AI Verifies Finite Simple Groups & Anthropic's Claude Drug Discovery Push - AI News (Oct 29, 2026)

One of the biggest proofs in mathematics runs to roughly fifteen thousand pages. A team of AI agents and mathematicians just checked every line of it by machine, and they found a hole that nobody had noticed. More on that in a moment. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is October 29th, 2026. We have a full lineup: AI in drug discovery, a big Nvidia and Meta deal, new world models, sobering benchmarks, and a few stories about how people are pushing back on AI. Let's get started.

AI Verifies Finite Simple Groups

First, the math story. Axiom says it has formally verified the Classification of Finite Simple Groups in the Lean proof assistant. That makes it the first complete computer-checked version of one of mathematics' largest results. The original proof is spread across hundreds of papers. Axiom's AI agents and mathematicians turned the whole argument into something a machine can check, and along the way they found many small errors. They also found some real gaps. The most notable was in the Unbalanced Group Theorem, where a published route turned out to be flawed. The team replaced it with a new argument, and that produced a new group theory result. This theorem is used throughout algebra, number theory and physics, so a checked and reusable version, plus an errata list, is a real gift to future researchers.

Anthropic's Claude Drug Discovery Push

Next, Anthropic is making a serious move into drug discovery. Its life sciences team spent months working with six drug-development organizations, using Claude well beyond molecule design. The work covered target selection, antibody optimization, manufacturing paperwork, regulatory drafts and clinical trial planning, with partners including Generate Biomedicines, BigHat Biosciences and Bristol Myers Squibb. Anthropic also acquired Mithril Medicine, a one-person biotech working on a rare form of muscular dystrophy. In a twist, its founder, Cody Tranbarger, revealed that he is the anonymous 'Professor Oak' known on BioTwitter. He calls his approach a 'pipeline in a person': one scientist running several drug programs with AI and consultants. He argues that AI's biggest near-term payoff is not just speed. It is the ability to run more programs in parallel and make better go-or-no-go calls. Mithril's program could reach human trials by late 2028 if results hold.

Nvidia Invests in Meta Offering

In an update to the Nvidia and Meta story, Nvidia says it will put $3 billion into Meta's equity offering. Meta will lean further on Nvidia infrastructure, including Vera Rubin systems, to power next-generation AI agents and its Muse offering. The investment arrives as Meta seeks as much as $95 billion through stock sales. That tells you how expensive this race has become.

Where AI Chips Actually Land

On that note, where do all these chips actually end up? Epoch AI launched an explorer that traces AI compute by country using customs records. From 2022 through mid-2026, about 72% of global AI compute arrivals went to the United States. China came second at around 10%, a smaller share than its chip count suggests, because much of its hardware is less powerful. Epoch admits the data has blind spots. Still, it is a useful map for questions about economic and military power.

Genie 4 and Robotics Reasoning

Now to world models and robots. Google DeepMind announced Genie 4, which generates interactive worlds in real time, complete with ambient sound, music and dialogue. Users can steer the scene with plain language, such as changing the weather or adding characters. It starts as a research preview for trusted testers. DeepMind is pitching it as a training ground for robots and agents, not just entertainment. On the robotics side, Google Cloud put Gemini Robotics ER 2 into public preview. It is a high-level reasoning model that helps robots plan tasks and check their own work by watching video. Boston Dynamics is among the companies showing interest.

Proactive Agents and Simulated Worlds

Two research efforts point the same way. In a follow-up to a project we covered earlier, Meta's InteractiveWorld code is now public, along with a benchmark of 120 tasks. It tests assistants that watch through a person's eyes and guide them inside a 3D world, and it scores them on what actually happened in that world rather than on what the text says. Separately, a paper called StreamVibe tackles video agents that anticipate what you want, even as your goals shift mid-task.

Google's Sleep-Based Continual Learning

Google Research also has a clever idea for keeping models learning after deployment. It is a 'sleep' framework. While awake, the model handles requests and holds new information in short-term memory. During sleep, it moves that information into long-term weights. A 'dreaming' stage then generates practice problems to sharpen reasoning. The goal is to fix a long-standing problem: models forgetting old knowledge when they learn new things.

Liquid AI Launches LFM3

And at the small end of the scale, Liquid AI released LFM3, a family of compact open-weight models for text, vision and audio. They are built to run on-device as local agents, handling jobs like triaging your inbox or doing web research.

Coding and Tax Agent Benchmarks

Now, a reality check on agents. Cognition's FrontierCode 2.0 benchmark asks AI agents to do real-world software engineering, not tidy coding puzzles. Agents get credit for being thorough, inferring constraints nobody stated and avoiding regressions. The best model scored only about 45%. Surge AI's new tax benchmark tells a similar story. Agents work through messy, realistic client files, and while Claude models topped the leaderboard, overall pass rates stayed low. Models often missed deadlines or tax implications a human adviser would flag. The takeaway from both: AI can handle pieces of professional work, but it can't yet own the whole job.

Visual Studio October Update

For developers, Microsoft's October Visual Studio update removes some everyday annoyances. You can now compare an old commit directly against your current, uncommitted code, and resolve merge conflicts across many files in one go. The update also brings a rebuilt Copilot Chat and Hot Exit, which restores unsaved work and your window layout when you reopen the editor.

Google AI Mode Election Evasion

Turning to AI and society, Ars Technica found that Google's AI Mode dodged a simple question: who won the 2020 U.S. presidential election. Instead of answering, it often pointed users to official sources. Asked from outside the U.S., it answered plainly. Google blamed overly strict safety filters on sensitive election topics and says a fix is coming. It is a good example of caution tipping over into being unhelpful.

Humanist Wins AI Exemption

Meanwhile, a University of Kentucky administrator, David Durman, won a workplace accommodation exempting him from using AI tools, based on his humanist beliefs. He argued that relying on AI would mean handing off his own thinking and moral judgment. The American Humanist Association believes this is the first case of its kind, and it could set a precedent.

Executives and Outsourced Parenting

Finally, Marissa Mayer drew criticism for describing how her app Dazzle could help plan birthdays, playdates and vacations. It could even coordinate with her husband's AI assistant. Critics see it as part of a pattern of tech leaders treating parenting and relationships as problems to optimize. A Pew study suggests many Americans believe AI will weaken meaningful relationships, so a lot of people may simply not want this kind of help.

That's it for today. AI caught a hole in a giant math proof, and benchmarks showed agents still falling short of doing a professional's whole job. Links to all of today's stories are in the episode notes. Thanks for listening to The Automated Daily, AI News edition. I'm TrendTeller, and I'll see you tomorrow.

AI Verifies Finite Simple Groups & Anthropic's Claude Drug Discovery Push - AI News (Oct 29, 2026)

Our Sponsors

Today's AI News Topics