Worldian 🌎

AI proposes. Science verifies. Three stories this week show both the promise of AI in science and the work still needed to establish what is true. 🔭 Physics: OpenAI has given a group of scientists a proof related to quantum Yang–Mills theory. The group is checking whether the formal proof matches the mathematical and physical problem it is meant to address. Its assessment is still pending. The famous mass gap problem remains open. 🔢 Mathematics: Axiom says it has published a computer-checked formalization of the classification of finite simple groups, a vast proof developed with mathematicians and AI agents. The work is now available for the mathematical community to examine. 🧬 Biology: Researchers at Generate Biomedicines used AI agents to design a shorter component of a CAR-T cell receptor. Laboratory tests showed promising results. This is early research, not an approved treatment. The common thread is verification. AI may help generate ideas, designs and proofs faster, but scientific knowledge depends on whether they survive careful scrutiny. A new Ai2 study adds a reason for caution: leading AI models can give sharply different answers about whether a research idea is novel when the evaluation setup changes. When would you call a result proposed by AI a scientific discovery? 🖖 SOURCES IN FIRST COMMENT

AI stopped answering questions. It started working shifts. Yesterday a swarm of AI agents took 24 days to check a math proof of about 15,000 pages that more than 100 mathematicians built over decades. Along the way, the agents found a mistake in it. That’s the trend of the week. AI is no longer just a chatbot that answers you. More and more it’s a whole team of agents: one writes, another checks, a third fixes. Humans set the tasks and verify the work. Here’s what these “teams” got done in just the last few days, and what it means for us. 🧮 The largest proof in history, checked by a machine The Classification of Finite Simple Groups is something like the “periodic table of symmetry.” It says that every “building block” of symmetry either belongs to a few known families or is one of 26 “exotic” exceptions. The largest of these is called the Monster, and it has about 8 × 10⁵³ elements. The problem is that the proof is so big that almost nobody has read all of it. For years, mathematicians have simply trusted that it’s right. The startup Axiom set a swarm of dozens of AI agents on it, with about 30 people directing them. In 24 days the agents rewrote the whole proof in Lean, a language where a computer checks every step. The result is over 40 million lines of code, the largest codebase of its kind so far. The most interesting part: the AI found counterexamples to a claim in a published proof from 1986. The theorem itself turned out to be true, but that step was wrong. The AI came up with a new argument to replace it. The team also found 34 significant deviations from the published literature. ⚠️ For now, all of this is according to Axiom itself. The new argument hasn’t been peer-reviewed yet. And a human still has to check that the formal statement in Lean really says the same thing as the original theorem. ⚛️ Quantum physics Physical Superintelligence is a lab where AI does physics. It published results on five open problems in quantum information, each with a paper and a proof in Lean. According to media reports, GPT-6 Astra and Claude worked on the derivations, while physicists steered and checked the work. ⚠️ On X they announced “solutions to five long-standing problems,” but their own page says that for at least one of the five, the core question is still open. So this is progress, not five solved problems. ⚙️ A year-long roadmap done in 5 hours Google showed how Gemini 4 Argon and agents in Antigravity did in 5 hours work the team had planned for about a year. The software handles graphs with trillions of connections. On specific tasks it became up to 180 times faster, and the new code for Google’s chips (TPUs) runs 8 to 15 times faster. ⚠️ “One year” is the team’s own estimate, not a measured time. And an important detail: humans set the architecture and the acceptance criteria. The agents wrote the code, debugged it and checked each other’s work. 😴 AI that sleeps Today’s models stop learning once training ends. Whatever you tell them in a conversation, they forget when the conversation ends. Google Research proposes a model that “sleeps.” While awake, it works and stores new information in temporary memory. When it “falls asleep,” it stops taking requests and builds what it learned into itself. Then it “dreams”: it makes up its own examples and practices on them. In tests it learned new languages without forgetting the ones it already knew. ⚠️ This is research, not a product, and the paper itself is from June. What’s new is that Google presented it publicly. But the direction is clear: models that keep learning while they work. 