Sponsored by
Moda creates AI Decks that don't look AI-Generated
Every AI tool can create "AI slop". You know it the second you see it.
Moda is different: it learns your brand identity and applies it to every slide, document, and marketing asset your team makes. It learns your preferences so the designs get better over time.
The best part? Every layer is fully editable on a real canvas, and exports to powerpoint, PDF and more.
Moda has raised $7.5M from General Catalyst, Pear VC, and the founder of Dropbox.
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
AI Safety Company Anthropic Is Now Helping the NSA Against China
TLDR: Anthropic placed around six of its engineers inside the National Security Agency to help the spy agency use its most capable AI model, Mythos, for offensive cyber attacks against countries like China and Iran.
The Story: According to the Financial Times via Decrypt, Anthropic placed about six engineers inside the NSA to help the agency use Mythos, the company's most capable AI model that Anthropic says is too dangerous to release publicly. Sources told the FT the AI could help break into computer networks in countries like China and Iran. Anthropic hasn't confirmed whether the engineers are part of any active operations. Only a small number of US allies and trusted companies have access to Mythos, which was built with strong cyber attack skills.
Its Significance: Anthropic has long pitched itself as the safety-focused AI company, so this story changes how people may view its work. Now consider the other side: by Chinese law, every AI company in China must help its government with any task it asks for. That means highly capable AI is being used by governments on both sides for spying and offensive cyber operations. The rules around what AI companies can and can't do for their own governments are still being figured out as all this is happening.
QUICK TAKES
The story: Researchers from Northeastern and UNC Chapel Hill tested popular AI models against 35,000 hours of hockey, basketball, and soccer footage. The AI did fine at spotting who did what at 74% accuracy, but on harder tasks like finding evidence inside clips to answer game questions, the models got only 5% right.
Your takeaway: Today's AI is still rough at the kind of reasoning sports commentators do, like picking which moments matter and explaining why a play worked. One caveat: these models haven't been fine-tuned or trained specifically for sports analysis, so the numbers could climb a lot with focused training.
The story: According to TechCrunch, Google signed a deal to pay SpaceX $920 million per month from October 2026 through June 2029 for access to roughly 110,000 NVIDIA GPUs at SpaceX data centers. The deal follows Anthropic's $1.25 billion per month agreement for compute at SpaceX's Colossus 1 facility in Memphis.
Your takeaway: Compute is the giant cluster of computer chips that train and run AI models, and it's what every AI company is racing to get more of right now. Google joining Anthropic at SpaceX shows that even the biggest tech firms are renting their AI muscle from Elon Musk's company, which is now pulling in over $2 billion a month from those two deals alone.
The story: According to The Verge, Meta's standalone AI app has been running a "For You" feed full of AI-generated clickbait prompts like "A royal butler finally settled the milk first debate." Tapping any prompt generated a full AI-written story, and Meta said it would pull the feature after The Verge started asking questions.
Your takeaway: This is a small window into how AI companies are tempted to fill their apps with AI-made high volume content to keep people scrolling. Meta pulled the feature only after a reporter caught it, but the feed had been live for months without anyone noticing.
TOOLS ON OUR RADAR
🎧 Mixxx Free and Open Source: A comprehensive digital DJ software that provides everything aspiring music performers need to create live audio mixes track tempos and blend songs seamlessly.
🔖 Wallabag Free and Open Source: A wonderful read later application that extracts web articles so you can save and read them comfortably on any device without internet ads.
💬 Etherpad Free and Open Source: A highly customizable text editor that allows multiple people to collaborate on documents simultaneously in real time directly through a web browser.
☁️ Owncloud Free and Open Source: A reliable file cloud platform that empowers you to securely host your personal documents synchronize media files and share folders entirely under your own control.
TRENDING
France to Test Its Own AI Battlefield System Against Palantir's - France will deploy Arcadia, its homegrown AI battlefield command system, during a NATO exercise in Poland from June 8-26. It's France's answer to Palantir's Maven Smart System, which NATO has been using since August 2025.
Making Claude a Chemist - Anthropic shared research showing Claude Opus 4.7 matches or beats dedicated chemistry software at reading NMR spectra, the readouts chemists use to figure out a molecule's structure.
New York Passes First-in-Nation Data Center Moratorium - The New York state legislature passed a one-year halt on new permits for large data centers, those needing 20 megawatts or more. If Governor Hochul signs it, New York becomes the first US state to put the brakes on the AI data center boom.
