Independent AI intelligence Two editions daily · ET
Fervor AI

AI Trending Briefing · November 21, 2026 · afternoon edition

A 17-million-line Lean proof that no person can read was certified because independent checkers agreed, and several other items today point the same way, output is cheap and the checker is the product.

LeanMathlibLevent AlpögeGPGMEMedChemBenchDreamCoderfrontier-modelsagent-securityagent-harnesslocal-aiai-skills

Trending AI Briefing: Saturday, November 21, 2026 (afternoon ET)

Saturday is quiet on vendor blogs and loud everywhere else. The day's biggest technical event is a Riemann Hypothesis proof that runs to 17 million lines of Lean, a size no human will ever read, and the Lean and Mathlib communities certified it anyway because several independent proof checkers agreed. Several other items rhyme with it: a benchmark that admits it cannot rule out memorized answers, a study showing 65 models giving the same advice, and an essay arguing that people who stop writing code lose the skill to review it. Generating the work is now the easy part. Checking it is where the trust lives.

What's hottest in AI news right now

The Lean and Mathlib communities published a joint statement on November 21 certifying Levent Alpöge's proof of the Riemann Hypothesis, released the day before as a Lean library of about 17 million lines. The statement, endorsed by the Lean FRO and the Mathlib Initiative, says the proved theorem matches Mathlib's own RiemannHypothesis statement, that "dozens of people" reproduced a clean build, and that the exported proof passed Lean's kernel plus the independent checkers con-leche and con-ron, which come with a formal proof that they cannot accept a false proof. The statement does not mention AI. That claim comes secondhand from Scott Aaronson, who wrote on November 20 that "Levent Alpöge says that he's gotten Anthropic models to prove the Generalized Riemann Hypothesis" (the stronger statement, which Alpöge's paper claims; the Lean library the communities certified proves the classical Riemann Hypothesis). Aaronson names no model, hedges with "apparently" in his title and adds "my bet is that the proof will stand." The honest catch: what got certified is formal correctness, not understanding, and Aaronson expects mathematicians to need years to explain the ideas. Lean community statement · Aaronson

GnuPG disclosed CVE-2026-129975 on November 20, a double free in GPGME triggered by an OpenPGP key with a crafted, very long user ID, fixed in GPGME 2.3.1 (dated November 19). Werner Koch's advisory calls it critical, says an attacker can use it to exfiltrate data from applications by sending them, for example, signed data, and says remote code execution "can't be ruled out." It affects software built on any GPGME before 2.3.0 (a version that was never publicly released), including Gpg4win, most mail clients, many XMPP clients, Samba and Flatpak; Gpg4win 5.1.2 ships the fix. The advisory ties the early disclosure, originally planned for November 25, to the ruby2shell bug. The catch for builders: statically linked copies stay vulnerable until each app is rebuilt, and any agent or pipeline that imports keys from strangers is exactly the input path described. GnuPG advisory

OpenAI published MedChemBench on November 20, a benchmark of 99 medicinal chemistry problems drawn from 117 papers and patents, covering 761 molecular structures. At the xhigh reasoning setting, GPT-6.2 Astra scored 84.5% against 58.2% for GPT-5.6 Sol, and OpenAI notes that its Astra model earned 95.4% of the credit on structure components but only about half on compound selection and numerical prediction (the page does not say which Astra version that breakdown covers). OpenAI's own limits section says "prior exposure to the published sources cannot be ruled out," and every response was a single attempt with no tools. OpenAI

Cody Collier's A90 experiment "AI is giving everyone the same advice," published November 20, asked 65 models from 30 companies the same open questions 127,918 times through OpenRouter. 98% of answers recommended an index fund for $10,000, 69% of 3,151 invented people were exactly 34, and "Elena Vance," the 34-year-old marine biologist Gemini keeps inventing, shows up in 684 new GitHub READMEs this year. Asking for ten options and picking one cut the chance a stranger gets your exact answer from 34% to 7%. Collier warns that 50 asks per question is enough to see a family resemblance, not to rank small differences. A90

