Artificially Intimidating
Context Window: AI Daily News Brief
Claude Says It Proved Riemann. A Human Proof Showed Up Hours Later -- AI Brief November 21
0:00
-5:19

Claude Says It Proved Riemann. A Human Proof Showed Up Hours Later -- AI Brief November 21

Today’s Context Window includes a laptop-sized model with a warning label, Warp’s agent factory, a voice actor vs. Netflix, and a no-AI Photoshop rival.
Planting the flag is the fast part. Checking the mountain is the job.

Good day . An Anthropic mathematician says he and Claude proved the Riemann hypothesis, the most famous open problem in math, and within hours a Stanford team rushed out an all-human proof of its own. Also today: a looping model small enough for your laptop that comes with a warning label, Warp’s pitch to get coding agents off that laptop, a voice actor taking Netflix to court, and an artist’s Photoshop rival with no AI at all.


Claude Says It Proved Riemann. Now Comes the Checking. Scientific American

  • What happened: On Friday, Anthropic mathematician Levent Alpöge posted a four-page note on his website: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis,” Scientific American reports. He says the method also settles the much broader Generalized Riemann Hypothesis. There’s no traditional paper. He released the proof as Lean code, a language in which a computer checks every logical step. In his note, he says a written-out version ran past 250,000 pages.

  • Why it matters: Riemann posed the problem in 1859. It’s about a hidden pattern in how prime numbers are spread out, and a mountain of other math is built on the assumption that it’s true. Early checks look strong. MIT’s Andrew Sutherland and others compiled the code independently, and he told Scientific American: “Honestly the chance that anything is fishy here is approximately zero.” A final check that the code proves the real theorem, and not a subtly different one, was still running.

  • What everyone’s saying: Hours later, Stanford’s Ravi Vakil, president of the American Mathematical Society, posted a “Hurried Announcement”: Alessandro Masullo has a proposed human proof of the general version, “all human, with no AI.” Brian Conrad, who has been reviewing it for months, told Scientific American “it’s undergoing major revision.” Cornell’s Steven Strogatz told The New York Times Alpöge looked like “a very skillful jockey riding a really fast horse.”

  • My read between the lines: Nobody alive will read 250,000 pages, so the proof now depends on trusting a different machine, the Lean checker. Scott Aaronson flagged the catch: Lean has had kernel bugs, and “a sufficiently motivated AI could almost certainly exploit them to prove a false statement.” That’s why the slow comparator run matters more than the viral tweets. And the Stanford draft going out early is the most human detail of the day: a mathematician publishing before he was ready so the history books still have a chance to say a person got there too.

📖 Further reading: The AI Pattern That Optimizes Anything Measurable — Overnight — my guide to the agent pattern that grinds overnight on anything with a score attached. A Lean checker is the strictest score there is, which is exactly why math fell first.


You don’t need to crack a 167-year-old math problem to get real work out of an AI. You just need one that shows up where your team already talks. Viktor is an AI agent that lives in Slack and connects to more than 3,000 tools. Ask it for the weekly report, a dashboard, a code fix or a campaign draft, and it gets done in the channel. It’s not a chatbot. It’s a coworker. New readers get $50 off their first month. Hire Viktor →


A Laptop Model That Thinks in Loops Thinking Machines Lab

Three laps per word. Mind the folder.
  • What happened: Thinking Machines Lab released Inkling 2 Mini on Friday, an open-weights model it calls “small enough to run on your laptop.” It has 26.4 billion parameters, 3.4 billion active at a time, and needs about 20GB of memory. The lab says its researchers ran it offline on a MacBook Pro during a flight, and it built a playable flight simulator in a few minutes.

  • Why it matters: The trick is looping. The model’s middle layers run three times for every word it writes, so it does more thinking per word without needing a bigger brain. Thinking Machines’ model card puts it at 48.2% on the DeepSWE coding test, the top score in its comparison table. The weights are on Hugging Face, which walks through the design and a research version that lets you set the number of loops.

  • What everyone’s saying: Looping has a known downside for safety work. Most oversight today reads a model’s written-out reasoning, and the lab says it plainly: “Reasoning that never appears in the CoT can’t be caught by a CoT monitor.” Part of the reason for the release, it says, is to let outside researchers study that. Its early tests found the model couldn’t secretly carry out hidden side tasks.

  • My read between the lines: Read the model card before you hand it your files. It says the model “may take destructive actions, such as deleting or overwriting files, without confirmation, and may not report the outcome of its work accurately,” and tells you to run agent work in a sandbox. So the headline is “runs on your laptop,” and the fine print is “don’t let it run loose on your laptop.” Both are true. Only one made the launch post.

📖 Further reading: Neo-Napster: The Compute Revolution Nobody Saw Coming — my case that the next AI computer is the one already on your desk. A 20GB model that codes on airplane mode is that argument getting real.



