Independent AI intelligence Two editions daily · ET
Fervor AI

AI Trending Briefing · November 22, 2026 · afternoon edition

Sunday's AI signal turns on who does the measuring, as OpenAI published its own pretraining loss as a proposed safety yardstick, an outside group had to estimate Anthropic's 2028 power mix, and a Ghostscript exploit list that reaches AI image tools still has no public full exploit.

OpenAIPerplexity and IntelligenceUpload2ShellBetter Data Center ProjectMuruworker755frontier-modelsagent-securitymulti-agentagent-infrastructurelocal-ai

Trending AI Briefing: Sunday, November 22, 2026 (afternoon ET)

A quiet Sunday for launches, a loud one for numbers. The stories that moved today all ask the same question: who measured this, and can anyone else check? OpenAI's chief scientist published the company's own pretraining loss figures and proposed them as a shared safety limit. An advocacy group built Anthropic's future power mix from permits and filings because Anthropic does not publish one. A security researcher lists ten applications where a Ghostscript bug ran commands, with AI image tools among the code paths traced, and every full exploit chain is marked "Not published." Self-reported figures are cheap. Checkable ones are still rare.

What's hottest in AI news right now

OpenAI published "Perplexity and Intelligence" on November 22, an essay by chief scientist Jakub Pachocki that puts the company's internal pretraining numbers on the record. Perplexity is how surprised a model is by text it has not seen; lower means a better predictor. On OpenAI's held-out dataset, log-perplexity was about 0.99 for GPT-3.5 and 0.91 for GPT-4, current models sit below 0.8, and a model near 0.76 started producing what he calls "groundbreaking results" on open math problems. He forecasts more transformative capabilities somewhere around 0.68 to 0.7, adds "I don't know what those capabilities will be," and proposes industry "common limits" on perplexity that labs agree not to cross until monitoring catches up. The catch is in his own footnotes: the dataset is a fixed snapshot of OpenAI's internal codebase, labs use different data so cross-lab comparison is hard, and chain-of-thought raises capability without lowering perplexity at all. It is a yardstick only OpenAI can read today. OpenAI, Hacker News

Upload2Shell, Paulos Yibelo's disclosure of CVE-2026-125396, reached Hacker News on November 21, where it drew 148 points. It is a use-after-free in Ghostscript's Type 1 font loader: a crafted document escapes the -dSAFER sandbox and runs commands as whatever process renders it. Ghostscript fixed it in 10.08.0 on September 8, Debian 13 got its first security update in late September and Ubuntu followed on October 21, and the CVE was published November 9; the page is a rolling disclosure last updated November 18. The researcher lists ten applications where commands ran, including Debian 13 CUPS, Roundcube 1.7.2, MediaWiki with PdfHandler and WordPress core's renderer, and says Google mitigated its own exposed path on November 10. For AI builders, two rows matter: Gradio 6.32.0 is marked "renderer reached," a live upload handed to Ghostscript, while InvokeAI, A1111 and ComfyUI are "source-mapped," meaning code paths were traced but not exploited. Every product case study and full command-execution proof is marked "Not published" (the write-up does publish a smaller memory-control proof that runs no shell), and a September teaser naming Box and Zoom is not backed by the page's own index. pwn.ai, Hacker News

The Better Data Center Project's estimate that Anthropic will draw about 87% of its power from gas and 11% from coal in 2028 became Sunday's most argued AI thread, about 700 points on Hacker News. The case study is dated November 16. It is the group's own model, not an Anthropic disclosure: a mid case that assumes 9.6 GW of contracted compute is online in 2028, about a quarter of it behind-the-meter gas at sites like Monarch in West Virginia and Colossus 2, and grid sites modeled with EPA's AVERT tool for marginal emissions. Its mid case lands at about 32 million tonnes of CO2e a year. The authors flag that the result leans on the marginal-emissions assumption, and with average grid intensity the carbon figure drops by roughly a fifth. Anthropic's own infrastructure page says it is "working to broaden our energy mix with geothermal, wind, solar"; the report says no US renewable commitment backs that yet. Better Data Center Project, Hacker News

Muru, Graham Dumpleton's AI-assisted Rust server for Python web apps, got a Simon Willison write-up on November 21. Dumpleton created mod_wsgi 20 years ago. Muru pairs a Rust edge and release controller with Python worker processes so WSGI and ASGI apps can deploy and roll back without downtime; the latest tag is 0.1.0-alpha.34, dated November 21. The interesting part is the paper trail: Willison links roughly 19 posts from November 4 to 22 (17 with Muru in the title) documenting how Dumpleton built it with AI, including one where he writes that "deciding what isn't good enough yet is still down to me." Per the repo, it is alpha software under Apache-2.0, and contributions are taken "as requirements, not as code." Simon Willison, muruhq/muru

worker755, a Show HN posted November 22, flips the multi-agent setup: you play the subagent and an LLM plays your manager. Instructions land in your terminal on a 90-second default clock, and a hidden check tracks whether the manager signs off on broken work. It runs locally on Node 24 with your own OpenRouter or Vercel AI Gateway key, or a local OpenAI-compatible server such as Ollama. It is a toy. It is also the most direct test on the board today of the thing every orchestrator design assumes and few measure: does the boss model notice when the worker lies? gulivan/worker755, Hacker News

New tools and features worth actually trying

worker755. Ten minutes as a subagent teaches more about orchestrator prompts than most blog posts; npm run mock plays offline before you spend a token. Honest tradeoff: the README says your key is "never logged or stored," yet the "Remember on this device" option keeps it in the browser, and real play needs a key or a local model.

