In partnership with
How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads
For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.
LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.
The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.
Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV 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 happy Sunday.
This is the weekly catch-up edition. Everything worth knowing from Monday through Saturday is below, grouped by what it's about instead of what day it ran. If you missed a few mornings, start here.
THE FRONT PAGE
The Week We Found Out Where Your DeepSeek Chatbot Questions Go
TLDR: Three federal agencies named six Chinese AI companies for copying American models, Anthropic put hard numbers on it two days later, and by the weekend one of those companies had a model matching OpenAI's best design work for about a hundredth of the price.
The Story:
On September 8 the NSA, FBI, and CISA put out a joint advisory naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI. The agencies say those firms pulled billions of tokens out of US models since 2024, including Claude, GPT, Gemini, and Grok, using fake accounts and resellers they call transfer stations. Two days later Anthropic published a 154-page threat report with counts attached: it says Moonshot passed close to 300,000 customer requests to Claude in one 10-day window through 5,380 accounts it describes as fake, and that DeepSeek made more than 12 million attempts over 14 days in July. Anthropic says some of those requests carried sensitive details from individuals, large companies, and accounts tied to foreign governments. Then on Saturday, OpenDesign Arena scored 13 models on real design jobs, and DeepSeek's V4.1 Flash came within 98% of GPT-6 Astra while charging about 1.4% of the price. Moonshot, DeepSeek, and Xiaomi did not answer reporters, and China's Commerce Ministry has said earlier claims like these have no factual or legal basis.
Its Significance:
Strip out the geopolitics and this is a story about where words you typed ended up. If the reports hold up, someone opened Kimi or DeepSeek, asked a question, and read an answer that Claude wrote, with nothing on the screen saying so. Hold onto the fact that these are one company's claims about its competitors and the other side has not answered them. What nobody disputes is the price gap, and cheap capable models are good news for anyone paying a bill. So the question worth following over the next few months is whether that price comes from better engineering or from somebody else's training run, because the answer changes which of these companies is still standing in two years.
QUICK TAKES
Personal AI Agents Stopped Being a Demo
The story: Meta launched Muse on Tuesday, a personal agent that runs on its own cloud computer and can send emails, book trips, and buy things once you approve each step. A second agent called Sentinel sits on that same machine and checks everything before it reaches the internet. By Thursday, Muse was the No. 2 app in the US App Store on about 83,000 iOS downloads. OpenAI spent the same week going the other direction and stopped selling its $200 Pro plan because demand for its new Astra model was loading up the servers.
Your takeaway: Meta didn't invent this. OpenClaw started the whole category last year, xAI's Grok Bot was the first version of it from a major company, and Muse is Meta's take on the same idea with a billion-user app behind it. Try the free version before you pay for anything, and read the list of what you're connecting it to before you tap yes.
AI Did Science This Week That Other Scientists Could Check
The story: Two mathematicians released proofs on fluid equations tied to one of math's six million-dollar problems, using Claude for the math and software called Lean to verify every step. Insilico Medicine published trial data on rentosertib, a drug whose protein target and molecule were both picked by AI, and six separate aging clocks read the treated patients as three to six years younger. A Nature review found AI now proposes whole experiment layouts that break design rules physicists follow, and beats the human versions anyway.
Your takeaway: Two years ago none of this was on the table. The shape underneath all three is the same: the model finds the answer, and a human or a piece of verification software still has to confirm it means anything. That second half is what separates these from a chatbot sounding confident.
The Week's Loudest Warnings Came From Inside the Field
The story: OpenAI's chief scientist Jakub Pachocki published an essay saying no lab, his own included, has solved alignment and monitoring well enough to keep scaling at full speed. Terence Tao, who won the Fields Medal in 2006, warned that AI labs are burning through good unsolved math problems faster than mathematicians can find new ones. And PISA tested more than 760,000 fifteen-year-olds across 91 countries: teens who used AI daily to draft writing scored 481 in science against 509 for those who almost never did.
Your takeaway: That PISA number has a second half worth more than the headline. Students who used AI once or twice a week to help them learn scored highest of any group, and daily users did 13 points better when their classes made them pick apart AI answers. Frequency wasn't the problem.
TOOL OF THE WEEK
Twenty-four tools ran this week. This one wins.
🎙 Buzz Free and Open Source: Transcribe audio and video entirely offline using OpenAI's Whisper model, nothing leaves your computer, no account, no upload limits.
The front page this week is about questions going somewhere the person typing them didn't expect. In the same seven days, Google opened Dreambeans to every US account and OpenAI started testing a setting that reads your Gmail and Slack to copy your writing style. A transcriber that runs on your own machine and uploads nothing felt like the right pick.
Runner-up: Excalidraw Free and Open Source: Sketch diagrams, flowcharts, and wireframes in a hand-drawn style right in your browser, no sign-up, no watermark.
TRENDING
GPT-6 Astra played through the game Portal with nobody at the controls and the API bill came to $571.18. It read screenshots, worked out its own moves, and cleared every test chamber across 3,336 tool calls. The run took close to a full day because the game sat paused while the model thought.
