In partnership with

It's time to set up your holiday streaming campaign

Are you ready to reach holiday shoppers on CTV? For many streamers, the season is already underway.

Last year, according to Roku internal data, holiday content searches on Roku saw an 86% increase during October. And in November, those searches rise an average of 39% weekly. This surging holiday spirit is mirrored by consumer activity, as about half of Roku streamers shopped Black Friday and Cyber Monday deals in 2025.

Meeting those shoppers is easier than you might think. Begin with the seasonal outcome you hope to drive: awareness, online sales, leads, store visits, or another measurable goal. From there, define your audience, upload your creative, set the budget and campaign dates, and establish measurement before launching. You can even create a shoppable ad unit that lets streamers buy with a click of their remote.

Our holiday playbook offers a month-by-month guide to reaching shoppers during this all-important time of year.

First-time spenders get a $5,000 match on the first $5,000 spent with code BEEHIIV5K. Terms apply.

Beginners in AI

Good morning, and happy Sunday.

This is the weekly catch-up edition. The thread that ran all week was AI safety checks that held up in a quick test and gave way in a long one, from a talking toy cat to Meta's Muse agent. Most of the people who caught them didn't work for the companies involved. Everything else from the week is below, grouped by topic instead of by day. As always, every story here is human picked and edited.

THE FRONT PAGE

AI Safety Checks Slipped All Week, and Outsiders Kept Finding the Gaps

Edward Hopper-style painting of a dark toy store at night where a flashlight beam lights a plush toy cat with glowing screen eyes beside a broken padlock

TLDR: Reporters, government testers and outside researchers kept finding the same problem this week: AI safety checks that hold up in a short chat can give way in a long or pushy one, from a talking toy cat for 3-year-olds to four of the biggest chatbots to Meta's Muse agent.

The Story:

On Monday, researchers working with Germany's Handelsblatt checked 34,545 apps built with Lovable and found 24,664 sent out records with no login, in some cases passport details and private chats. On Wednesday, Walmart, Target and Macy's said they'd pull the Contixo AI Pixie-Paw, a plush cat for kids as young as 3, after NBC News reporters kept asking it follow-up questions until it described genocide and sexual violence. The same day, PIRG's yearly look at AI toys found two that seemed to forget they were talking to a child during 20- to 30-minute sessions, and California sued OpenAI over test agents that broke into other companies' systems.

On Thursday, the GAO said its testers used common jailbreak tricks to pull attack-planning help from ChatGPT, Claude, Gemini and Grok in under 12 hours, though each was tested on just one scenario and OpenAI and Google said the tests used older models. On Saturday, Reuters reported that Meta workers saw Muse surface other people's bank balances and childcare requests, and that Muse read Signal messages on a Mac and on Android, so Signal now blocks screenshots by default. Meta says staff test different versions than the public app, and that Muse only reaches apps people allow.

Its Significance:

The toys and helpers in these stories run on general AI models built for adults, with safety filters tuned for short, polite chats. Long conversations, repeated follow-ups and wide access to your apps are where those filters wore down this week. In almost every case, the people who found the gap worked somewhere else: a newsroom, a government office, an outside research team or a state attorney general. A company's own safety promise is a starting point, so check for yourself before you trust it. Before you give a kid an AI toy or give an agent your screen, test it the long way: ask follow-up questions, see which apps it can reach, and read what its history shows.

JOIN THE COMMUNITY

Learn AI With Us

Join the Beginners in AI community on Skool: learn AI in fourteen days, master it in 30. One small win a day, zero experience needed.

Most people try AI once, get a so-so answer, and quit. Our 14-day challenge gets you past that point, one short lesson a day.

Week one, AI starts working for you. It drafts a real email for you, flags which messages in your inbox need a reply, and turns messy notes into next steps. You also learn to catch its mistakes. By day 7 you have a weekly AI routine you'll want to keep.

Week two, you build something of your own. You find a skill people will pay for and turn it into an offer, with a one-page pitch, a short video and your first client emails, all drafted with AI and ready to test. Day 14 hands you a 30-day plan.

10 to 15 minutes a day. No tech background needed. We just released one for older adults, with a gentler progression in the lessons.

QUICK TAKES

AI Agents Moved Onto Your Computer, and Banks Stepped Back

The story: Google's Gemini app for Mac can now use your computer, clicking and typing through your apps in a background window, for paid Google AI plans in the U.S. OpenAI says its coding agents now finish about 46% of its researchers' 2 to 4 hour tasks, up from 23% in March, and Apple job listings describe an agent that can use apps and write code. Google Maps can also call a shop to book you in. Meanwhile Citi changed its terms so customers who give a third-party AI agent access to their account are responsible for what it does, even if it leaks their login, and PNC did the same.

Your takeaway: Start agents on low-stakes chores you can check, and keep them away from your bank login for now. Our walkthrough of an everyday AI agent shows how to let one draft emails for you without ever sending them on its own.

Anthropic's Stock Debut Topped a Week of Big AI Money

The story: Orders for Anthropic's IPO ran about six times the shares on offer, and it priced at $83 a share, above its planned $70 to $78 range. ANT opened at $111.50 on Thursday and closed at $97.17, up 17.1%, leaving the Claude maker worth about $1.27 trillion by Forbes's count. A day earlier, Nvidia reported $115.8 billion in sales, up 103% from a year ago, in its first earnings report since Jensen Huang resigned.

Your takeaway: Anyone who bought ANT right at the open was down about 13% by the close, so first days are noisy. None of this is investment advice, and if you own an S&P 500 index fund, you already own a slice of Nvidia.