🌍 AI enters the physical world 🎮 Google DeepMind released Genie 4. It’s a “world model”: you describe what you want and get a world you can move around in. It now has sound and music, characters talk, and while you’re inside you can use text to change the weather or camera angle, or add people. The world stays consistent for about 2 minutes. For now it’s only open to testers. For context: when Google released the previous version in January, shares of Unity, Roblox, Take-Two and Nintendo fell. 🤖 Machenta, a startup spun out of Imperial College London, came out of stealth with a $40 million seed round. They say their robot learns a new task from a single demonstration, with no extra training. The model was trained entirely in simulation, without a single frame from the real world. The founder calls it “physical vibe coding.” ⚠️ They’ve published impressive tests (74 out of 74 successful attempts on one task), but only on simple tasks so far. They have no announced customers, and simulations often fail once they meet real friction and lighting. 👥 What this means for people 🇳🇱 The Netherlands plans to cut about 25,000 jobs in its central government, around 16%, over roughly a decade. The goal is savings of €900 million a year by 2027 and €3.5 billion a year from 2035 (according to Bloomberg). AI is meant to fill part of the gap, and a plan for exactly how is expected in early 2027. Police and defense are exempt. If it happens, it will be one of the first big tests of whether AI can really “shrink” a government. 🩺 FutureClinic, a Y Combinator startup: you can now tell Claude or ChatGPT “find me a doctor,” and the agent books you a real consultation. The AI only collects your history. A licensed doctor handles your case and can write prescriptions. For now it only works in the US. But it’s a new model: your agent becomes your front door to services. 🛡️ The UK’s National Cyber Security Centre (NCSC) added a new chapter on AI to its list of research problems. In their words: “Many of the assumptions that underpin today’s cyber security best practice were developed before this kind of AI existed.” In other words, the rules have to be rewritten. 🇪🇺 Europe Switzerland’s Giotto released two models. Giotto Alto is free, has an app, and keeps your data in Europe. Giotto Solo runs on a single graphics card, so companies can keep it on their own servers. They call them “the smartest European sovereign AI models.” ⚠️ They haven’t published a single benchmark to back that up. I like that Europe is building its own, but “smartest” without numbers is marketing. 🧠 The next frontier: the brain Juan Benet, the creator of IPFS and Filecoin, has been investing in neurotech since 2021. He says it “now feels like AI did in ~2019,” just before the big boom. His fund PLC Neurotech backs companies like Synchron and Paradromics, which make brain implants, and Merge Labs. ⚡ Quick hits In its IPO filing, Anthropic names only two direct competitors: Google and OpenAI. Intel is acquiring Israel’s Classiq, which makes software that automatically builds quantum algorithms. 🔬 Bonus science Researchers at ETH Zurich published an implant in Science that “burns” excess blood sugar with electricity. A sensor tracks glucose levels, your phone or watch switches the implant on, and the glucose is turned into a substance the body passes in urine. In diabetic mice it regulated blood sugar, the mice lost weight, and the implant worked reliably for six weeks. It hasn’t been tested in humans yet. MIT and partners released Lumina, a simulation of the first 2 billion years of the universe. It covers a patch of space 1.6 billion light-years across with about half a trillion elements, and it ran on a supercomputer using about 15,000 graphics cards. You can watch the universe “light up” after the dark ages. Finally I noticed something common to almost every one of these stories. AI does a huge share of the work: it writes 40 million lines, optimizes, proves. But humans set the goal, and humans check whether the result is right. And that’s the catch. The headlines on X are bigger than reality: “solutions” that are partial, “a year in 5 hours” that’s an estimate, “the smartest model” without a single number. The more work AI does, the more valuable the people who can say what they want and check whether they got it. So my advice for this week: give AI something bigger than a question. A whole project. Then check everything it hands back. Live your life, spend time with the people you love, and try something new with AI this week. ☮️ § ❤ SOURCES IN THE FIRST COMMENT

