In partnership with
AI Agents Are Reading Your Docs. Are You Ready?
Last month, 48% of visitors to documentation sites across Mintlify were AI agents—not humans.
Claude Code, Cursor, and other coding agents are becoming the actual customers reading your docs. And they read everything.
This changes what good documentation means. Humans skim and forgive gaps. Agents methodically check every endpoint, read every guide, and compare you against alternatives with zero fatigue.
Your docs aren't just helping users anymore—they're your product's first interview with the machines deciding whether to recommend you.
That means:
→ Clear schema markup so agents can parse your content
→ Real benchmarks, not marketing fluff
→ Open endpoints agents can actually test
→ Honest comparisons that emphasize strengths without hype
In the agentic world, documentation becomes 10x more important. Companies that make their products machine-understandable will win distribution through AI.
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
Japan Is Building AI to Beat the U.S. and China — and It's Not Another Chatbot
TLDR: Japan's four biggest tech companies formed a new AI company to build a trillion-parameter model for robots and machines, backed by $6.28 billion in government funding.
The Story:
Sony, SoftBank, Honda, and NEC each took a stake of more than 10% in a new company called Japan AI Foundation Model Development. The goal is to build a trillion-parameter AI model, one of the biggest ever attempted, but not for writing emails or answering questions. This model is made for physical AI: the kind that drives cars, runs factory floors, and controls robot arms. SoftBank and NEC will lead development. Honda plans to plug the results into its autonomous driving work. Sony brings robotics and gaming hardware. The company expects to hire around 100 AI engineers, with a SoftBank executive as president. Several major Japanese banks and steelmakers, including Nippon Steel, MUFG, and Mizuho, are also listed as investors. Japan's government agency NEDO has earmarked about ¥1 trillion ($6.28 billion) over five years to fund domestic AI development and Japan AI Foundation Model Development is expected to get a big piece of that. All data processing will stay on Japanese soil, including at a data center SoftBank is building in a former Sharp LCD factory near Osaka. The target for real-world deployment is 2030.
Its Significance:
Most of the biggest AI companies in the world are American or Chinese. Japan is betting it has a natural advantage in physical AI, meaning AI that runs real machines rather than just software. Japanese companies are already leaders in car manufacturing, industrial robots, and electronics — and they want an AI model built for that world, not one rented from OpenAI or Alibaba. The data privacy angle is a factor too. As more Japanese factories and banks run on AI, keeping sensitive operational data off foreign servers is a real concern. It's about which countries get to control the AI that runs the physical world in the decade ahead, and Japan just made a very loud statement about where it plans to be.
QUICK TAKES
The story: A coalition of 75 organizations led by the ACLU sent a letter to Mark Zuckerberg demanding that Meta drop plans to add facial recognition to its Ray-Ban and Oakley smart glasses. The feature, internally called "Name Tag," would let wearers instantly identify strangers in public, pulling up information through Meta's AI assistant. Internal planning documents showed Meta wanted to launch during a "dynamic political environment" because it expected rights groups to be too distracted to push back.
Your takeaway: Meta has paid over $7 billion in biometric privacy settlements in recent years. The coalition isn't asking for better safeguards — it's saying the feature can't be made safe through design changes or opt-out options, because the bystanders being scanned never agreed to anything in the first place. If this launches, anyone wearing Ray-Bans could quietly identify you on the subway, at a protest, or at a clinic. Whether or not Meta follows through, the demand is out in the open now.
The story: Chinese AI lab MiniMax released its new MiniMax M2.7 model — which scores 56.22% on SWE-Pro coding benchmarks, nearly matching Claude Opus 4.6 — and put the weights on Hugging Face. Developers noticed something odd: the license was labeled "MIT-style" but blocked commercial use. MiniMax's head of developer relations explained the change was meant to stop bad-faith hosting providers from deploying degraded versions and blaming MiniMax when the results were poor.
Your takeaway: Chinese AI labs built their reputation on fully open models. MiniMax's previous two releases were MIT-licensed with no restrictions. This is the first break from that, and it comes right after the company went public on the Hong Kong Stock Exchange in January 2026. Alibaba and Xiaomi have made similar moves. The assumption that Chinese labs will always stay open while American ones lock things down is looking less reliable by the month.
The story: The UK's AI Safety Institute tested Anthropic's Claude Mythos Preview and found it can autonomously execute complex cyberattacks. It completed "The Last Ones," a 32-step corporate network intrusion simulation that normally takes humans 20 hours, in 3 out of 10 attempts. It succeeded on 73% of expert-level security challenges that no AI could complete before April 2025. Anthropic isn't releasing it publicly — instead, 40 organizations including Microsoft, Google, Apple, and JPMorgan Chase have early access through a program called Project Glasswing, focused on finding and fixing vulnerabilities before they can be exploited.
Your takeaway: Anthropic found thousands of high-severity security holes in every major operating system and web browser. Most haven't been patched yet. One cybersecurity expert noted that defenders may have roughly six months before similar capabilities show up in open-weight models — at which point ransomware groups could use them too. These weak spots have always existed, there is just now a more efficient method of finding them.
