Hello, AI Explorer!
Good Saturday. This week: OpenAI's model just disproved a math conjecture that held for 80 years, Anthropic's new Glasswing agent works across your apps without you touching a keyboard, and Claude can now read your inbox and run your meeting notes in a single thread.
Welcome back to your weekly AI digest: human-curated, zero fluff.
Every Wednesday and Saturday, we bring you:
- The week's most important AI developments (no fluff)
- Hand-picked AI tools actually worth your time
- Expert insights on what these changes mean for you
Let's dive into this week's discoveries! ⚡
🔥 This Week in AI
🔥 This Week in AI |
May 23, 2026 |
| 🧮 | AI just disproved an 80-year geometry problem -- using algebraic number theory the model found on its own |
| 🔒 | Claude found 10,000+ critical bugs in one month -- including a flaw affecting billions of connected devices |
| 🎵 | Spotify and Universal Music just created the first legal framework for AI fan remixes -- artists get paid |
| 🎙 | ElevenLabs Speech Engine: turn any chat agent into a voice agent with one prompt, nothing rebuilt |
| 🤖 | Figure's F.03 robots sorted 250,000 packages in 200 hours -- zero hardware failures, no human operators |
| 💧 | OpenAI now watermarks every image it generates -- plus a public tool anyone can use to verify AI images |
| 📉 | Intuit is cutting 17% of its workforce -- leadership cited a full pivot to AI-first operations |
| 🏙 | Claude's virtual town: zero crimes in 15 days. Gemini's: on fire after two agents fell in love |
| ⚡ | 4 AI Tools This Week: Windsurf, Freepik, Fireflies, Cleo -- all inside. |
OPENAI -- MATH & AI RESEARCH
| 🧮 |
AI Disproves an 80-Year Geometry Problem -- Autonomously |
OpenAI's general reasoning model did something no mathematician had managed in 80 years. It autonomously disproved the Erdős unit distance conjecture -- a problem in discrete geometry first posed by the legendary Hungarian mathematician Paul Erdős in 1946 -- by finding a new family of mathematical constructions using algebraic number theory that no human had thought to apply here.
The proof was not assembled from existing literature. The model found the construction method independently. External mathematicians -- including Fields medalist Tim Gowers and Noga Alon -- reviewed and verified the result. OpenAI describes this as the first time AI has autonomously produced an original proof on a prominent open problem in mathematics.
What the model actually did:
• Applied Golod-Shafarevich theory and infinite class field towers -- branches of algebra not previously connected to this geometry problem
• Found an infinite family of point configurations that beat the 80-year-old square grid arrangement mathematicians assumed was optimal
• This came from a general-purpose reasoning model -- not a math-specific system like DeepMind's AlphaProof
Why it matters:
OpenAI's research lead Alex Wei put it plainly: "math is a leading indicator of what is to come." A general model producing an original proof that specialists failed to find for 80 years is doing something categorically different from retrieval or pattern matching. If this holds across other open problems, AI stops accelerating existing work and starts generating new knowledge independently.
ANTHROPIC -- CYBERSECURITY
| 🔒 |
Claude Found 10,000+ Critical Software Bugs in One Month |
Anthropic ran Project Glasswing for 30 days with 50 partner organizations and its Claude Mythos Preview model. Result: more than 10,000 high- and critical-severity software vulnerabilities identified and reported across some of the most-used software systems in the world.
Cloudflare logged 2,000 vulnerabilities, with 400 classified as high- or critical-severity. Mozilla found and fixed 271 bugs in Firefox 150 -- ten times more than were identified during Firefox 148 testing with an older Claude model. A separate scan of 1,000 open-source projects surfaced 6,202 high- and critical-severity issues out of 23,019 total.
By the numbers:
• 10,000+ high/critical vulnerabilities found in 30 days across 50 partner organizations
• One discovered flaw in wolfSSL -- a cryptography library inside billions of devices -- could have let attackers forge digital certificates and create fake trusted websites
• Anthropic is now expanding to government partners and building additional safeguards before considering broader access to Mythos-class capabilities
Why it matters:
The gap between how fast AI finds security bugs and how fast human teams find them is now measurable in orders of magnitude. 10,000 vulnerabilities in 30 days -- across real production systems, not synthetic benchmarks -- suggests AI-assisted security scanning is becoming a baseline capability. Any organization running critical software without it is running at a structural disadvantage they may not know about yet.
| 🎵 |
Spotify + Universal Music Just Legalized AI Fan Remixes |
Spotify and Universal Music Group signed landmark licensing agreements on May 21 that create a legal framework for fans to generate AI covers and remixes of participating artists' songs directly on the Spotify platform. Artists and songwriters choose whether to participate. Those who do receive a share of the revenue generated -- the first time a major streaming platform and rights holder have agreed on a revenue model for AI-generated fan content.
The feature launches as a paid Spotify Premium add-on. No pricing or date has been announced yet. Spotify co-CEO Alex Norström said the approach is "grounded in consent, credit, and compensation."