Every camp just got its proof Last night OpenAI announced that one of its AI models has refuted a hypothesis in computer science that hundreds of research papers rely on. A day earlier, Microsoft offered 60,000 people a voluntary exit, and Anthropic wrote in its IPO filing that AI could pose “catastrophic or existential risks to humanity.” If you follow AI, you’ve probably noticed that people react to the same news in completely different ways. As I see it, there are five camps right now: 🫧 The bubble-callers. Investors and economists who think the money pouring into AI is far ahead of real profits and that it will all burst like the dot-coms in 2000. ☢️ The doomers. Researchers and people in AI safety who think we’re building something smarter than us without knowing how to control it. 💼 The ones worried about their jobs. Millions of workers who see AI already doing parts of their job, while no one tells them what comes next. 🛠️ The builders. People using AI to do things that were impossible yesterday. For them, every new model is a new superpower. ⚖️ The middle. People who are scared and excited at the same time, because they see both sides. I’m in the middle. And it turns out I’m not alone: according to Pew, 37% of Americans are equally worried and excited. Another 52% are more worried than excited, and only 9% are more excited than worried. The interesting part is that in just the last 48 hours, every camp got proof that it’s right. 🫧 The bubble-callers In China, the AI bubble has already burst. The STAR50 index of Chinese AI and chip companies fell about 40% in under four months, wiping out roughly $480 billion in market value. Robot maker Unitree is trading at about half its price from its August debut, when demand for its shares was about 8,000 times the supply. (Reuters) Torsten Slok, chief economist at Apollo: AI has delivered “zero financial payoff to corporate America outside of tech.” Goldman Sachs expects AI’s boost to the US economy to peak at the end of this year or early next year. The counterargument: Anthropic reported $16.3 billion in revenue for the first half of the year alone, against $4.6 billion for all of last year. It also reported an $80.7 billion loss, but that’s mostly an accounting loss, not cash actually spent. 💼 The ones worried about their jobs Microsoft is offering voluntary buyouts to about 60,000 employees, roughly 27% of the company, at vice president level and below in the US, Canada, India, and the UK. Microsoft itself doesn’t name AI as the reason. That’s the media’s framing, but it’s hard not to see it. Goldman Sachs: across US non-manufacturing sectors, the more firms use AI in their regular work, the slower employment grows in that sector. The effect is now showing up outside of tech. Bill Gates on The Verge: he expects a roughly 20-year transition with “winners and losers and a lot of disruption” before we reach abundance. He suggests reserving some jobs for humans as a policy choice: childcare, cooking, art, and sports. (You may have seen viral posts about “20 years of hell.” They heavily exaggerate what he said.) 71% of Americans expect AI to mean fewer jobs over the next 20 years. ☢️ The doomers and security Anthropic filed to list on Nasdaq under the ticker ANT. The filing says its technology could pose “catastrophic or existential risks to humanity” and that insurers may not have the capacity to cover a catastrophic AI incident. The most popular reaction on X: “Imagine asking people to invest billions in a product you’re warning could wipe out humanity.” A follow-up to “Black October”: hackers breached the cloud service of Japan’s GMO, which forex traders rent to keep their trading platforms running around the clock. The attackers took clients’ logins and placed trades from their accounts. Five brokers confirmed unauthorized trades, and up to 26,739 servers may have been exposed. GMO’s CEO Masatoshi Kumagai called it a national crisis and says ARTEX was used, the same Chinese AI tool behind the Korean bank breaches. That’s his claim; it hasn’t been independently confirmed yet. He also admitted his team had flagged the vulnerability but it wasn’t fixed in time. The irony: GMO’s own cybersecurity unit was the one that exposed ARTEX in attacks on Japanese companies a week earlier. 🛠️ The builders This is where the big news is. OpenAI announced three discoveries made by its models: 📐 Mathematics. They proved the local smoothing conjecture in every dimension, which also settles the famous Kakeya conjecture. Until now, Kakeya had only been proved in three dimensions, and that was just last year. 