Muru. If you run Python services and want zero-downtime releases from one tool, uv tool install muructl gets you the controller. Honest tradeoff: it is an alpha with a single maintainer steering the design, so treat it as something to read and test, not to put in front of paying users yet.

PcapAnalyzer. A local web app that reads .pcap captures, flags 27 kinds of network problems with severity and next steps, and can hand a capture to an AI for investigation (158 points on Hacker News on November 21). Honest tradeoff: the "captures never leave your network" promise holds only for the self-hosted copy, not the hosted demo it also links, and beyond a no-key demo mode the AI step needs your own Anthropic key (the README says it sends packet metadata, not the capture file).

Trending AI repos on GitHub today

Read from Trendshift's daily board at about 15:12 ET; its ranks are momentum scores, not star totals, and star counts below come from cache-busted shields reads.

  • pingdotgg/t3code (#6): a single interface that drives the coding agents already on your machine (Claude Code, Codex and others) from desktop, web or phone. Why now: a nightly build landed today. MIT, about 71k stars, latest is the nightly 0.1.12-nightly.20261122.5098. Caveat: it does not resell tokens, so you need your own signed-in agents, and the T3 Connect remote option has no stated price.
  • firecrawl/firecrawl (#17): turns websites into clean markdown for agents, as a local binary, a self-hosted server or a hosted API. Why now: back on the board with a local "one binary" mode. AGPL-3.0 for the core (SDKs MIT), about 207k stars, release feed not readable this run. Caveat: local mode falls back to Firecrawl Cloud for pages it cannot load unless you set FIRECRAWL_CLOUD=off.
  • pithings/unkernel (#20): the real Linux kernel compiled to WebAssembly, running unmodified arm64 Alpine and Debian binaries in a browser tab or Node, booting to a shell in under a second. Why now: v0.2.0-261121-e69adda shipped November 21. GPL-2.0-or-later, with the JavaScript SDK under MIT, about 1.1k stars. Caveat: the speed figures are the project's own, the syscall numbers were taken on an earlier riscv64 user space, and there is no seccomp or ptrace.
  • FlorianBruniaux/claude-code-ultimate-guide (#21): a large Claude Code reference covering hooks, skills, MCP servers and templates. Why now: v3.46.0 today. CC BY-SA 4.0 (templates CC0), about 6.9k stars. Caveat: it promotes the author's consulting and gates a PDF behind an email sign-up.
  • HKUDS/CLI-Anything (#24): generates command-line harnesses so agents can drive existing desktop software such as GIMP and Blender. Why now: steady climb on the board. Apache-2.0, about 63k stars, last release v0.4.0 on June 25. Caveat: the README gives four different test counts and two different app counts, and the agent benchmark is still an unchecked roadmap item.
  • zhouxiaoka/autoclip (#25): cuts long videos into short clips with captions, covers and post copy, with editing done locally. Why now: AutoClip Desktop v1.5.8 is three weeks old and the repo is climbing. MIT, about 13k stars. Caveat: its speed figures are three self-run tests on different footage, and the README carries sponsor referral links.

What actually matters from today's signal

Measurement is turning into the product. OpenAI's essay matters less for its numbers than for the move: a frontier lab offering one internal metric as the basis for cross-lab limits. Builders should track three things this week. Whether any other lab answers with its own loss figures. Whether your image, PDF or document pipeline, including the AI front ends people self-host, still ships a Ghostscript older than 10.08.0. And whether your orchestrator agents check subagent output or just approve it, which worker755 lets you feel from the wrong end of the leash.

The counter-signal is that every headline number today is self-reported or modeled. OpenAI's perplexity comes from a dataset nobody else can see. The Anthropic power mix is an outside estimate built on assumptions Anthropic has not confirmed or disputed. The Ghostscript exploit list has no public full chain. None of that makes them wrong. It makes them unverified, and the industry keeps treating unverified as settled. The cheapest correction is the one the AI tools in this brief already allow: run the check yourself, on your own systems, before you repeat the number.


Source access notes: Read Sunday November 22 from about 15:09 ET. The workspace shell and the Mac shell reached npm and raw GitHub only; vendor pages, Hacker News (Algolia API), Trendshift and repo files were read through web-fetch subagents. OpenAI, Anthropic, Cloudflare, GitHub changelog, Hugging Face, Microsoft Foundry and Agent Framework, LangChain, Mistral and xAI all loaded; their November 20 to 21 items (MedChemBench, Anthropic's IPO close, Cloudflare's WAF posts, Foundry VNet checks, LangChain knowledge factory, Inkling 2 Mini, harness adapters) were covered in the last two briefs and skipped. Anthropic's newsroom shows a stray "November 22, 2026" label with no post attached; treated as a page artifact. Claude Code is unchanged (2.1.330 on latest, 2.1.323 on stable); Codex CLI is still 0.171.0. Google's AI blog rendered no story cards, and DeepMind shows month-level dates only. Hugging Face papers served Friday's November 20 list. Product Hunt search returned only index pages, skipped. Firecrawl's releases.atom was rate limited on five tries. Codex (Part 2 article check) noted the pwn.ai write-up does publish a smaller memory-control proof of concept that runs no shell and that HN item pages do not prove front-page placement; both briefing lines were corrected. Adversarial pass (Sonnet 5.5, 18 tool uses) caught: Muru post count (17 Muru-titled, about 19 in all, running to Nov 22), Debian's first fix in late September not October, Gradio listed as renderer reached rather than source-mapped, "a dozen" exploited targets corrected to ten applications, the HN count updated to about 700, and two Muru facts re-attributed to the repo. Its note on PcapAnalyzer's demo mode was folded in. The OpenAI footnote caveats on cross-lab data and chain-of-thought come from the first full fetch of the essay; the second fetch was truncated and could not confirm them, so they stay as stated by the author.