Presented by
Good morning, AI enthusiasts. A machine just solved math problems that professional mathematicians have worked on for decades, for a few hundred dollars each.
AlphaProof Nexus doesn't just move a benchmark. The same capability, AI-driven formal verification, applies to any system where proving something is correct has always been the expensive part.
In today's recap:
Google AI cracks 9 unsolvable Erdős math problems
OpenAI's $5.7B quarter and imminent IPO filing
Get brutally honest feedback from ChatGPT
The "Vibe Slop" alarm: Geohot vs. AI agents
4 new AI tools, prompts, and more
GOOGLE DEEPMIND
Google AI autonomously solves 9 Erdős math problems
Recaply: Google DeepMind just published AlphaProof Nexus, an AI that solved 9 decades-old Erdős math problems and 44 OEIS conjectures, using a large language model paired with the Lean formal proof verifier at a few hundred dollars each.
Key details:
AlphaProof Nexus runs an AI model that writes proofs, then passes them to Lean, a formal checker that rejects any step that doesn't hold. Failed proofs go back to the AI in a loop until the system solves the problem or gives up.
9 of 353 open Erdős problems were solved (2.5%), and 44 of 492 OEIS conjectures were proved (9%). Some of the Erdős problems had gone unsolved for more than 30 years.
Results appeared in arXiv preprint 2605.22763v1 on May 21, 2026, with all formal proofs in a public GitHub repository. The work builds on AlphaProof's silver medal at the 2024 International Mathematical Olympiad.
There's no public product yet. DeepMind sees applications in combinatorics, algebraic geometry, and mathematical optimization.
Why it matters: Erdős problems aren't designed to be tractable. Professional mathematicians have worked on some of them for decades without progress. Cracking 2.5% of them at a few hundred dollars each changes what's feasible in mathematical research. The same capability, AI-driven formal verification, applies directly to software and cryptography. Making things provably correct has always been expensive. AI is changing the cost equation.
PRESENTED BY VIKTOR
Last week Viktor wrote a brief, built a landing page, and opened a pull request.
Last week, Viktor wrote a campaign brief, built a landing page, opened a pull request, generated a board-ready PDF from live Stripe data, and sent a follow-up email to a churned customer. All from Slack. Same colleague that also pulls your reports and monitors your dashboards. 5,700+ teams. 3,000+ integrations.
OPENAI
OpenAI's IPO era begins as revenue nears $6B
Recaply: OpenAI just outlined plans to file for a confidential IPO within days, with Goldman Sachs and Morgan Stanley preparing paperwork as Q1 revenue reached $5.7B, driven by Codex, enterprise sales, and early advertising.
Key details:
Goldman Sachs and Morgan Stanley are handling the paperwork. A confidential filing keeps details private until 21 days before the roadshow, targeting a public offering in September 2026.
OpenAI has 905M weekly users and 55M paying subscribers. Q1 revenue hit $5.7B, but the company spent $1.22 for every dollar it made, posting a -122% operating margin.
A court dismissed Musk's lawsuit to block the IPO, though he's still appealing.
OpenAI raised $122B at an $852B valuation in March 2026. The confidential filing is expected this week, with a public offering targeting September.
Why it matters: OpenAI is burning $1.22 for every dollar it earns, but investor appetite hasn't slowed. Going public changes that. Quarterly earnings will force accountability that private fundraising never required. Codex is becoming a real revenue driver, and enterprise sales are accelerating. There's a real case margins will improve. But once the prospectus drops, the world gets to check every 90 days.
GUIDES
Get brutally honest feedback from ChatGPT
Recaply: In this tutorial, you will learn how to disable ChatGPT's default sycophantic behavior using a custom system prompt, getting direct, unfiltered feedback that actually improves your work.
Step-by-step:
Open ChatGPT and go to Settings → Personalization → Custom Instructions.
In the "How would you like ChatGPT to respond?" field, paste: "No filler phrases. No praise. No unnecessary agreement. Mirror my communication style. If I'm wrong, tell me directly."
Start a new chat and test it by submitting something real, a draft email, a strategy, a business idea. The response should skip the usual opener and lead with substance.
If the feedback is still too soft, add "Lead with what's wrong" to your instructions. Some users also add "Don't soften criticism with praise."