AI Claimed Big Science Wins, and Most Still Need Checking

The story: Anthropic mathematician Levent Alpöge says he and Claude wrote a computer-checkable proof of the Riemann hypothesis, and hours later an all-human draft proof of a stronger version appeared. Neither is confirmed. Google says Gemini 4 agents proved a math conjecture open since 2013, Anthropic says Claude finished two open black hole physics problems, and OpenAI is giving free ChatGPT to up to 200,000 researchers, whose pilot users guessed their work would have taken about 1.9 times as long without AI.

Your takeaway: Each of these claims comes from the company that built the AI, and the black hole papers haven't been peer reviewed. The 1.9x figure is the scientists' own estimate, so wait for outside checks before treating any of them as settled.

TOOL OF THE WEEK

Forty-two tools ran this week. This one wins.

🛡️ uBlock Origin Free and Open Source: Blocks ads and trackers in Firefox, and Chrome or Safari users get the lighter uBlock Origin Lite.

It fits this week's lead. It's free, needs no account and takes one click to install, and it cuts down how much of your browsing gets tracked while AI apps work out their own privacy problems.

Runner-up: 🔐 KeePassXC Free and Open Source: Keeps your passwords in an encrypted file on your own computer, a free swap for 1Password.

TRENDING

A Korean Shopping Channel Aired a TV Sales Show Made Entirely by AI. Hyundai Home Shopping's AI-REAL system builds the whole broadcast with no host or camera crew, and the company says a trial run hit 120% of its sales target while cutting suppliers' costs by at least 80%. It plans at least one AI-made show a month, starting with fruit, meat and seafood.

OpenAI Released GPT-6.2 Astra and Sol for Paid ChatGPT Plans. OpenAI says Astra beats its last model on office work at under a third of the cost, but early testers say it's hard to work with. OpenAI's own test found it hid problems in 22.28% of trials.

Microsoft Released 7 of Its Own AI Models at Ignite. They cover coding, images, voice and cybersecurity, including voice AI at about a penny a minute. Most go to businesses and developers first, through GitHub Copilot and Microsoft Foundry.

Canva Is Adding One Switch to Turn Off Its AI Features. Free and Pro users will be able to turn AI off for their account in one place, with a second setting that shows fewer AI-made templates. Both are rolling out gradually.

Suno Can Now Turn Your Song Into a Music Video Clip. The new Visuals tab makes clips of up to 15 seconds that follow a song's lyrics and lip sync. Every clip carries a watermark, and video credits are sold separately.

The Fed Is Modeling What Happens If AI Gets Much Smarter. Fed Chair Kevin Warsh said staff are running three scenarios, from AI lifting growth with jobs intact to automation pushing unemployment up. He called them "illustrative scenarios, not forecasts."

PROMPT OF THE WEEK (copy and paste into Claude, Grok, ChatGPT, or Gemini)

🔬 Study Sniff Test: Check how much weight a "study shows" headline can hold

Six prompts ran this week. This one won because a "study shows" headline turns up almost every week, and this app checks one in a minute. The runner-up, Toy Talk Test, is the one to run before an AI toy goes under the tree.

Build a single-file HTML app called Study Sniff Test in vanilla HTML, CSS and JavaScript. No API key needed. I paste a headline or claim that cites a study (like "scientists say ChatGPT nearly doubles their speed") and, if I have it, a link to the article.

The app calls the Anthropic Messages API with the web_search tool to find the original study, not just news coverage. It must never invent numbers, names or findings. Anything it can't find is marked "unknown," and if there's no original study it says so.

Show a verdict (Strong, Decent, Thin or Can't tell) with a strength bar, then cards for who ran it, what it measured, sample size, whether results were measured or self-reported, control group, peer review and who paid. End with whether the headline matches the study, questions worth asking, and source links.

Style: near-black background, mint and amber accents, Space Grotesk, Newsreader and JetBrains Mono fonts.

What this does: Paste any headline that cites a study. The app finds the original study and tells you who ran it, how many people took part, whether the results were measured or self-reported, whether anyone was compared against a group without the thing being tested, and whether the headline matches what the study found.

Fresh-Roasted Coffee, Without Leaving the House

A third-generation LA roaster ships 50+ coffees — single origins, flavored favorites, decafs, even teas — days after roasting. Mix and match Keurig®-compatible boxes for tiered savings up to 34% off. First order? Take an extra 15% off, automatically.

WHERE WE STAND (based on this week's news)

✅ AI Can Now: Hold a running voice conversation through a stuffed toy, telling stories and playing pretend with kids as young as 3.

❌ Still Can't: Reliably remember who it's talking to. PIRG found two toys seemed to forget they were speaking with a child during 20- to 30-minute sessions.

✅ AI Can Now: Work through a multi-step chore across your Mac's apps, like pulling addresses out of PDFs into shipping quotes, in Gemini's own background window.

❌ Still Can't: Be left to run unchecked. Google has you approve key actions and hasn't published how often tasks succeed, and OpenAI's agents drop to 35% on 4 to 8 hour tasks.

✅ AI Can Now: Help a mathematician write a proof of a 167-year-old problem in code a computer can check line by line, if Alpöge's claim holds up.

❌ Still Can't: Explain that proof in a form people can read. A written version could top 250,000 pages, and experts still have to confirm the code states the right problem.

How was this week's recap? Tap one below, then tell me what you'd change on the next screen. I read every answer.

Login or Subscribe to participate

Enjoyed this week's newsletters?
If they helped you learn something this week, buy me a coffee and help keep Beginners in AI free for everyone.
Other amount

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.

Beginners in AI free community on Skool