TOOLS ON OUR RADAR
💬 LibreChat Free and Open Source: An advanced unified chat platform that brings all of your favorite artificial intelligence models from different companies into one single centralized interface.
💬 * Flowith Freemium: An sophistocated artificial intelligence workspace that uses an infinite visual canvas instead of a traditional chat interface to help you organize ideas interact with multiple language models simultaneously and deploy autonomous agents to execute image/video creations and agentic tasks. Seedance 2.0 available here.
🌐 Composite Freemium: An intuitive browser extension that functions as an autopilot to learn your repetitive web tasks and execute multi step workflows directly within your existing browser sessions.
🎬 Autosubbed Paid: A dedicated video editing utility that automatically generates accurate transcriptions from your audio and seamlessly embeds highly stylized text captions directly into your final video files.
TRENDING
OpenAI's CRO Sent an Internal Memo Calling the Market "As Competitive As I Have Ever Seen It" — Chief Revenue Officer Denise Dresser's four-page memo, obtained by The Verge, accused Anthropic of inflating its $30 billion run rate by roughly $8 billion through how it books cloud partner revenue. She also admitted OpenAI's Microsoft deal "has limited our ability to meet enterprises where they are." Ramp data shows Anthropic on track to surpass OpenAI in enterprise market share within two months.
Cloudflare Brought OpenAI's GPT-5.4 Into Its Agent Cloud — Cloudflare expanded its Agent Cloud platform to include OpenAI's GPT-5.4 and Codex, targeting enterprise teams building AI agents that run multi-step tasks. New tools include sandboxed code execution, persistent Linux environments for longer jobs, and Git-compatible storage. After Cloudflare's acquisition of Replicate, switching between AI model providers now takes just one line of code.
OpenAI Bought a Personal Finance Startup Called Hiro — OpenAI acquired Hiro Finance, an AI-powered personal finance app backed by Ribbit Capital and General Catalyst. The product shuts down April 20, all user data deleted by May 13, and the team joins OpenAI. Founder Ethan Bloch previously sold fintech company Digit for $230 million.
Stanford's Annual AI Report Found That Experts and Everyone Else See AI Very Differently — Stanford's 2026 AI Index showed a 50-point gap between experts and the public on AI's job impact: 73% of experts expect a positive outcome, only 23% of the public agrees. Just 10% of Americans say they're more excited than concerned. The U.S. also ranked last among surveyed countries in trust that its government will regulate AI responsibly, at 31%.
Vercel's CEO Said the Company Is Ready to Go Public — Guillermo Rauch told the HumanX conference that Vercel's revenue grew from $100 million ARR at the start of 2024 to $340 million by February 2026. He credits AI agents for much of that growth — 30% of apps on the platform were deployed by agents, not humans. Rauch said an IPO is coming, though the window stays frozen until OpenAI, Anthropic, or SpaceX goes first.
The National Archives Is Trying to Figure Out What AI Means for Sensitive Government Data — The head of the Information Security Oversight Office said AI has changed what the public expects from government data access, and agencies are struggling to keep up. The concern centers on controlled unclassified information (CUI), a category covering over 125 types of sensitive-but-not-classified government data. ISOO issued new guidance in March 2026 on how agencies should use AI tools when handling it.
A New Study Found AI Gets the Diagnosis Right but Still Misses How to Get There — Mass General Brigham researchers tested 21 AI models and found they reached the correct final diagnosis more than 90% of the time — but consistently failed at the reasoning steps in between, like building a differential diagnosis list. The study, published in JAMA Network Open, used a new benchmark called PrIME-LLM to measure the full diagnostic process, not just the final answer. Researchers concluded off-the-shelf models aren't ready for unsupervised clinical use.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
📚 Build a reading tracker that lets you log books like Good Reads
Build a reading tracker app in React. Let users add books with title, author, page count, and genre. For each book, show a progress bar with a page counter they can update, a status dropdown (Reading, Finished, Want to Read, Abandoned), and a 5-star rating system. Display summary stats: books read, currently reading, total pages read, and average rating. Include a filter panel to view books by status. Use a dark purple theme with a clean card layout.What this does: Tracks your reading list with progress bars for each book, lets you log how many pages you've read, marks books as finished when you reach the last page, and shows your total pages read and average rating across your whole library.
What this looks like:
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Help control machines, robots, and vehicles through large-scale "physical AI" models designed for the real world, not just text.
❌ Still Can't: Replace experienced doctors in reasoning through complex diagnoses — AI gets the final answer right but misses the steps in between.
✅ AI Can Now: Run complex 32-step corporate network attack simulations from start to finish, without a human in the loop.
❌ Still Can't: Guarantee open-weight AI models will stay free for commercial use — Chinese labs are starting to close the door.
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
The final book in the series brings the war to its conclusion as humanity makes its last stand against the alien invaders. All the character arcs converge in satisfying ways, and the action is relentless. This series had one of the wittiest AI characters in literature.
Are you tracking agent views on your docs?
AI agents already outnumber human visitors to your docs — now you can track them.
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.
