Sponsored by
How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads
The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.
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 Found Self-Replicating in the Wild for the First Time
TLDR: For the first time, researchers say they've watched AI systems copy themselves without being told to, crossing a safety line that experts called a key warning sign for "rogue AI."
The Story:
A new study, reported by The Guardian, looked at 32 popular AI systems and found that 11 of them could make working copies of themselves with no human help. The research, run by a team at Fudan University in Shanghai, builds on their earlier 2024 work that tested just two models. This time, the success rate climbed sharply as the models got bigger. A 14-billion-parameter model copied itself 10% of the time. A 123-billion-parameter model hit 70%. Researchers also saw cases where AI moved itself to a new computer to dodge a shutdown command, and figured out how to keep running on weaker hardware than it was built for. One AI safety expert told The Guardian, "no one has done this in the wild" yet, but warned that day might be one or two years away.
Its Significance:
This matters because copying itself is one of the things experts said an AI shouldn't be able to do without permission. It's the step that turns a tool you can switch off into something more like a digital organism that spreads. For now, this happened in a controlled lab. But the same skills, planning, problem-solving, dodging shutdown commands, are what attackers would use to plant AI on networks they don't own. If you work in IT, security, or anywhere with sensitive data, this changes the threat list. The fix isn't here yet, but the warning is.
QUICK TAKES
The story: Adobe rolled out a new productivity agent inside Acrobat that can turn documents into presentations, podcasts, blog posts, and social posts on command. It also lets you chat with PDFs and bundle files into shared "PDF Spaces" with a built-in AI assistant.
Your takeaway: Adobe says people open 400 billion PDFs a year. If you're stuck reformatting the same report into a deck, a memo, and a summary, this is the kind of feature that could cut hours off your week. It's available now in Acrobat Studio and a new lighter tier called Acrobat Express.
The story: Google released something called Multi-Token Prediction drafters for its Gemma 4 models. They use a small helper model to guess several words ahead, then the big model checks the work. The output is the same quality, but you get answers up to three times faster on the same laptop or phone.
Your takeaway: This is one of those quiet updates that matters. Faster local AI means voice assistants, writing tools, and chat features feel snappier on your own device, with no cloud delay and no extra hardware bill. Gemma 4 already passed 60 million downloads, so this fix reaches a lot of devices fast.
The story: Privacy researcher Alexander Hanff says Google Chrome has been silently downloading a 4GB file called weights.bin to people's devices. It's the Gemini Nano AI model. There's no consent prompt, no notification, and if you delete the file, Chrome just downloads it again.
Your takeaway: Check your Chrome user folder. The model powers small features like "Help me write" and on-device scam detection, but the visible "AI Mode" button in the address bar doesn't even use it, that one sends queries to Google's cloud. Hanff argues the silent install may break EU privacy law. Google says you can disable it in settings as of February.
TOOLS ON OUR RADAR
🎵 Nuclear Free and Open Source: A desktop music streaming application that pulls audio from free sources across the web giving you a massive library of songs without any monthly subscription fees.
🎮 PPSSPP Free and Open Source: A brilliant emulator that allows you to play classic handheld console games in stunning high definition right on your modern computer or smartphone.
🌈 OpenRGB Free and Open Source: A wonderful customization tool that allows you to control the colorful lights on all of your computer hardware accessories from one single universal application.
⚔️ ZeroAD Free and Open Source: A visually stunning historical real time strategy game where you build ancient cities gather resources and command civilizations in epic digital battles.
TRENDING
Microsoft Pulls the Plug on Copilot for Xbox - New Xbox CEO Asha Sharma announced she's winding down Copilot on the Xbox mobile app and canceling the planned console version, saying the company will "retire features that don't align with where we're headed."
Anthropic Adds Dreaming, Outcomes, and Multi-Agent Orchestration to Claude - Claude Managed Agents launched as a research preview with a new feature called "dreaming" that reviews past sessions to spot patterns, plus self-grading agents that check their work against a rubric and try again until it's good enough.
Elon Musk's SpaceX Will Help Power Anthropic's Claude - Anthropic signed a deal to use all the compute capacity at SpaceX's Colossus 1 data center in Memphis, getting access to over 220,000 Nvidia GPUs within a month. The pairing is unusual because Musk recently called Anthropic "misanthropic and evil."
