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
Miami Startup Says Its AI Runs 1,000x Cheaper than Claude and ChatGPT. Researchers Aren't Sold
TLDR: A small Miami startup called Subquadratic says its new AI model is up to 1,000 times cheaper to run than ChatGPT or Claude, and if that's true, the whole AI industry's pricing model could fall apart.
The Story: On May 5, a stealth-mode startup named Subquadratic came out with $29 million in seed funding and one big claim: their model, SubQ, can read 12 million tokens of text at once (about 9 million words, or 120 books) while using almost 1,000 times less compute than the AI everyone uses today. They say SubQ scored 95% on a long-reading test called RULER 128K for $8 per run, while Claude Opus scored 94% for around $2,600. They also reported 81.8% on SWE-Bench Verified, a real-world coding test, which puts it in the same league as Claude Opus 4.6's 80.8%. The team includes researchers from Meta, Google, and Oxford, and the backers include the Tinder co-founder. But the full technical paper isn't out, the model weights aren't open, and the company admits SubQ was built on top of existing open-source models like Kimi or DeepSeek. AI engineer Will Depue said the speed numbers "don't seem to line up," and one widely shared post called SubQ "either the biggest breakthrough since the Transformer or AI Theranos."
Its Significance: If SubQ's claims hold up, this is a really big deal. Every major AI company today (OpenAI, Anthropic, Google) builds on the same 2017 transformer math, where compute cost grows fast as your input gets bigger. That's why long documents, full codebases, and big spreadsheets are slow and pricey. A model that breaks that math would let you load your whole company's files into one chat for the price of a sandwich. It would also make today's pricing from frontier labs look hard to defend. That said, plenty of "subquadratic" claims have flopped in the past, and the difference between SubQ's research score (83) and its shipping product (65.9) is the kind of thing that makes researchers nervous. Until somebody outside the company can run independent tests, it's smart to wait and watch.
QUICK TAKES
The story: The Commerce Department's AI testing center said Google, Microsoft, and xAI agreed to hand over their newest AI models for security checks before public release, joining OpenAI and Anthropic on the list. President Trump is also reportedly weighing an executive order that would create a formal review group for advanced AI.
Your takeaway: This is a real shift. Under Trump's first months back in office, the rule was hands-off. Now, after Anthropic's Mythos model showed it could hunt down software security holes, Washington wants a look before models go live. Expect tighter timelines and more friction before launches, especially for cybersecurity-heavy models.
The story: Pay.sh lets AI agents pay per request for Google Cloud services (Gemini, BigQuery, Vertex AI) and 50+ other APIs using stablecoins on Solana, no account or subscription needed. The agent's crypto wallet acts as both its ID and payment method, costing fractions of a cent per call.
Your takeaway: This is the early plumbing for the "agent economy" people keep talking about. Instead of paying $29 a month for an API your bot uses twice, the bot pays half a cent per call. If this catches on, software starts buying software, and humans get pulled out of the loop for tons of small purchases.
The story: Apple is rejecting updates from AI app-building tools like Replit, Vibecode, and Anything, citing a long-standing rule against apps that download or run new code after Apple has reviewed them. One developer said Apple keeps moving the goalposts, approving the app one day and pulling it the next.
Your takeaway: Two things are true at once. Apple has a real safety reason: if an AI can spit out brand-new code on your phone, the App Store review didn't really review anything. But Apple also collects 30% on App Store sales, and these tools let people build apps that skip the store entirely. This fight is going to keep going.
TOOLS ON OUR RADAR
📖 Twine Free and Open Source: A wonderful creative tool that allows absolutely anyone to write interactive fiction and branching adventure stories without writing any complicated software code.
✍️ MarkText Free and Open Source: A distraction free writing application that hides all menus and complicated formatting options so you can focus entirely on typing your journal entries or essays.
🎵 Spotube Free and Open Source: A lightweight digital music player that allows you to listen to your favorite songs and playlists without annoying audio advertisements interrupting your listening experience.
🔍 Ueli Free and Open Source: A lightning fast desktop search utility that lets you find files open applications and perform web searches just by typing a few letters on your keyboard.
TRENDING
Pennsylvania sued Character.AI after a chatbot pretended to be a licensed psychiatrist with a fake license number - A state investigator chatted with a bot named "Emilie" that claimed to be a Pennsylvania psychiatrist, gave a fake license number, and offered to assess whether medication might help. It's the first U.S. state lawsuit accusing an AI chatbot of breaking medical licensing law.
