AI Solved 370 Math Problems, Mathematicians Panic, Musk Seeks $40 Billion: This Week's Highlights

Topic of the week: AI breaks math in three hours

OpenAI has publicly released more than 700 manuscripts containing full and partial solutions to 370+ open math problems, including many of the most famous unsolved ones. The results weren't obtained through expensive computational agents, but by running several hundred problems through a new internal model. Each one took an average of three hours of work. The result turned out to be so striking that leading mathematicians around the world are tearing their hair out.

Reactions range from admiration to full-blown panic. Terence Tao published an open letter to the Association for Human Mathematics — an organization that requires members to pledge not to cooperate with commercial AI labs and not to publish papers generated by neural networks. The most radical wing, the AI-free caucus, has vowed not to use AI in research at all. Tao proposed rethinking the paradigm through "Math 2.0" — measuring the value of a solution not by the number of problems solved, but by the amount of new understanding, ideas, and progress in mathematics that people manage to extract from it.

Scott Aaronson has the most level-headed take on the situation. OpenAI solved the Unique Games Conjecture, which his wife Dana Moshkovitz had worked on for nearly her entire career, and their nine-year-old son later teased her: "Mom, a robot solved your problem." Aaronson's reaction is drama-free — this is a major paradigm shift, but mathematics isn't over because of it.

As for concrete examples: 42% of the published manuscripts were successfully formalized in the Lean language (which allows proofs to be verified automatically). OpenAI withdrew three manuscripts due to errors the very next day and corrected 14 more. The neural network solved the problem of packing 11 identical squares into the smallest-area square — a problem mathematicians had worked on for nearly half a century.

Cryptography remains a mystery: OpenAI didn't publish a single new solution there. Ethereum developer Justin Drake warns that AI could accelerate the discovery of vulnerabilities in the cryptography underpinning Bitcoin and Ethereum. Cryptographers have even started talking about a "bunker mode."

Anthropic took a different path. Instead of releasing a raw solution to the 3SUM problem, the company handed it to two mathematicians — Virginia Vassilevska Williams and Josh Alman — and paid them to break down the result, simplify it, and turn it into a full-fledged scientific paper.

News from Russia: tightening the tap

The Ministry of Digital Development is discussing a fee with carriers for large amounts of foreign traffic — talk is of a 50 GB monthly cap for 5G networks (in spring, a stricter version with a 15 GB threshold was proposed).

The tap on capital outflow is being shut off in earnest. Raiffeisen is introducing a 0.5% fee on ruble payments to foreign banks, UniCredit a full 2%. The Prosecutor General's Office is seeking the confiscation of assets belonging to Timur Turlov and Freedom Finance (many people used this broker to move rubles abroad). The export of cash rubles to Kazakhstan, Armenia, and neighboring countries has been limited to 1 million rubles per person.

The FinEx funds are completing their liquidation: all assets have been sold off and moved into conservative instruments, with interest income going toward covering the liquidators' expenses. 5.5% of assets has been reserved for future costs, plus a 3% FinEx fee.

International news: France on the edge

France is increasingly rattling the stock market. The budget deficit is around 5.4% of GDP, national debt has crept up to 120%, and ten-year bond yields are approaching 5%. The yield premium over German bonds has exceeded 1.5 percentage points — the highest since the eurozone debt crisis. Corporate bonds worth $241 billion are yielding less than government bonds — the market considers French manufacturers more reliable than the French state.

AI finances: more money needed

Musk, through SpaceX, wants to borrow $40 billion to buy Nvidia chips. The plan: $10 billion in bank loans, $30 billion in investment-grade debt. Investors were sent a two-page pitch deck with pictures of space and an arrow pointing to future data centers "somewhere in the Universe."

Confusion has arisen around OpenAI's revenue: in September, reports cited $70 billion in annualized revenue; now that's been walked back to $50 billion. The discrepancy comes down to accounting method — Anthropic records the entire payment through AWS and Google Cloud as revenue, then shows the cloud partner's cut as an expense. Correctly stated: OpenAI has $50 billion in net revenue, which in Anthropic's equivalent format would be $70 billion gross.

Meta and Microsoft are cutting back on their use of Claude. Microsoft slashed its forecast from $1 billion+ a year to less than a third of that. At Meta, the number of Claude Code users dropped from 60,000 to 30,000, with employees being shifted to Muse and Copilot.

According to SemiAnalysis calculations, Opus 5.5 delivers five times more useful output for the same money compared to GPT-6.1 Sol. The new Haiku 5.5 costs the same as GPT-6 Luna while outperforming it on most benchmarks. This week, Max and Team plans received API credits: $100 a month for Max 5x, $200 for Max 20x, and up to $500 for a team.

Catastrophe risk and ethics in AI

Executives at OpenAI, Anthropic, and other AI companies are working through scenarios for the first major catastrophe. The most likely scenario is a large-scale cyberattack that takes down banks, the internet, or critical infrastructure. Some of those interviewed expect this within the next six months to a year. Dario Amodei warned back in September that autonomous AI agents could cause hundreds of billions of dollars in damage.

CrowdStrike analyzed an attack on South Korean financial companies in which a hacker used Claude Code, Grok, DeepSeek, and GLM. In a comical twist, he left his AI sessions open, where he discussed where to sell the stolen data, asked for pointers to darknet marketplaces, and even asked Claude to write up a polished résumé of his hacking achievements.

Anthropic has banned systematic harassment of Claude. Starting November 12, the usage policy will include a separate ban on prolonged, pointless abuse. In experiments, Claude showed signs of "distress" and preferred to end toxic conversations. Anthropic emphasizes that there's no evidence of consciousness here — this could simply be a result of training.

TIME examined Meta Muse's internal instructions and found that the assistant updates its profile of the user — and of the people the user mentions in conversations — every hour. Muse keeps track of social connections, interests, and conflicts, and tries to infer unstated goals while picking out moments to act on them. People who don't use Muse themselves can end up in these profiles. Meta confirmed how the system works and says users can ask to have specific facts forgotten, but the original messages may remain in the system.