw README and LICENSE, releases.atom). A Sonnet 5.5 adversarial pass then caught an unsupported HN comment count on the Karpathy thread and an unsupported "front page" claim for QGIS (both removed), an overstated "precise zero" employment line (reworded with the paper's bounds), unconfirmed exact policy wording for QEP-428 (now paraphrased) and the uclaros attribution; it confirmed the paper's figures, star counts and licenses. Its releases.atom reads were robots-blocked, so release tags rest on the first subagent's reads.