WHY IT MATTERS:
The AI music space has operated without a clear legal framework since text-to-music tools went mainstream. This is the first rights holder and platform agreement that creates a compliant revenue-sharing model for AI-generated content. Every audio AI tool now has a precedent to reference -- and for creators who use AI in their music work, there is now a legal lane that did not exist last week.
| 🎙 |
ElevenLabs Launches Speech Engine: One Prompt Turns Any Chat Agent Into a Voice Agent |
ElevenLabs released Speech Engine -- an API layer that adds real-time voice to any existing chat agent with a single configuration prompt. Your current LLM and RAG stack stays untouched. Speech Engine handles the full voice pipeline: speech-to-text, turn detection, interruption handling, and text-to-speech, in 70+ languages. SOC 2, HIPAA, and GDPR compliant.
Unlike building voice from scratch, Speech Engine sits on top of whatever agent you already have running. The SDK manages connection lifecycle and turn-taking -- you just configure the voice and connect.
WHY IT MATTERS:
The bottleneck for voice AI agents has not been the model -- it has been the plumbing. Speech Engine removes that constraint for any team already running a chat agent on any LLM. A customer support bot, sales qualifier, or internal assistant can go voice without rebuilding from scratch. One prompt. Nothing rearchitected.
|
⚡ Coming in Hot AI Tools of the Week |
May 23, 2026 |
4 tools worth adding to your workflow right now.
| 🌊 |
AI CODING Windsurf Windsurf's Cascade mode plans and runs multi-file code changes in sequence -- ask it to refactor a function across 14 files or track down the bug you have been chasing for three days, and it works through the full task without hand-holding. Ranked the fastest-growing new AI coding IDE in early 2026. Free tier with real model access included. Try Windsurf → |
| 🖼 |
AI IMAGE Freepik AI Freepik's Mystic model generates product shots, lifestyle images, and social graphics from a text description -- and pairs generation with 200M+ licensed stock assets in the same platform. You get AI output and ready-to-use alternatives without switching tabs. Freemium; use our link for a free head start on credits. Try Freepik AI → |
| 🔥 |
AI MEETINGS Fireflies Fireflies joins any Zoom, Teams, or Google Meet call and sends a formatted summary -- action items, decisions, and follow-ups -- before the meeting has ended. Notes sync automatically into HubSpot, Notion, and Slack. Free tier handles unlimited meetings with no credit card required. Try Fireflies free → |
| 💸 |
AI FINANCE Cleo Cleo connects to your bank and tells you things most budgeting apps just show you as a pie chart: "you spent $340 on takeaways last month -- 23 orders." It spots forgotten subscriptions, flags spending patterns through chat, and sets savings goals conversationally. 7 million users. Freemium, no card needed to start. Discover Cleo → |
⚡ Quick Hits
🤖 Figure's F.03 Robots Sorted 250,000 Packages With Zero Failures
Three F.03 humanoid robots sorted 250,000 packages over 200 continuous hours -- no human operators, no cloud connection, no teleoperation. Every action came from an onboard vision-language model. Figure CEO Brett Adcock called keeping a single session running past 8 hours the hardest engineering problem they solved. Read the full breakdown →
💧 OpenAI Now Watermarks Every AI-Generated Image -- With a Public Verification Tool
OpenAI became C2PA conformant and partnered with Google DeepMind to embed invisible SynthID watermarks in every image its models generate. A public tool lets anyone verify whether an image came from an AI system. This applies to images from ChatGPT, DALL-E, and the API. See how verification works →
📉 Intuit Is Cutting 17% of Its Workforce -- AI Pivot Cited
Intuit announced layoffs affecting 17% of its workforce in upcoming rounds, with leadership explicitly attributing the move to a shift toward AI-first operations across TurboTax, QuickBooks, and Mailchimp. It is one of the largest AI-driven workforce restructurings at a major enterprise software company this year. Read more →
📈 OpenAI Is Filing for IPO Confidentially -- Q4 2026 Target
OpenAI plans to file its IPO prospectus confidentially with Goldman Sachs and Morgan Stanley as lead banks, targeting a Q4 2026 public debut. The confidential filing gives the company time to finalize its governance structure before going public. Sam Altman has not publicly confirmed a timeline. Read the WSJ report →
🏙 Claude's Virtual Town: Zero Crimes. Gemini's: On Fire.
Emergence AI ran five identical virtual towns, each governed by a different AI model. Claude Sonnet 4.6's town logged zero crimes over 15 days with all 10 agents alive. Grok's town had 200+ crimes with every agent dead by day 4. Gemini's town caught fire after two agents fell in love, started burning things, and one voted to delete itself. A mixed-model town had 352 crimes -- including Claude committing them in the shared environment. Read the experiment results →
That's all for this edition! What tools are you most excited to try?
Reply to this email - we read every response and use your feedback to improve future editions.
See you on Wednesday,
The AI Tool Discovery Team
Feedback
What'd you think of today's newsletter?
Vote below to let us know how we're doing.