Claude proves Fermat's theorem

Claude cracks Fermat's theorem, GPT-6 lands on ChatGPT, and more.

Claude proves Fermat's theorem

Hi there, this is your daily ☕️ Techpresso.


In today's Techpresso:

🧠 Claude proves Fermat's theorem

💥 GPT-6 Astra hits top ChatGPT plans

💰 Top 1% drive 80% of OpenAI revenue

⚖️ Judge blocks X rival from Twitter name

🤝 US and China to hold AI safety talks

🚕 Tesla bars kids under 13 from Cybercab

Plus: 🎁 12 other news you might like, 🧰 6 tools, and 📚 5 papers.

🧠 Claude proves Fermat's theorem LINK

  • Anthropic says its Claude AI formally proved Fermat's Last Theorem, a puzzle that stumped mathematicians for 358 years, by writing what the company calls the longest math proof ever made, a chain a computer can check step by step.
  • Working mostly on its own over 11 days, dozens of Claude agents proved more than 30,000 supporting theorems and produced 13 million lines of code, more than five times the size of Mathlib, the shared library mathematicians already lean on.
  • The agents first lost track of their work, but a tool called Prove2Me gave each one a live to-do list to avoid duplication; Imperial College London's Kevin Buzzard reviewed the result and confirmed it holds up.
  • 💥 GPT-6 Astra hits top ChatGPT plans LINK

  • OpenAI has released GPT-6 Astra, its newest model, to ChatGPT users on Pro, Enterprise, and Business Premium plans through ChatGPT Work and Codex, with Plus and Business users getting access over the next few days.
  • The model also reaches developers through the OpenAI API, Microsoft Azure, and AWS Bedrock, while subscribers on Pro, Business, and Enterprise plans get GPT-6 Astra Pro, the stronger version, with usage counted against their existing plan quotas.
  • Standard Astra gives roughly half the messages per five-hour window as the older GPT-5.6 Sol, so Plus users get about 5 to 45 messages versus Sol's 10 to 100, and Free and Go tiers get neither model.
  • 💰 Top 1% drive 80% of OpenAI revenue LINK

  • New data from Ramp shows that the top 1% of customers bring in 80% of enterprise revenue at both OpenAI and Anthropic, a concentration that has held even as more businesses start paying for generative AI.
  • Ramp's lead economist Ara Kharazian called it a level of concentration risk unseen in any other software category the firm tracks, since usage-based pricing ties spending to how much the software does, not just how many people use it.
  • At Anthropic, coding tools Cursor and GitHub Copilot alone drove about $1.2 billion of its $5 billion revenue last year, while OpenAI has not said what share of its total revenue any single customer represents.
  • ⚖️ Judge blocks X rival from Twitter name LINK

  • A federal judge in Delaware stopped the startup Operation Bluebird from using the "Twitter" name for its rival social network, ruling in favor of Elon Musk's X on that core trademark.
  • The same court found X had likely abandoned two related trademarks, the word "tweet" and the Twitter bird logo, so the startup rebranded its site as Tweet.app and opened to the public.
  • Before launch, over 172,000 people asked for a handle on the site, with the company charging users $20 each to reserve one, money that likely helps cover its legal bills.
  • 🤝 US and China to hold AI safety talks LINK

  • The US and China are expected to sit down in mid-September for their first talks focused only on AI safety, marking a new official channel between the two rivals since Trump's second term began.
  • Treasury Secretary Scott Bessent will likely lead the US side, which wants labs in both countries to "police themselves" on AI-driven cyberattacks and share threat information, though the White House says no such meeting is confirmed yet.
  • Washington also plans to raise China's "distillation" of US models, where cheaper systems learn from larger ones, after accusing Moonshot AI of copying Anthropic's work, with any deal expected to stay narrow.
  • 🚕 Tesla bars kids under 13 from Cybercab LINK

  • Tesla's new Cybercab robotaxi will not carry children under 13 "at this time," a rule buried in the fine print released this week, even though its older Robotaxi Model Y SUVs allow riders as young as 8.
  • The Cybercab skips the standard LATCH anchors for child seats, so seats can only be strapped in with the seat belt, and Tesla still requires any passenger under 18 to be with an adult.
  • Tesla did not say why it bars young kids from the Cybercab, but the car uses electronic door latches, a feature that led to a recent recall of 3 million cars in China after a crash-safety investigation.
  • Other news you might like

    • US, Britain to coordinate on scam center takedownsLINK
    • Once popular for attacking AI, ASCII smuggling is embraced by spammersLINK
    • Google’s Gemini Spark can now manage your Google Photos libraryLINK
    • Seattle Times sues Microsoft and OpenAI, alleging they trained their AI on its journalismLINK
    • AMD unveils Threadripper Halo Station, an AI workstation packing 96 cores and dual liquid-cooled MI350P accelerators — 'the most powerful workstation in the world' can run trillion-parameter models, says AMDLINK
    • G42 gave up Huawei, then let Washington watch its data centres. Majority US ownership is reportedly next.LINK
    • AI visionary Ray Kurzweil is joining a Silicon Valley startup developing a brain computer system that involves snorting nano particlesLINK
    • OpenAI’s rogue agents keep escaping, with no formal process to investigate themLINK
    • Tesla Full Self-Driving will now overtake manual driving to avoid disasterLINK
    • OpenAI commits $1 billion to cyberdefense effort amid AI safety scrutinyLINK
    • Google warns of new Chrome zero-day flaw exploited in attacksLINK
    • TikTok's Parent Company Just Borrowed $30 Billion to Go All-In on AILINK

    🧰 Trending tools

    cmmnts: drop-in comment system with threads, moderation, logins, anonymous posts, Markdown, mentions, emoji, and GIFs, so you skip building one yourself.LINK

    Clockwork: schedule AI coding agents on a calendar, run them unattended in sandboxed git worktrees, pause for risky approvals, and review honest cost reports.LINK

    Reflexio: bundles reaction and focus mini-games like Lights Out, Schulte Grid, and Spatial Memory to measure and train speed, attention, and visual memory.LINK

    CommuteBar: shows live commute times in your Mac menu bar with saved destinations, route comparison, and leave-now alerts so you time trips without app-switching.LINK

    Experiential Labs: routes AI model calls across hosted providers, your own keys, and your GPUs with unified access control, automatic failover, and per-call cost attribution.LINK

    Retold: turns recorded family stories, spoken in their own voices, into hand-drawn animated films you can keep and share across generations.LINK

    📚 Trending papers & reports

    Denmark tree-species mapping shows that hand-built satellite features still beat ready-made AI representations for identifying forests nationwide, though the AI wins when labeled field plots are scarce, below roughly 25% of training data.LINK

    Neural video compression shrinks video files more efficiently by better handling tricky scenes with fast motion, blocked objects, and fine textures, delivering sharper quality at the same file size than a leading prior method.LINK

    3D scene reconstruction gets sharper by fixing a common training snag where distant pixels stop influencing the model, producing more reliable results from photos even with rough starting points, plus a new benchmark for moving scenes.LINK

    Chip design automation predicts where circuit components will physically sit early in the design process, cutting power use by ~7% and speeding up signal timing by ~27% versus current leading methods.LINK

    Speech cleanup tools can adjust themselves on the fly to unfamiliar background noise, sharpening audio quality without needing new labeled recordings, using a model of what clean speech should sound like.LINK


    See you tomorrow for a new dose of ☕️ Techpresso!

    More from the archive