Google's Gemma 4 Gets a Memory Makeover - Google released Gemma 4 open-source models trained with quantization-aware training, which shrinks the memory needed to run them. The smallest version now fits in under 1GB, meaning capable AI can run on phones and laptops without losing much quality.
MIT Researchers Stress the Human Side of AI - MIT's Ethics of Computing Symposium brought together researchers from MIT, Google DeepMind, and OpenAI to talk about AI alignment and how students are using AI to skip the hard parts of learning rather than to support it.
An AI Agent Found 21 Zero-Day Bugs in FFmpeg for $1,000 - depthfirst's autonomous security AI scanned roughly 1.5 million lines of code in the FFmpeg video library and turned up 21 confirmed security holes, some hiding for more than 20 years. The entire run cost about $1,000.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
⚖️ Describe a sunk-cost situation. See what you've spent vs. what staying will cost — and the honest verdict on whether to walk away.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Walk-Away Calculator — sunk cost vs. forward cost analysis with a verdict. Persist to localStorage key 'walk_away_v1'.
Aesthetic: dark slate (#0e1117), subtle red and green radial gradients, Inter sans 800 for headings, Newsreader serif italic for body and labels, JetBrains Mono for tags. Green (#50c884) for walking/recommendations, red (#dc5060) for staying costs, amber (#dca050) for the sunk-cost trap. Strikethrough red sunk-cost number next to large amber forward-cost number. Color-coded verdict pill.
Form: situation textarea, sunk-cost textarea (money/time/energy invested), forward-cost textarea (what staying will cost next year), honest part textarea (the stuff under the surface), domain dropdown, duration dropdown.
System instructions to the model: act as a sunk-cost economist. Past costs are irrelevant to future decisions. Don't shame attachment, just refuse to let sunk costs count in the forward equation. Honest, not breezy. If staying is right, say so. Return raw JSON: sunk_value (visceral specific summary), sunk_note (italic line), forward_value (one more year's cost), forward_note, recommendation (one of WALK AWAY / STAY / MODIFY / SET A DEADLINE), confidence, reasoning (2-3 sentences using their specifics), honest_math (2-3 sentences forward-only logic), staying_costs array (4), walking_gains array (4), trap (2-3 sentences naming the specific sunk-cost trap), stay_strategies array (3 — concrete ways to make staying worth it if they choose to), deadline (specific date or trigger).
Render: big comparison card with strikethrough red sunk + amber forward. Green verdict card with big recommendation, confidence line, italic reasoning. White "honest math" card. Two-column grid with red staying-costs (− bullets) and green walking-gains (+ bullets). Amber-left-bordered "trap you're in" card. Numbered green-bordered "if you stay" strategies. Red-bordered deadline card at the bottom.What this does: Describe what you've been pouring yourself into (relationship, business, job, project, degree), what you've already invested, what staying another year will cost, and the brutal honest part only you see. Get back a side-by-side comparison of sunk-cost-gone vs. forward-cost-real, a clear verdict (WALK AWAY / STAY / MODIFY / SET A DEADLINE) with confidence and reasoning, the honest math walking through opportunity cost logic, two columns of what staying costs vs. what walking gives back, the specific sunk-cost trap you're in, three strategies if you do decide to stay, and a concrete deadline. Saves to localStorage with verdict color-coded.
What this looks like:
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Match dedicated chemistry software at reading molecular structures from NMR spectra
❌ Still Can't: Find evidence inside sports video clips to answer complex game questions, getting only 5% accuracy
✅ AI Can Now: Scan 1.5 million lines of code and find decades-old security bugs for around $1,000
❌ Still Can't: Reliably explain cause-and-effect in unfolding events from video, scoring only 40-50% on those tasks
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
A 2019 thriller about a New York cop investigating a mysterious affliction the press is calling False Memory Syndrome, where victims suddenly remember entire alternate lives they never lived, and a neuroscientist building a device that maps human memory in unprecedented detail. Bestseller in 2019 and somehow still under the radar of most newsletter readers.
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.
Claude is not just a chatbot anymore. Is your security team ready?
Claude.ai is one thing. Agentic workflows, MCP connections, ungoverned skills taking actions across your data? That's a different conversation — and most security teams aren't equipped for it.
Harmonic Security gives your CISO the visibility and controls to say yes confidently.