Christian Ekrem's essay "If you stop writing code, you also get worse at reviewing it" (November 20) reached about 366 points on Hacker News. His argument is analogy rather than data: skills decay without practice in every other field, so engineers who only review Claude's output should expect their review to decay too, and he asks readers whether they are "really reading half of it." It has no numbers, and it lands hard anyway on a day when review is the whole story. cekrem.github.io · HN

New tools and features worth actually trying

DreamCoder. Mickey Petersen, author of Mastering Emacs, announced it on November 20 after 15 months of work: a browser-based coding harness and IDE with sub-agents, MCP servers, skills, hooks and a tiling layout, with downloads for macOS, Windows, Linux and Docker, and config import from Pi, OpenCode and Codex. He pitches it at open-weight models and singles out Qwen 3.8 27B. Honest tradeoff: the announcement states no price or license, full Emacs key bindings are "parked," and it runs its own tool loop, so your existing harness habits only partly carry over. Mastering Emacs

OpenComputer. A Show HN on November 20 pitched diggerhq's Apache-2.0 project as an open-source alternative to Devin, with a coding agent you talk to in Slack that works on its own sandboxed machines, in your cloud or theirs. Honest tradeoff: its speed comparisons against Devin are the vendor's own benchmarks, self-hosting leans on Cloudflare Workers and Durable Objects, and you still bring an agent backend such as Claude Code, Codex or OpenCode. GitHub

LookAtMyQuant. An MIT quantization compiler that measures how much each tensor suffers when compressed, then picks a format per tensor to fit a size budget; its author reports 12 times lower KL divergence than Q4_K_M at the same file size on Gemma 4 E2B. Its README also reports shrinking Thinking Machines' new Inkling 2 Mini from 18.89 GB to 5.38 GB (1.63 bits per weight), which is what sent it up Hacker News today. Honest tradeoff: every number is self-reported, and compiling a plan needs datacenter GPUs (the README cites about $34 of Modal time for its second recipe). GitHub

ClawDOS. A GPL-3.0 coding agent that runs on an IBM PC XT with an 8088 at 4.77 MHz, talking to the Claude Messages API through a TLS tunnel on another machine, with a safe mode that asks before changes. Its author says Claude wrote all of the code. Honest tradeoff: it is a delightful stunt rather than a tool, and the tunnel box holds your API key. ClawDOS

Trending AI repos on GitHub today

Read from Trendshift at about 15:15 ET; its figures are momentum scores and printed fused with neighboring digits today, so ranks only.

  • XinaTian/MROffice (#3): browser-based document, spreadsheet, slide, PDF, handwriting and code editors. Why now: an AGPL office suite with AI hooks is climbing fast. AGPL-3.0 plus a NOTICE requiring MR Office attribution, about 1.4k stars (shields), release MR Office 0.1.5 this week; caveat: Word, Excel and PowerPoint support, charts and PDF save need a closed-source desktop component, and embedding needs the hosted commercial service.
  • elementalsouls/Claude-BugHunter (#5): a bundle of Agent Skills and slash commands that teach Claude Code web bug hunting and red-team methods. Why now: security skills are the fastest-growing skill category on the board. MIT code with CC BY 4.0 content, star count unconfirmed (the badge and Trendshift disagree), release v2.1 (feed dated October 16); caveat: the README contradicts itself on skill counts and carries two sponsors.
  • rohanarun/alpoge-rh-replay (#7): a volunteer effort to replay the Lean build of Alpöge's proof module by module and record evidence. Why now: it is the crowd-sourced version of the checking described above. MIT, about 70 stars, no releases; caveat: its README counts 93,051 of 175,187 modules with some recorded evidence, far fewer with a full kernel replay, and no complete verification yet, and it needs a large Linux build machine.
  • atuizz/posewear (#8): a static gallery of pose, outfit and portrait references with bilingual prompts and an in-browser AI try-on using your own key. MIT for code, about 1.3k stars, v1.0.0 (October 30); caveat: image licensing is unclear, and the Doubao route goes through the site's proxy by default despite the README saying keys go straight to the provider.
  • kunkundi/crossdesk (#15): a lightweight cross-platform remote desktop with a web client. LGPL-3.0, star count not read this run, release v1.7.6-20261112 (November 12); caveat: the README calls it experimental, and self-hosting means installing a self-signed certificate on every client.
  • bookwjitchieni1986/blender-ai-mcp-addon (#18): a Blender add-on with an AI chat assistant and a built-in MCP server. Listed as a warning, not a pick: MIT license file carrying another developer's copyright (the author of the original blender-mcp), about 2.2k stars two days after its v1.0.0 (November 19), a keyword-stuffed README and a ZIP name that differs between README and release. Those are lure markers; nobody here inspected the ZIP, so do not install it without doing that first.

