A 23-year-old with no advanced math training just solved a problem that stumped the world's top mathematicians for 60 years. His only tool: a ChatGPT Pro subscription and a free afternoon. In today's Beginners in AI 👇

The AI didn't follow the standard playbook. It used a formula well-known in other areas of math — but nobody had ever thought to apply it to this kind of problem. UCLA's Terence Tao said researchers had been making "a slight wrong turn at move one" for decades.

Stanford's 2026 AI Index just dropped. The finding that matters most: human scientists with PhDs still outperform the best AI agents by 2:1 on complex research tasks. --> An AI "agent" sounds like a robot that does your job. What it actually is: a language model connected to tools (web browser, code runner, file system) that tries to complete multi-step tasks by deciding what to do next at each step. --> The problem is the "next step" part. Each decision has a small chance of going wrong. String 20 decisions together and the errors compound. A human researcher holds the full problem in their head and course-corrects naturally. An agent loses the thread around step 8. --> This is why your AI coding assistant is great at writing one function but struggles to architect an entire app. Same technology, different task length. --> AI agents will get better. But right now, treat them like a junior researcher: great at fetching, summarizing, and first drafts. Don't hand them the whole project unsupervised. What's the longest task you've successfully handed to an AI agent start to finish?

80% of enterprise workers are either avoiding or actively rejecting their company's AI tools. The number that explains it: only 9% of workers trust AI for critical decisions. 61% of executives do. That's a 52-point trust gap. --> Trust follows experience. Experience follows training. A third of the workforce has never used AI at all. --> Think of it like handing someone a power drill and saying "build a deck" without showing them which bit to use. They'll either avoid it or strip every screw. --> Companies are spending $54 million on average for digital transformation. 40% of that spend is underperforming because of adoption failures. The tools work fine. The rollout doesn't. --> If you're the person avoiding AI: pick one 10-minute task you do every day and try it in Claude or ChatGPT this week. One task. That's the entire on-ramp. Which side of the 52-point trust gap are you on?

Anthropic just launched Claude Design. You describe what you want, it builds slides, prototypes, and one-pagers. The part worth understanding: what changes when "design" becomes a conversation instead of a skill. --> Traditional design tools assume you already know what good design looks like. You pick layouts, adjust spacing, choose colors. The tool executes your vision. Conversational design flips that. You describe the outcome ("a one-pager for our Q2 investor update") and the AI handles layout, hierarchy, spacing, and typography. --> Your job shifts from pixel-pushing to describing what the audience needs to walk away understanding. The underlying tech is the same language model that writes code and analyzes documents. It learned visual patterns from millions of existing designs the same way it learned sentence patterns from text. --> Claude Design exports to PDF, PPTX, or directly into Canva. Available now on Pro, Max, Team, and Enterprise plans. --> A polished client deck in 10 minutes instead of 2 hours in PowerPoint. A product mock for a stakeholder meeting without waiting 3 days for the design team. Which of those would change your week?

Anthropic made an AI model that can hack into any major operating system and web browser. Then they refused to release it. Here's why that actually matters, and what "zero-day vulnerabilities" are in plain English.

So what did Anthropic do? They didn't release it. They gave it to about 40 organizations (Amazon, Apple, JP Morgan, Goldman Sachs) so they could find and fix their own vulnerabilities first. They're calling it Project Glasswing.

Here's what this means if you're not a hacker or a bank CEO: AI models are now good enough at writing and reading code to find flaws that human experts miss. That capability won't stay locked up forever. Other companies are building similar models right now.