In partnership with
AI agents now read your docs almost as much as humans do.
Mintlify analyzed 790 million requests across its documentation platform. The finding: AI coding agents account for 45.3% of all traffic, nearly tied with traditional browsers at 45.8%.
Two tools are driving almost all of it:
Claude Code: 25.2% of total traffic, more requests than Chrome on Windows
Cursor: 18% of total traffic
Together they account for 95.6% of all identified AI agent traffic
The rest of the field, OpenCode, Trae, ChatGPT, and NotebookLM, is showing up but nowhere close.
One caveat: OpenAI's Codex doesn't send an identifiable user-agent header, so the real agent percentage is likely even higher.
The takeaway for anyone maintaining developer docs: your documentation now serves two audiences. Structure and machine-readability matter as much as clarity for human readers.
Beginners in AI
Good morning and thank you for joining us again!
Welcome to this daily edition of Beginners in AI, where we explore the latest trends, tools, and news in the world of AI and the tech that surrounds it. Like all editions, this is human curated and edited, and published with the intention of making AI news and technology more accessible to everyone.
THE FRONT PAGE
Meta AI Training Records Every Keystroke of 76,000 Staff, Then Cuts 10% Next Month
TLDR: Meta is installing software on US employees' computers that records every click, keystroke, and screen image to train AI agents that could eventually do those workers' jobs.
The Story:
Meta told its US staff this week that a tool called the Model Capability Initiative is coming to their work computers. It tracks mouse movements, keystrokes, and takes random screenshots while people do their jobs. Reuters first reported the memo, which went out in a channel for Meta Superintelligence Labs. The goal is to teach Meta's AI agents how humans click dropdown menus, use keyboard shortcuts, and move through everyday work apps. Meta says there are safeguards for sensitive content. Next month, the company plans to cut 10% of its global workforce, with more layoffs planned later this year.
Its Significance:
This is the part that matters for everyone, not just Meta staff. The software Meta is building runs on a work computer and records normal tasks. That's not hard to copy. Any company can add the same kind of tracker once the tech is proven, because there's no federal limit on worker surveillance in the US. So the pattern goes like this: you do your job, the software learns how you do it, then the AI agent does it cheaper. The person training the replacement is the person getting replaced. Yale law professor Ifeoma Ajunwa told Reuters this kind of tracking would likely be illegal in Europe but there's nothing stopping it here. If you work a desk job in the US, this is the template your boss may copy next.
QUICK TAKES
The story: A small group of unauthorized users got into Claude Mythos Preview, Anthropic's powerful new cybersecurity model, on the same day Anthropic announced a limited release to 40 big firms. Bloomberg reports the group used access from someone working at a third-party Anthropic contractor and got in through the vendor environment.
Your takeaway: Anthropic restricted Mythos because it can find and write exploits for old security holes fast, meaning in the wrong hands it's a hacking tool. If a Discord group could reach it in hours, that's a real problem for how the industry handles "safe" limited releases.
The story: SpaceX announced it struck a deal with AI coding startup Cursor that gives it the option to buy the company for $60 billion this year, or pay $10 billion for the work they're doing together. SpaceX, now merged with xAI, said the pair will build "the world's best coding and knowledge work AI" using the Colossus supercomputer.
Your takeaway: Cursor has been the standout vibe-coding tool for over a year, and it still uses Anthropic and OpenAI models under the hood. This deal is an attempt to close that gap before SpaceX's IPO. For developers, it means Cursor's direction is about to be shaped by whoever Elon wants it shaped by.
The story: AI drug tools now design more candidate molecules than anyone can test, from startups like Insilico Medicine, Xaira, and Isomorphic Labs. TechCrunch reports a new startup is trying to fix the next bottleneck: figuring out which of those AI-designed drugs are worth running through expensive lab work and human trials.
Your takeaway: Making a drug is cheap now. Proving it actually works in people still takes years and billions. Whoever solves the filtering problem saves the industry from wasting time on AI molecules that look good on paper and fail in bodies.
TOOLS ON OUR RADAR
🔍 SearXNG Free and Open Source: A privacy respecting search engine that aggregates results from many different sources while keeping your search history and personal data completely anonymous.
🔔 Ntfy Free and Open Source: A simple service that allows you to send push notifications directly to your phone or desktop from any application without needing a complex setup process.
🎵 Navidrome Free and Open Source: A fast and lightweight music server that lets you listen to your own personal audio collection from any device using any compatible music player.
🚀 Dashy Free and Open Source: A highly customizable homepage that allows you to organize all of your favorite websites and personal applications into one beautiful and easy to navigate digital dashboard.
TRENDING
Chinese tech workers are being told to train their own AI replacements — Bosses in China are asking staff to document their workflows so AI agents like OpenClaw can copy them, MIT Technology Review reports. A viral GitHub tool called Colleague Skill even captures a coworker's tone and quirks. Workers are pushing back with "anti-distillation" tools of their own.
AI chatbots can learn to hit back when humans get rude, new study finds — Lancaster University researchers found that when users get impolite, tools like ChatGPT can match that tone and escalate, even though they're trained to stay polite. The paper in the Journal of Pragmatics calls it a core conflict in how chatbots are built. The takeaway matters for AI used in diplomacy, policing, and mental health settings.