Save these as your default custom instructions so every new conversation starts with this filter applied.
Pro tip: After getting feedback, follow up with: "Steel-man the strongest objection to this, then tell me if you agree with it." This forces ChatGPT to argue against your idea before giving its final verdict.
TOGETHER WITH HITOUCH
AI ads that look and feel like your brand
Most AI tools fall short because they lack context. They generate in a vacuum.
Hightouch Ad Studio uses your data and brand guidelines to produce high-quality creative. Refresh ads based on performance, react to trends, and respond to competitors instantly.
Less time prompting. More time launching.
AI RESEARCH
AI agents can't build quality software
Recaply: George Hotz just argued that AI agent adoption will be "one of the most costly mistakes" in software history, warning large organizations are producing buckets of slop while high performers carefully read every agent-generated line.
Key details:
Agents get started fast but consistently struggle to finish the job. Hotz spent 6 months testing the theory on real tinygrad code and a USB/PCIe chip reversing project. "The agent frontloads all the progress, then gives you a slot machine lever to pull," he wrote.
Hotz tested every major model, harness, and prompt combination before reaching his conclusion. He now aligns with the LeCun/Marcus camp, which questions whether large language models can ever truly program in a meaningful sense.
Large organizations have slower feedback loops and less instinct for error correction. Bottom performers won't filter the slop, and average quality deteriorates. "It is a golden era for buckets and buckets of slop, and a dark age for gems of quality," Hotz wrote.
The post went up May 24, 2026. Hotz frames a concrete test: will macOS get better or worse in the next 2 years as Apple pushes AI tools on all its engineers?
Why it matters: Hotz's strongest point isn't the claim but the mechanism underlying it. High performers catch bad output and correct it naturally. Large organizations with slow feedback loops and misaligned incentives often can't do the same, and the gap compounds over time. If he's right, AI adoption creates an asymmetric result: it raises the ceiling for high performers while pulling down average output everywhere else. The macOS trajectory over the next two years is a concrete test worth watching.
TOOLS
Trending AI Tools
🎙️ Personal Podcasts - Spotify's new feature that lets AI agents generate and save private audio briefings directly to your Spotify library.
🎥 Gemini Avatars - Google's new feature that clones your likeness in 5 minutes for AI-generated video
🧠 The Path - AI therapy app co-founded by Tony Robbins, scoring 95/100 on clinical validation versus 65 for consumer chatbots, with $14.3M in seed funding.
⚙️ Hermes Agent - NousResearch's open-source agentic framework with 166K stars and 27.4K forks, designed to grow with the user's workflow.
NEWS
What Matters in AI Right Now?
Hark raised $700M in a Series A at a $6B valuation, with Nvidia, AMD Ventures, ARK Invest, and Intel Capital investing.
Researchers from Zhejiang University demonstrated "AudioHijack," a technique hiding malicious commands inside podcasts, music, and YouTube videos to silently hijack AI voice assistants.
Microsoft CEO Satya Nadella disbanded the company's senior leadership team (SLT), replacing it with smaller, flatter teams reporting closer to the action.
Trump postponed signing an AI executive order hours before a planned photo op with tech CEOs, telling reporters he "didn't like certain aspects." Adviser David Sacks and Elon Musk pushed back, with one source calling the order "just something doomers wanted."
Anthropic's new enterprise consulting venture, backed by Blackstone and Hellman & Friedman, acquired San Francisco's Fractional AI as its operational centerpiece, ending Fractional's 11-month partnership with OpenAI. The deal targets midsize companies across thousands of private equity portfolio companies.
Starbucks retired its AI-powered inventory counting tool nine months after rolling it out across North American stores, after the app repeatedly miscounted milk types and other products. "Beverage components and milk will now be counted the same way you count other inventory categories," read an internal newsletter.
99% of CEOs expect AI-driven layoffs within the next two years, according to Mercer's Global Talent Trends report. Most anticipated headcount reductions are expected to focus on early-career positions, with only 32% of respondents believing the workforce can blend human and machine capabilities optimally.
Anthropic announced that Project Glasswing, using Claude Mythos Preview, has identified more than 10,000 high- or critical-severity vulnerabilities across major open-source projects, with 97 patched upstream and 88 advisories issued.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