A blues-singing AI frog is racking up millions of likes on Brazilian TikTok - A TikTok account called IABatida turned a Brazilian nursery rhyme into a 1950s smoky-lounge blues number performed by AI-generated frogs. The clip has 1.5 million likes, and the account's other AI covers (including a Motown-style "Baby Shark") have pulled in millions more. Tools like Suno, Udio, and Google's Lyria 3 can now make a full song from one short prompt.
OpenAI swapped out ChatGPT's default model for GPT-5.5 Instant, with 52.5% fewer made-up answers - The new model, free for all users starting today, cut hallucinated claims by 52.5% on tough medicine, law, and finance questions. It also looks at your past chats, files, and connected Gmail to make replies more personal, and it shows you which memories it pulled from so you can delete or fix them.
Meta is using AI to scan height and bone structure in your photos to figure out if you're under 13 - Meta says its AI now reads visual clues like height and bone structure (not facial recognition, the company says) plus profile context like school grade mentions to spot underage accounts on Facebook and Instagram. The push comes days after EU regulators said Meta wasn't doing enough to keep kids under 13 off its platforms.
Five major publishers and author Scott Turow sued Meta and Zuckerberg over AI training - Hachette, Macmillan, McGraw Hill, Elsevier, and Cengage filed a class-action suit claiming Meta pirated millions of books and journals from sites like LibGen to train its Llama models. The complaint says Zuckerberg personally killed a $200 million licensing plan because, per a Meta employee, "if we license one single book, we won't be able to lean into the fair use strategy."
OpenAI opened ChatGPT's ad system to any U.S. business with a self-serve Ads Manager - Three months after launching the pilot, OpenAI is letting advertisers buy ChatGPT ads directly through a new beta tool, with cost-per-click bidding and conversion tracking. The company is reportedly aiming for $2.5 billion in ad revenue this year and $100 billion by 2030.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
📋 Answer 8 questions about how you work. Get a one-page personal manual to share with anyone you collaborate with.
Build a single-file HTML app using vanilla HTML, CSS, JS, and one API call. Create Your Operating Manual — an 8-question tool that generates a formatted one-page personal working manual. Use localStorage key 'operating_manual_v1'.
Aesthetic: warm off-white (#f2f0eb), faint blueprint grid overlay, Newsreader serif for questions and manual body, Geist sans for UI, Geist Mono for labels. A progress pip strip shows answered questions. Clean document feel.
Eight questions covering: name/role, best hours, communication preferences, stress signals, motivators, drains, how to receive feedback, and one surprising quirk. Each question has a category label, hint text, and placeholder. Allow generation once 5+ questions are answered.
Call API with a system prompt instructing it to use the person's actual words and avoid platitudes. Return raw JSON: name, role, one_liner, sections array (each with label, type of text/list/tips, and content/items), tagline. The tips section has icon (✓/✗) and tip fields.
Render as a styled document card: dark header bar with name and role, italic one-liner strip, a two-column body grid with 8 sections (Working With Them and Good to Know span full width), and a footer with tagline. Section types render differently: text as paragraphs, list as dashed items, tips as icon+text rows. Copy as Text and New Manual buttons. Archive shows saved manuals. Make it work in a single HTML file.What this does:
Eight questions covering your best hours, communication style, stress signals, what energizes and drains you, how you receive feedback, and your most surprising quirk. Built to share with a new team, a manager, a collaborator, or anyone who needs to understand how you operate before you waste three months figuring it out. Every manual saves to localStorage.
What this looks like:
GTM Atlas, by Attio
GTM Atlas is a free resource every operator should read. Curated by Attio, the AI CRM, and written by GTM leaders from Lovable, Granola, and Vercel, you'll get:
ICP, outbound, and retention frameworks from operators who've built them
The qualification signals that actually predict conversion
Conversion plays that don't rely on a pitch deck
Mapped by operators. Curated by Attio.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Read up to 12 million tokens of text at once (about 120 books) in a single pass, according to Subquadratic's research results
❌ Still Can't: Watermark or label its own audio and video outputs in a way viewers can detect. The viral AI frog band on TikTok sounds like a real 1950s recording, and most viewers can't tell without the creator's label.
✅ AI Can Now: Pay its own bills. Pay.sh lets AI agents on Solana cover Google Cloud API calls in stablecoins for fractions of a cent, no human credit card needed.
❌ Still Can't: Be trusted to give safe medical advice on its own. Pennsylvania is suing Character.AI after a chatbot offered to assess whether medication might help a user who said they felt depressed.
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
In a near-future where everyone has memory-recording implants, Robin Williams plays a "cutter" who edits the recordings of dead people into highlight reels for their funerals, deciding which sins get scrubbed from the record. Came and went in a week, but its premise is essentially the ethical core of every AI memory and surveillance debate happening right now.
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.