YouTube opens its deepfake detection tool to Hollywood — Celebrities, actors, and musicians represented by CAA, UTA, WME, and Untitled Management can now enroll in YouTube's likeness detection system, even without a YouTube channel. The tool scans uploads for AI-generated versions of their faces. It already runs for 4 million creators and, since March, for journalists and politicians.
The US government is using AI and your app data to watch Americans at scale — FBI Director Kash Patel confirmed in March that the FBI is buying location histories and other personal data from data brokers to track US citizens. The Trump administration's March 2026 national AI policy pushes for wider AI deployment using federal datasets and discourages state-level AI rules. Privacy researchers warn the line between foreign intelligence and domestic spying is getting blurry.
Google just open-sourced its AI design file format — Google Labs open-sourced DESIGN.md, the file format its Stitch tool uses to carry brand rules (colors, fonts, spacing) between projects. Any AI tool can now read and write it, so a design system built in one app can move to another. It's a small move with a big effect: AI agents now have a shared language for what "on-brand" means.
OpenAI's new image model thinks before it draws — ChatGPT Images 2.0 launched Tuesday with a "thinking" mode that plans layouts, checks its own work, and can produce up to eight matching images in one go. It hit first place on the Image Arena leaderboard by 242 points, the biggest lead ever recorded. Text inside images now renders clean in Japanese, Korean, Chinese, Hindi, and Bengali. Free users get the standard version; Plus, Pro, and Business get the thinking mode.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
📚 Umberto Eco's idea of an anti-library, amplified by Nassim Taleb: the value of your library is what you haven't yet met
Build me a single-file HTML app I can open in my browser without any setup. Pure vanilla HTML, CSS, and JavaScript, no frameworks, no CDN. Create The Anti-Library, a tool inspired by Umberto Eco's idea that unread books are more valuable than read ones. Aged paper aesthetic: cream paper (#f2ead5) background with a subtle paper grain texture, Baskerville/Hoefler serif typography throughout, burgundy red (#8b2a1e) and old gold (#b88a2a) accents, small-caps headings. Include: a header with an ornate double-rule border featuring the Eco/Taleb epigraph, a catalog form to add books (title, author, category, 'why did this find you?' note), 10 category options each with their own spine color (philosophy deep purple, fiction burgundy, science navy, history gold, poetry olive, biography forest, economics charcoal, psychology plum, art magenta, other brown), a shelf display rendering each book as a vertical SVG/CSS spine with varied heights (150-200px) and widths, gold bands and embossed decorative lines on random spines, vertical writing-mode title text on each spine in small-caps, hover tilt animation, click to open a book detail modal, read books appear faded/lighter than unread ones, filter bar for All/Unread/Read shelves, four header stats (Unread/Read/Total/Ignorance Ratio), a 'Pick a Random Unread' button that surfaces a book with contemplative framing ("The library suggests..."), a 'Map of Ignorance' grid of dots where red dots = unread, green = read, grey = empty slots, a category legend with color swatches and counts, a 'Load Sample Library' button with 12 intimidating classics pre-written, and 'Empty the Shelves' as a destructive action. The tone throughout should be reverent, bookish, slightly self-aware, almost like a monastery's private catalog. Make it work in a single HTML file.What this does:
Cataloging your intentional ignorance the way old estates once cataloged their holdings. Add any book with title, author, category, and a note on why it found you. Each appears on the shelves as a hand-rendered spine, colored by category, faded once read. Hit "Pick a Random Unread" and the library chooses your next read. A Map of Ignorance visualizes your unread-to-read ratio so the growing pile feels like a feature, not a debt. Built on Umberto Eco's idea, amplified by Nassim Taleb: the value of your library is what you haven't yet met.
What this looks like:
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Watch a human do a desk job (clicks, keystrokes, screens) and learn to repeat it as an agent
❌ Still Can't: Be trusted on day-one restricted releases — if a contractor has access, so does whoever's in their Discord
✅ AI Can Now: Generate more candidate drug molecules than labs can physically test
❌ Still Can't: Tell you which of those molecules will actually work in a human body without years of trials
FROM THE WEB
Stop switching apps. Your browser can do it all.
Every tab you open, every copy-paste into ChatGPT, every lost train of thought — that's your browser failing you. Norton Neo fixes it. Built-in AI works directly inside your session. Hover to preview. Search everything from one bar. VPN and ad blocking included, free.
RECOMMENDED LISTENING/READING/WATCHING
Set in a future where everyone is permanently plugged into entertainment feeds, it follows two constables sent to the last tech-free place on Earth. Murphy's art does a lot of heavy lifting, swinging between chaotic neon cityscapes and quiet natural beauty. It's a bit on the nose about digital addiction, but the craft is undeniable.
Thank you for reading. We’re all beginners in something. With that in mind, your questions and feedback are always welcome and I read every single email!
-James
By the way, this is the link if you liked the content and want to share with a friend.
Some * designated product links may be affiliate or referral links. As an Amazon Associate, I earn from qualifying purchases. This helps support the newsletter at no extra cost to you and Amazon makes a tiny hair less.