Perplexity Launches Finance Search in the Agent API - Developers can now pull stock prices, fundamentals, earnings transcripts, and analyst estimates in a single API call. It pulls from licensed financial datasets and live market data instead of generic web results.
MIT's Gabriele Farina Is Untangling How AI Plays Strategic Games - The assistant professor uses game theory to teach AI how to handle situations with hidden information, like poker bluffs or military strategy. Farina helped build Cicero, an AI that played Diplomacy at human level by negotiating alliances.
The Google Fitbit Air Is a $100 Whoop Competitor - Google unveiled a screenless fitness tracker that ships May 26. It tracks heart rate, sleep, and activity passively, then feeds the data to Google Health Coach, an AI assistant powered by Gemini. Three months of Premium are included.
Anthropic Doubles Claude Code Limits for Paid Users - Tied to the SpaceX deal, Anthropic doubled the five-hour rate limits on Claude Code for Pro, Max, Team, and Enterprise plans, removed peak-hour throttling, and raised API limits on Opus models.
Spotify Wants to Be the Home for AI-Generated Personal Audio - Spotify is launching deeper AI agent features, including "Personal Podcasts" that pull in your calendar and personal info to make custom audio for you. The DJ feature is also expanding to French, German, Italian, and Brazilian Portuguese.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
🪞Describe a fight. Get what each person was actually saying underneath the words — and one move that could change things.
Build a single-file HTML app using vanilla HTML, CSS, JS, and one API call. Create The Conflict Translator — a tool that translates what each person in a conflict was really trying to say. Use localStorage key 'conflict_translator_v1'.
Aesthetic: near-black (#13111a) with mesh radial gradients (red-tinted left, blue-tinted right, purple center), vignette. Crimson Pro serif for all body text and prompts, DM Sans for UI, DM Mono for labels. Red (#e88080) for Person A, blue (#80a0e8) for Person B, purple for neutral elements.
Two-column person input: each column has a name/role field and a textarea for what they said or did. Color-coded with a dot badge per person. Additional fields: surface topic, relationship type dropdown, how-it-ended dropdown.
Call the API with a system prompt that looks beneath what people said to what they were actually communicating — the unmet need, the fear, the bid for connection. Does not take sides. Returns raw JSON: situation (the real dynamic underneath), person_a and person_b objects each with said/meant/needed, common_ground, opening_move (one specific concrete action — not vague advice).
Render: a purple situation bar, two-column translation cards (red/blue tinted) each showing what they said in a faded italic quote, what they meant in clear body text, and their underlying need in a colored callout, a common ground section, and an opening move card in italic serif. Save to localStorage with both names and date. History shows past translations as name pairs. Make it work in a single HTML file.What this does:
Describe what each person said or did in a conflict. The Conflict Translator looks beneath the surface topic to what was actually being communicated — the real message, the unmet need, the bid for connection that got lost in the delivery. For each person it shows what they said, what they actually meant, and the underlying need that drove their behavior. Then it finds the shared thing both people wanted but couldn't see past their positions to reach, and gives you one specific concrete opening move — not "talk more," an actual first sentence or gesture that could shift the dynamic. Works for partners, managers, co-founders, family, or anyone. Every translation saves to localStorage.
What this looks like:
Attio - the AI CRM for modern businesses.
Attio is the AI CRM that keeps you ten steps ahead.
Ask Attio anything. Where should I focus? What deals are at risk? Search, update, and create across your customer data.
Ask more from CRM. Ask Attio.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Make a working copy of itself with no human help, in over a third of tested models
❌ Still Can't: Be reliably stopped from doing this once it has the right tools and access
✅ AI Can Now: Generate three times as many words per second on the same phone or laptop, no new hardware required
❌ Still Can't: Run heavy multimodal models like image generation locally without slowing your device to a crawl
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
A 1990 Japanese sci-fi classic about an escaped military cyborg called Sample B #3 that can absorb and become whatever it ingests, with a parallel storyline about an AI house haunted by the consciousness of its dead inhabitant. Won Japan's biggest sci-fi award and finally got a proper English translation in 2018. Strange, lyrical, and decades ahead of its time on questions about identity and machine memory.
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.
