OpenAI Solved a $1M Math Problem. A Mathematician Says They Fought Dirty.

Yesterday might have been the strangest 24 hours in AI so far this year. OpenAI says it solved one of math's hardest problems in under four days, then got accused of racing dirty to beat two people who got there first. Meanwhile the NSA told American AI labs it's fine to quietly serve worse answers to suspected spies without telling them, and Meta shipped an agent that can spend your money after you've closed the app. Let's get into it.

The 8-hour workday is becoming a 4-hour workday for people who know how to use AI.

Everyone else is still catching up.

This AI work playbook shows you exactly how to cut your work hours in half using AI.

Sign up for Superhuman AI and get:

Claim your AI playbook

What happened: OpenAI announced that an internal model, more capable than its new GPT-6 Astra, ran roughly 10,000 AI agents for 88 hours (2.7 million messages, 130 billion tokens) and produced a proof for the Navier-Stokes existence and smoothness problem, one of math's seven Millennium Prize problems, which Astra then verified in proof-checking software called Lean. OpenAI says it isn't claiming the $1 million prize. NYU mathematician Tristan Buckmaster says he and an Anthropic researcher, Levent Alpöge, had quietly worked the same unusual approach for a year and posted a related proof the night before OpenAI's announcement, and that OpenAI only started chasing the problem after learning about their progress. OpenAI denies seeing their work directly, but admits it can't rule out that some training data came from how the mathematicians used its Codex product.

Why it matters to you: AI labs are now racing each other to claim famous "firsts," and the incentive to move fast and grab credit is stronger than the incentive to get the credit right. When one of these tools makes a big capability claim, that confidence isn't the same thing as an independently checked result.

What to do about it: Before you take an AI company's big announcement at face value, check whether independent experts have actually verified it, not just the press release.

What happened: Meta launched Muse this week, a personal AI agent built to handle real tasks: booking a table, buying something, filling out forms, sending email, using its own secure cloud computer to browse websites the way a person would. It runs through a phone app or WhatsApp, keeps working after you close it, and only comes back to you when it needs approval or something changes. Payments run through one-time-use card numbers tied to your account, and Meta says a separate "Sentinel" agent has to approve anything Muse sends out to the internet before it happens. It's US only for now, with a free tier and paid plans reportedly around $20 and $100 a month.

Why it matters to you: This is the clearest sign yet that "AI agent" is moving from chatbot to something with your money and your accounts. Convenient if it works as described, but you're trusting a company's internal gatekeeper agent to catch every mistake before it reaches your bank account.

What to do about it: If you try it, or anything like it, start with tasks that don't touch a payment method, and watch what it actually does before you hand it one.

What happened: The NSA, FBI, and CISA published a joint advisory saying China-based AI companies are running industrial-scale operations to copy the capabilities of US models, using tactics like fraudulent account pools that run around the clock and prompts designed to expose a model's hidden reasoning. The advisory's suggested fix for American AI companies is unusual: quietly serve degraded answers, shorter reasoning, different phrasing, lower quality, to accounts suspected of copying their models, and don't tell those users they've been switched to a worse version.

Why it matters to you: If you're a paying customer of any major AI tool, this is now official guidance that the company might silently hand you a downgraded version if their system flags your account, correctly or not, and you won't be told. It's aimed at suspected foreign copycats, but there's no guarantee the flagging is perfect.

What to do about it: If an AI tool's answers suddenly feel worse for no reason, it might not be you, and it might not be a bug.

What happened: TechCrunch reports that Anthropic is warning Claude users that hackers are using common infostealer malware to lift login sessions off people's computers, then using those sessions to access Claude accounts and burn through the victim's usage without them knowing. One Max subscriber noticed his token usage climbing while he wasn't even working; Anthropic later found a stolen session key had been used to generate unauthorized Claude Code access tokens. Anthropic has responded case by case, signing people out, invalidating logins, and issuing some refunds, but it still doesn't give users an itemized log of exactly what consumed their tokens.

Why it matters to you: If you or your team pay for Claude, or any AI subscription, this is a reminder that account theft doesn't always look like someone posting embarrassing messages. Sometimes it just looks like your bill or your usage cap creeping up for no clear reason.

What to do about it: Keep your devices free of junk downloads, and if your usage jumps without explanation, change your password and contact support before you assume it's a fluke.

What happened: Apple held its biggest event of the year on September 9, and for the first time under new CEO John Ternus, AI was the headline instead of an afterthought. The new Siri AI starts a beta rollout with iOS 27 on September 14 (English first, five more languages in October), the iPhone 18 Pro and Pro Max ship with a new A20 Pro chip built for more on-device AI work, and Apple also unveiled its first foldable phone, the iPhone Duo. Apple is pairing its AI photo tools with a verification system, SynthID plus a new Reference Image feature, meant to prove what a photo actually captured.

Why it matters to you: Apple has spent two years being the AI industry's biggest skeptic by necessity, and this event is Apple admitting it has to compete on AI terms now, not just camera and design terms. If a lot of your customer-facing work runs through iPhones, expect the AI features in the next update to actually matter this time.

What happened: Jacob Coxon, who spent three years doing pretraining research at OpenAI and then Anthropic, publicly resigned this week and said both labs are racing toward self-improving AI systems recklessly. A colleague at Anthropic, Evan Hubinger, echoed part of his concern, saying the team believes there's more than a 10% chance AI causes serious harm to humanity within the next decade, and that the company doesn't yet have a real plan to keep a future superintelligent system under control.

Why it matters to you: This isn't a random outside critic, it's someone who was inside one of the top labs saying the safety work isn't keeping pace with the capability work. You don't have to agree with the specific odds to take the basic warning seriously: the people building this stuff aren't all convinced it's safe.

What happened: French AI company Mistral raised €3 billion (about $3.6 billion) at a €21 billion valuation, the largest funding round any European tech company has ever closed, led by Samsung Electronics with EQT, PSG Equity, a16z, Nvidia, Salesforce Ventures, and others also participating. Mistral has never claimed to beat OpenAI or Anthropic on raw capability. Its pitch is sovereignty: European banks, governments, and defense contractors that don't want their data sitting on American servers.

Why it matters to you: "Which country's AI company built this" is becoming its own selling point, not just a footnote. If you sell to European clients or plan to, expect data residency and AI vendor nationality to come up in more contracts going forward.

Here's today's loop: a math problem got "solved" by a company that's also fighting over who deserves credit for it, spy agencies told AI labs it's fine to quietly serve you a worse product without saying so, and the newest AI agent on the market can spend your money after you've stopped watching. None of that happened because AI got scarier overnight. It happened because speed keeps beating honesty as the thing that gets rewarded. You don't have to panic about any single headline here. Just keep asking who benefits from you believing the biggest claim in the room, and keep your hand near the off switch on anything that touches your money or your accounts.

If this issue saved you from reading through 7 different AI newsletters to find the stories that actually mattered, forward it to one person who needs to stop scrolling and start reading this instead. New here? Subscribe and get the next issue tomorrow.

Talk tomorrow,
Mark Shilensky

Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social