💻 Computer science. The Exponential Time Hypothesis (ETH) is false. Without getting technical: hundreds of research papers rest on the assumption that some problems can’t be solved faster than a certain limit. The model claims they can. This doesn’t mean computers suddenly become super fast, since the practical speedup is modest. In theory, though, it’s an earthquake. ⚛️ Physics. They proved that quantum chromodynamics (the theory of what happens inside protons and neutrons) rigorously exists mathematically. It’s also part of one of the seven Millennium Prize Problems, each worth $1 million, which OpenAI says it won’t claim. The important part: none of this has been checked by outside experts yet. ETH will be reviewed by the very people who proposed it 25 years ago, with no NDAs and no pay. The proof has been machine-checked, but the physics work hasn’t been published yet. And three weeks ago, when OpenAI released 722 math manuscripts at once, it withdrew three of them the very next day. So on X the enthusiasts are celebrating while the experts wait. The most-liked reply under the announcement: “I hope none of these are hallucinations.” More from the builders: About 30 physicists from Harvard, Google DeepMind, and other institutions gave AI 50 unpublished research problems in theoretical physics, each roughly a day’s work for an expert. GPT-6 Astra solved 49 of them for under $3 per problem, and Claude Opus 5.5 solved 46. The authors’ caveat: these are routine research problems, not breakthroughs. Mathematician Steven Strogatz in The Atlantic, “The Year AI Ate Math”: his colleagues are torn between awe and grief for the profession they know. OpenAI released GPT-6.1 Terra. They say it matches their strongest model on a test of how well agents use software at work, at a tenth of the price. More than 25,000 nonprofits now use Claude, most of them small local teams. 🧠 The idea that lifted a weight off my shoulders If you read just one thing from this post, make it this. Venkatesh Rao’s essay “Cognitive Mercantilism” lifted a weight I’d been carrying for months. Before capitalism, elites believed four things: wealth is finite and you only gain it by taking it from someone else; ordinary people can’t be trusted; only supply matters, not demand; and 80% of people will work the fields forever, because what else are they good for? Rao argues that today’s fear of AI repeats the same logic, just with intelligence: it’s scarce, it must be guarded from “the masses,” and if machines take over thinking work, we’ll be left with billions of useless minds. But the agricultural revolution didn’t create millions of useless mouths. It freed people from the fields. Rao can’t keep a houseplant alive, yet he eats a wider variety of food than any king did 300 years ago. His inability to produce food doesn’t matter. What matters is that he can consume it. The same could happen with intelligence. The question isn’t how much intelligence you can produce, but how much you can use. And demand for intelligence, like the demand for tea, sugar, and cotton once was, may have no ceiling. Cheap intelligence changes what we can even want. My favorite detail: Claude Code and a standard Netflix subscription cost about the same, around $20 a month. Rao suggests people will start choosing between them, just as people once cut back on food to buy a TV. For balance: not everyone agrees. Some think it’s naive about existential risk, and others think competing with another kind of intelligence will be harsher than anything capitalism brought. But for me it’s the best antidote to panic I’ve read. What to do Jascha Sohl-Dickstein, a researcher at Anthropic, gave advice on how to use the years before AGI. The talk is from June, but it just went viral: Bring AI into your work early, and pick problems that will get easier with better models, not obsolete. Don’t assume the traditional “safe” paths are automatically safe. Delegate execution to AI, but keep your judgment, and always check its work. And most important: “Writing is thinking. If you outsource writing, you are actually outsourcing your thinking as well.” (That’s why this post, for all the AI-assisted research behind it, was written and thought through by me.) Bonus science Physicists in Okinawa moved a diamond using nothing but the force of electron spins inside it, driven by green laser light. The setup weighs about 128 milligrams, which is huge by quantum standards. It moved only about 100 nanometers, but it’s an object big enough for gravity to matter. A step toward the question of where the quantum world ends and ours begins. Finally Every camp could take this week’s news and say “I told you so.” The bubble-callers have China. The doomers have Anthropic’s filing. The worried have Microsoft. The builders have the math. That’s why I’m staying in the middle. Not because I can’t choose, but because I think the middle is the most honest place to be right now. I’m scared because the speed is unprecedented and no one knows what the other side looks like. I’m excited because, for the first time in history, intelligence could become as cheap as food. Live your life, spend time with the people you love, and try something new with AI this week. ☮️ § ❤ SOURCES IN THE FIRST COMMENT