What actually matters from today's signal

The trend to track is verification as a separate, independent layer. The Lean result works because the thing that wrote the proof and the things that checked it share nothing: several kernels, written in different languages, two of them formally proved sound, all reading an inert export rather than the original source. Builders can copy that shape directly. The highest-signal areas this week are deterministic checkers in front of agent output (type checkers, test suites, policy hooks that fail closed), replay and reproduction tooling like the RH replay repo, benchmarks that publish their contamination limits the way MedChemBench does, and input hygiene for anything an agent parses from strangers, which is what the GPGME bug punishes.

The counter-signal is the human checker. A Lean kernel does not get tired or lose practice; a reviewer does. Ekrem has no data, but the A90 numbers show what happens when nobody checks: a 98% consensus on index funds and one invented marine biologist's name spreading through hundreds of repos. Machine checkers exist for proofs and types. For most product code and almost all advice, the checker is still a person, and that person is reading less of each diff every month. The risk is not that agents write bad code. It is that the only reviewer left stops being able to tell.


Source access notes: Read Saturday November 21 from about 15:09 ET. The workspace shell reached npm and raw GitHub only; vendor pages, Hacker News (via the Algolia API) and Trendshift were read through web-fetch subagents. OpenAI, Anthropic, GitHub changelog, Hugging Face, LangChain (now at langchain.com/blog), Cloudflare, Microsoft Foundry, Agent Framework, Mistral and xAI all loaded; nothing was dated November 21, and their November 19 to 20 items (Performance Factory, Inkling 2 Mini, harness adapters, LangSmith Tool Gateway, ruby2shell, Foundry Managed Compute, GitHub advisory permissions) were already covered in this morning's or yesterday's briefs and were skipped. Claude Code is unchanged since this morning (2.1.330 on latest, 2.1.323 on stable); Codex CLI still 0.171.0. Google's AI blog served a stale snapshot and DeepMind shows month-level dates only. Product Hunt was skipped. A Hacker News item claiming new models "flop at Simple Bench" was dropped because the Simple Bench page carried no scores for them. The morning brief skipped the Riemann Hypothesis story because AI involvement was unverified; it is the lead here because Aaronson's post now attributes the work to Anthropic models, as Alpöge's claim. Trendshift's printed figures were fused, so ranks only. GitHub rate-limited some release feeds for the repo-verification subagent; MROffice and crossdesk releases were re-read from the Mac. Trendshift rank #17 (a MetaTrader bot) returned not-found and was dropped. A Sonnet 5.5 adversarial pass then ran on the draft. Applied: restored "Generalized" to Aaronson's quote and separated Alpöge's GRH claim from the RH statement the Lean communities certified; softened the thesis from "accepted" to "certified" and from "the rest of today's signal" to "several other items"; tied MedChemBench's structure-versus-selection split to OpenAI's own wording and flagged its ambiguous Astra naming; corrected the GPGME exfiltration wording, affected range (before 2.3.0) and the ruby2shell timing attribution; fixed the LookAtMyQuant line (its README does report Inkling 2 Mini at 5.38 GB); HN points to about 366; Elena Vance is a marine biologist, not a novelist; narrowed the RH replay progress figure; dropped an OpenComputer CLI command two reads disagreed on. Held: star counts and Trendshift ranks (from the repo-verification pass and the morning-style live read).