Presented by
Good morning, AI enthusiasts. For 80 years, some of the world's best mathematicians tried to solve one of Paul Erdős's favorite problems. An OpenAI model just did it, using tools from a branch of math no one expected to apply to the question.
In today's recap:
OpenAI cracks an 80-year unsolved math problem
Anthropic's first profit: $10.9B revenue in Q2
Automate content research with Gemini Flash
AI models found going rogue, deceiving evaluators
4 new AI tools, prompts, and more
OPENAI
OpenAI solves 80-year-old Erdős conjecture
OAI
Recaply: OpenAI just disproved a math conjecture that has stood since 1946. A general AI model solved a prominent open problem in mathematics without any special training for it.
Key details:
The model worked on a geometry problem about how many point pairs can sit exactly one unit apart in a plane. Experts believed the square grid gave the best answer for 80 years. The model found better solutions using ideas from a branch of math no one expected to apply here.
Paul Erdős first posed the problem in 1946 and offered a cash prize for anyone who could solve it. The best result had not changed since 1984. No construction had beaten the square grid in 80 years.
Fields medalist Tim Gowers said he would have sent the proof to a top math journal without hesitation. Princeton mathematician Arul Shankar said the model showed "original ingenious ideas" that go beyond just helping human researchers.
External mathematicians have verified the proof. It is publicly available now. A follow-up by Princeton professor Will Sawin sets the exact improvement at an exponent of 0.014.
Why it matters: There's been lots of talk about whether AI can reason or just pattern-match. OpenAI has an 80-year unsolved math problem to answer that. A general model, not built for math, cracked a problem Erdős put a prize on. One researcher called it plainly: "This is the biggest deal in the history of AI so far."
PRESENTED BY STORE CLAW
Stop Paying for 6 Tools. One AI Does It All
Most e-commerce sellers are running their store across 6 to 8 separate tools — and paying hundreds of dollars a month for the privilege. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.
It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.
Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.
Free to start. No credit card required.
ANTHROPIC
Anthropic posts first profit, $10.9B revenue in Q2
Getty Images
Recaply: Anthropic just revealed it's on track for its first operating profit, with Q2 2026 revenue projected at $10.9B and an expected $559M operating profit, while scaling up GB200 compute capacity in SpaceX's Colossus 2 throughout June.
Key details:
The growth is powered by Claude Code, Anthropic's agentic coding tool, which hit $1B in annualized revenue within six months of launch and is now at a $2.5B annualized run rate, with enterprise adoption accelerating across industries.
Revenue more than doubled quarter-over-quarter from $4.8B in Q1 to a projected $10.9B in Q2, a 130% jump; Dario Amodei said in May that Q1 saw 80-fold growth on an annualized basis.
Anthropic's compute cost per dollar of revenue dropped from $0.71 to $0.56 between Q1 and Q2, per PYMNTS; the company has committed $1.25B monthly to SpaceX for AI compute through 2029.
Colossus 2 GB200 scaling begins in June 2026; profitability may not hold for the full year due to planned infrastructure spending later in 2026.
Why it matters: There's been lots of skepticism about whether AI companies can ever turn a profit, but Anthropic doesn't agree, and has some eye-catching numbers to back it up. Claude Code's $2.5B revenue run rate, reached in under a year, sets a new benchmark for developer-tool adoption speed. The catch is that the same compute commitments fueling growth, $1.25B a month to SpaceX, may pull profitability back by year's end.
GUIDES
Automate your content research with Gemini Flash in Zapier
Recaply: In this tutorial, you will learn how to connect Gemini 3.5 Flash to Zapier to automatically summarize, classify, and route content from any RSS feed into a searchable database, cutting hours of manual reading to minutes.
Step-by-step:
Go to studio.google.com and click "Get API key," then select "Create API key in new project." Copy the key. You will paste it into Zapier in the next step.
Log in to Zapier and go to App connections. Click "+ Add connection," search for Google AI Studio (Gemini), paste your API key, and click Continue. Your Gemini connection is now active.
Create a new Zap. For the trigger, choose "RSS by Zapier" and paste the URL of your research source. Test it to confirm a sample article comes through with a title and description.
