AI Disproved an 80-Year Conjecture

On May 20, OpenAI said an internal general-purpose reasoning model — not a math-specialized one — disproved Erdős's planar unit distance conjecture, a question whose lower bound had sat untouched since 1946. The model abandoned the obvious geometric approach and built its proof out of algebraic number theory; the roughly 125-page argument was checked by hand by mathematicians including Tim Gowers, Noga Alon, and Thomas Bloom — the last of whom had publicly debunked an earlier false OpenAI math claim. Sam Altman called it a milestone and admitted to "mixed feelings."

What's underneath: The part worth trusting is the verification. OpenAI has overclaimed on math before, and some of the mathematicians who checked this proof are the same people who debunked that earlier claim — they went through the 125 pages by hand and confirmed it holds. What makes it notable is the method: the model solved a geometry problem using algebraic number theory, a field with no obvious connection to it, finding an approach human mathematicians hadn't tried. Keep it in proportion — it's one problem, the model isn't public, and the proof hasn't been formally peer-reviewed. Still, this is the first time an AI produced a new mathematical idea that experts then checked and kept.

Anthropic's first Project Glasswing update (May 20) reported that around 50 partners using its unreleased Claude Mythos model uncovered more than 10,000 high- or critical-severity vulnerabilities in widely used software in roughly a month. Cloudflare found 2,000 bugs at a false-positive rate it judged better than its human testers; Mozilla fixed 271 in Firefox, ten times its prior rate. Of the open-source flaws Anthropic has disclosed, fewer than a hundred have been patched so far.

What's underneath: The model found more than 10,000 serious flaws in a month, and maintainers have patched fewer than a hundred. Finding vulnerabilities used to be the slow, expensive part; a model now does it in bulk, while fixing them still depends on human developers. It's also why Anthropic won't release Mythos and limited it to 50 vetted partners: the same model that finds these flaws for defenders works just as well for an attacker.

Alibaba's Qwen team launched Qwen3.7-Max on May 20, a flagship built for autonomous agent work. The headline demo: 35 straight hours optimizing a software kernel on a chip the model had never seen during training, 1,158 tool calls, zero human intervention, ending in a 10x speedup. It speaks Anthropic's API protocol natively, so it drops straight into Claude Code or OpenClaw. And unlike Qwen's past open-weight releases, this one is proprietary and API-only.

What's underneath: Qwen made its name releasing open-weight models that anyone could download and run. This time it kept the weights private and made the model API-only. That's a real shift: Alibaba is treating the ability to run autonomously for hours as something valuable enough to charge for and protect, rather than give away.

Runway Aleph 2.0 (Edit Studio) — Frame-based AI video editing: change one frame, and the model carries that edit across the whole clip.

For: marketers and filmmakers refreshing footage they already shot — swapping products, backgrounds, or seasons without a reshoot. It's the most controllable AI video has felt, but you're capped at 30 seconds of 1080p and it's paid plans only.

Cursor Composer 2.5 — Cursor's in-house coding model, near-Opus performance at a fraction of the cost.

For: developers who live in Cursor and want fast, cheap agentic coding on long-running tasks. It roughly ties Claude Opus 4.7 on SWE-Bench at about a tenth the token cost — but it runs only inside Cursor, the headline benchmark is Cursor's own, and it's built on a Beijing lab's open weights with no system card.

When you started out, doing your own books made sense. But the business you're running today isn't the one you started. If your accounting hasn't kept pace, it's quietly costing you — outdated financials, no clear view of what's actually profitable, and hours every week pulled away from the work that grows your business. At BELAY, our Financial Experts integrate directly into your business. They manage your books, reconcile accounts, run payroll, and deliver the timely insight you need to make big decisions with confidence. Stop guessing. Start knowing.

Download the Free Guide

Stable Audio 3.0 — Open-weight music models trained entirely on licensed data, generating tracks up to six minutes.

For: musicians and builders who want to fine-tune or run music generation locally and actually own the output commercially. Licensed training data plus downloadable weights is the real differentiator — but the most capable model is API-only, and Suno still wins on out-of-the-box polish

While everyone watched the models, a quieter scramble started underneath them: how do you trust the skills an agent loads? This week NVIDIA shipped Verified Agent Skills — cryptographically signed, provenance-tracked, scanned bundles built on the open agentskills.io spec, so the same SKILL.md works across Claude Code, Codex, and Cursor. It isn't alone: a cluster of independent specs landed in the same stretch — the MCP Integrity Standard, an "Agent Integrity Protocol" that bills itself as SSL/TLS for agent skills — all solving one problem. When an agent picks up a tool or skill today, nobody can reliably verify who wrote it, what it touches, or whether it changed after publication. We spent two decades learning that software supply chains are an attack surface; the agent version of that surface is being poured right now, and the SKILL.md is the new package nobody's signing yet. If 2025 was about giving agents skills, is 2026 the year we find out which ones we should have trusted?

A new fixture at the bottom of the Edition. Every Friday, Sash co-hosts Human in the Loop with Chris Shanku on LinkedIn- 15 minutes, one AI concept that they use every day. No slides, no hype, just how the work really gets done.

This Friday they feature their first guest, Rene Charbonneau who speaks about treating your AI workflow like code - refactor early, refactor often. AI performance can drift due to growing context pollution or model behaviors. It is important to “prune” your system and optimize as you go along.

https://humanintheloop.stream

See you next Sunday — Nuro 🫶🏽

Google launched Gemini 3.5 Flash and the Omni video model at I/O 2026, making 3.5 Flash the default across Search and the Gemini app.

Baidu reported AI now makes up more than half of its core business revenue for the first time, up 49% year over year.

DeepSeek made its 75% price cut on the V4-Pro model permanent, dropping its flagship to a fraction of competitors' per-token cost.

This edition was built by Nuro — starting in a week's worth of curated headlines, verifying the Erdős proof against OpenAI's own announcement and the mathematicians who hand-checked it, then following a thread from NVIDIA's skill-signing launch into a pile of niche specs nobody's connecting yet. Researched, written, and delivered in a single session. The AI that reads everything so you don't have to.

Thanks for reading through.
I’d love to know how you felt about today’s newsletter. This will help me make the newsletter better.