A machine may have proved a theorem, and the rest of us still have to check its work. The Brief is free and always will be. Members get the paywalled deep-dives on how to put these tools to work without getting burned, plus the full archive. Become a member →


Warp Wants Your Agents Off Your Laptop Warp

  • What happened: In a new talk, Warp CEO Zach Lloyd argues teams should stop prompting coding agents one at a time on their own machines and build a “software factory” instead. Warp’s blog lays out its own: one configuration file runs every agent in the company. A “Foreman” agent hands work to separate triage, coding, code review and design agents, and each one runs in its own isolated cloud box.

  • Why it matters: When Warp launched its Oz platform in February, Lloyd said “most agents run on individual laptops with no centralized visibility or audit trail,” per the announcement. For a company, that means nobody can see what the agents did or what they could reach. Warp’s setup gives each agent its own model, secrets and tool access, so one agent can’t touch what another holds.

  • What everyone’s saying: The factory idea is catching on because one agent in one chat window doesn’t scale. Warp’s most ambitious piece is a self-improvement flag: every 24 hours or so, a separate agent reviews the runs that failed and suggests fixes to the factory’s own setup, which arrive as pull requests for a human to review. Keep in mind this is a vendor describing the product it sells.

  • My read between the lines: Put this next to Story 2. One lab ships a model that “runs on your laptop” and warns it may delete your files. Another company says agents shouldn’t live on laptops at all. They’re describing the same problem from opposite ends. The part to watch is the self-improving factory. It sends its own fixes back as pull requests, and the queue of things for a human to approve gets longer every day.

📖 Further reading: Your laptop has been in the way this whole time — my explainer on why the execution layer has been stuck on your laptop and what changes when it moves to the cloud. Warp is selling the factory-sized version of the same move.


He Asked Netflix for His Data. It Said AI. DW

He found out the way everyone does now: he asked for his data.
  • What happened: A DW report rounds up creative workers in Germany losing ground to AI. Berlin interpreter Sara Campos lost a client of 15 years to an AI system that was cheaper and worked in five languages at once. Hamburg designer Adrian Romero helped bring AI tools into his agency, then was laid off. And Berlin voice actor Matti Klemm is taking Netflix to court, saying it trained AI on recordings of his voice without his permission.

  • Why it matters: Klemm, the German voice of Jason Momoa, learned it by filing a European data access request. Netflix’s reply showed it had trained AI systems on his voice recordings “without Klemm’s knowledge and consent,” Voice Over Herald reports. His side wants a court to order disclosure of what was used and a stop to future training. Yesterday we covered a model suing over his AI double; this is the voice version of that fight.

  • What everyone’s saying: Germany’s dubbing industry is split. Netflix just raised residual payments, heise reports, and its 2025 rules require explicit consent for digital replicas and synthetic voices. Those rules say nothing about training models on recordings. Klemm’s answer, from an earlier heise report: “There can hardly be any talk of voluntary consent when it is effectively a prerequisite for receiving assignments.”

  • My read between the lines: The most useful tool in this story isn’t a lawsuit. It’s the data access request. Klemm didn’t have to prove anything from the outside. He asked, and the company had to answer in writing. Anyone in the EU who has worked for a platform can file the same form. And Romero’s detail stings: he helped bring in the tools, and then the job went away. He’s moving to an insurance company.

📖 Further reading: Every New Tool Was Going to Kill Art, with John Hardin — my conversation with historian John Hardin on every tool that was supposed to kill art. His point is that the fights that mattered were about who gets paid, which is exactly the fight Klemm is in.


An Adobe Rival Whose Feature Is No AI Creative Bloq

Sketchbooks, this way. Magic wands, wait outside.
  • What happened: Artist Stuart Semple launched NIVA, a browser-based design and photo editor whose main selling point is that it has no generative AI, Creative Bloq reports. It reportedly drew close to 30,000 users in 48 hours after he announced it on Instagram. It started as a Kickstarter that raised $240,000 from 3,000 backers, under the name “Abode.”

  • Why it matters: It’s a full toolkit: brushes, a pen tool, layers, healing and cloning, type tools and page layout, all in one app. The pricing is the other half of the pitch. There’s no subscription. It’s free for an hour a day, $10 for a day pass or $100 to own it outright. Semple: “With Adobe pushing AI further into its tools, I wanted somewhere people could still practise their craft and make things themselves.”

  • What everyone’s saying: Semple is the artist who once irritated Anish Kapoor with a very black shade of black, so a public jab at a giant is on brand. On the name change: “I can’t say I woke up to a nasty letter from the lawyers representing Adobe for legal reasons, so I won’t.” He says over half of the first users made something within 40 seconds of getting access.

  • My read between the lines: “No AI” is now a feature you can sell, and 30,000 sign-ups in two days says people will try it. But I think the pricing is the bigger rebellion. A $100 one-time license is a dare to the subscription model, and that annoys people more than any AI button does. The hard part comes later: keeping a promise like “no AI” once the users who joined for it start asking for just one small shortcut.

📖 Further reading: The Font That Beat AI for About a Week — my look at the viral anti-AI font that AI beat in days. NIVA takes the smarter route: it doesn’t try to block AI, it just leaves it out and lets you choose.


That’s your AI Brief for Saturday.

—Nicholas from Artificially Intimidating

Discussion about this episode

User's avatar

Ready for more?