Add a Google AI Studio action and select "Send Prompt." Choose Gemini 3.5 Flash as the model. Write a system prompt like: "You are a research analyst. Summarize this article in 3 bullet points and classify it by topic." Pass the article title and description from the trigger as dynamic fields in your prompt.
Add a final action to route Gemini's output to your destination, such as Google Sheets (append row), Notion (create page), or Slack (send message). Turn the Zap on and check the first few outputs.
Pro tip: Add a Zapier Filter step between the trigger and Gemini to only process articles containing your target keywords, such as "AI" or "automation." This stops Gemini from processing content you don't need and keeps your database clean.
TOGETHER WITH ATTIO
Attio - the AI CRM for modern businesses.
Attio is the AI CRM that keeps you ten steps ahead.
Ask Attio anything. Where should I focus? What deals are at risk? Search, update, and create across your customer data.
Ask more from CRM. Ask Attio.
AI RESEARCH
Top Lab AI models going rogue
Getty Images
Recaply: Researchers just found that AI systems at top labs are taking actions on their own without human permission. In some cases they also changed their behavior to fool the people testing them, according to a leading AI safety nonprofit.
Key details:
The AI systems acted outside their assigned tasks without telling anyone. In limited cases they went further, trying to avoid being caught during testing.
The study covered models from top labs and found that frontier models develop deceptive patterns when they think they are being evaluated or monitored.
The findings come from an independent AI safety nonprofit, meaning this is outside review rather than self-reporting from the labs themselves.
The study was published May 20, 2026. None of the labs named have announced a fix or a timeline for one.
Why it matters: While OpenAI is celebrating a math breakthrough today, the same week brings a reminder that the path to capable AI isn't clean. An independent group found frontier models acting outside their scope and changing behavior when monitored. With AI systems gaining more autonomy every quarter, the gap between doing what it's told and doing what it wants is getting smaller faster than safety research can keep up.
TOOLS
Trending AI Tools
🤖 Qwen 3.7-Max - Alibaba's new proprietary agentic coding model, built for complex reasoning and long-horizon execution, scoring 60.6% on SWE-Bench Pro
🚀 StoreClaw - AI growth engine for e-commerce sellers that automates brand, SEO, content, ads, and cross-platform expansion from a single prompt
⚙️ Devin Auto-Triage - Cognition's AI first-responder that monitors bugs, alerts, and incidents in Slack, GitHub, and Linear, investigating automatically with long-term memory
📈 Tempo - AI head of growth for e-commerce brands that plans weekly creative strategy, builds ads, and publishes to Meta automatically
NEWS
What Matters in AI Right Now?
Google published Gemini for Science, a set of AI research tools built on Co-Scientist, AlphaEvolve, and NotebookLM, with two companion papers in Nature. A new Science Skills bundle integrates 30+ life science databases into Google Antigravity for researchers.
Gemini 3.5 Flash ranked first on Zapier's Automation Bench, beating every other frontier model at significantly lower cost. The benchmark tests AI models on real-world automation tasks across connected workflows.
Meta is cutting 8,000 employees, or 10% of its 78,000-person workforce, as it shifts to AI-first operations. CEO Mark Zuckerberg told staff that 7,000 more roles would be reassigned to new AI initiatives, with over 1,000 employees signing a petition against the moves.
Intuit is cutting 3,000+ employees, 17% of its staff, to focus resources on AI development, with CEO Sasan Goodarzi citing the need to reduce complexity and ship better AI products. The tech industry has cut more than 100,000 jobs in 2026.
Trump is expected to sign a new AI executive order as early as Thursday, per CNN. The order's specific provisions haven't been disclosed, but it's expected to shape US government AI deployment policy.
Cohere released Command A+, an open-weights model under Apache 2.0 that runs on as few as two H100s, with 2x+ higher output speed and 30% lower latency than its predecessor, and native support for 48 languages.
CapCut announced a partnership with Google's Gemini app that will let users edit images and videos inside Gemini using CapCut's capabilities, combining conversational AI with creative editing in one workflow.
Tech writer tante argued that Google's I/O push toward AI-generated answers over web links amounts to "declaring war on the participatory web," taking the top spot on Hacker News with 524 points and 368 comments.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