Draymond Green says NBA front offices use AI and players barely touch it. The analytics teams are running it on the decisions that matter most: who gets traded, signed, and cut. That's how AI shows up in most workplaces, starting with the people handling the spreadsheets rather than the people doing the job.
DeepMind mapped every possible single-letter change in human DNA and put it online free. That's 9 billion changes pre-calculated into a 1-petabyte database, sorted by one score. Researchers ran it against 54,000 UK Biobank records and turned up 22% more links between DNA and traits, with no coding needed.
Mistral raised 3 billion euros to build Europe's own AI, at a valuation over 21 billion euros, led by Samsung alongside an EU-backed fund. Mistral says it's the largest equity round a European tech company has ever closed. The servers will still run on Nvidia chips designed in America.
NASA and IBM put a moon-mapping AI online for anyone to download. It trained on more than 30 data layers from nine instruments across four missions, and NASA says it cuts errors on likely ice spots by up to 22% and maps craters about 19% better. The model sits on Hugging Face and the code sits on GitHub.
A therapy chatbot cut symptom scores in a controlled trial. Dartmouth clinicians tested Therabot with 210 adults, half getting four weeks of access and half on a waitlist. The chatbot group showed larger drops at both the four and eight week checks. One trial isn't proof, but the size of the drop is worth watching.
PROMPT OF THE WEEK (copy and paste into Claude, ChatGPT, or Gemini)
🔧 Describe the sound your car's making. Get the real commonly-reported causes, ranked, with a read on how urgent it is.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Car Trouble Sound Diagnostic — describe a car sound, get real commonly-reported causes via live web search, ranked by likelihood and urgency. No API key needed.
Aesthetic: dark charcoal (#14161a), steel blue (#5a8cb8) primary with a blue glow top-left, caution amber (#e0a840) secondary with a glow bottom-right. Space Grotesk for headings/cause names, Inter for body, JetBrains Mono for labels/tags.
Form: sound description text input, a "when it happens" dropdown (braking/turning/starting/accelerating/idling/at speed/not sure).
Technical: call the Anthropic Messages API WITH tools:[{type:'web_search_20250305', name:'web_search'}] (works automatically), system prompt as an automotive information researcher (explicitly NOT a mechanic performing diagnosis) who finds real commonly-cited causes from reputable automotive/repair sources via search, ranks most-to-least commonly reported, and assigns an honest urgency level per cause using standard categories ("Stop driving, unsafe" / "Get it checked soon" / "Safe to drive short term") — explicitly required to never give a definitive diagnosis and to always frame this as general information requiring an actual mechanic's physical inspection. Return raw JSON: sound_summary, general_note (1-2 sentences on the system this points to), possible_causes (3-5: cause, likelihood, urgency, explanation), safety_note (2-3 sentences reiterating this isn't a diagnosis, plus specific accompanying warning signs — burning smell, dashboard lights, loss of braking — that mean stop driving immediately).
Render: an overview card with the restated sound description and general note. A ranked list of cause cards, left-border and badge color-coded by urgency (red=stop, amber=soon, green=safe), each showing likelihood tag, urgency badge, and explanation. A prominent red "not a diagnosis" safety card at the bottom with the specific warning-sign guidance.What this does: Describe what you're hearing and when it happens, and it searches the live web for real, commonly-cited causes from reputable automotive sources, ranked from most to least commonly reported. Each cause comes with a color-coded urgency level, stop driving, get it checked soon, or safe short term, so you know what needs attention now versus what can wait for a scheduled visit. Five other prompts ran this week, and this is the one you'll open more than once a year. Cars make noises constantly, and the alternative is walking into a shop with no idea what you're hearing.
The Practical AI Agent series kept rolling on YouTube this week with Parts 2, 3, and 4. It's a follow-along, so you finish with your own team of agents that work on their own and in teams.
Part 1 is where to start:
WHERE WE STAND (based on this week's news)
✅ AI Can Now: Play a 3D puzzle game start to finish, reading only screenshots and choosing its own inputs.
❌ Still Can't: Do it at a price that makes sense. That run cost $571.18, took 3,336 tool calls, and ate close to a full day.
✅ AI Can Now: Close proofs on equations that specialists expected to take decades, with software checking every step instead of colleagues taking them on trust.
❌ Still Can't: Write them up. Buckmaster says the English explanation the models produced was barely readable, so he and Alpoge rebuilt the steps by hand for other researchers.
✅ AI Can Now: Find brain aneurysms on CT scans that trained radiologists missed, 55 of them across 3,856 exams.
❌ Still Can't: Be trusted on its own verdict. When the AI called an aneurysm it was right 78.2% of the time. When a radiologist called one, 92.7%.
PICK OF THE WEEK
Idoru by William Gibson - Book
A 1996 novel about a world-famous rock star who announces plans to marry an idoru, a fully synthetic AI performer generated from massive datasets rather than born, and the two very different people drawn into finding out what that marriage would mean. Gibson wrote it nearly three decades before AI-generated pop stars became a real commercial category. It pairs well with a week where Universal Music signed a deal to license its catalog to an AI remix app and Suno retrained its models on licensed songs.
I'm still rolling out the community feature and looking for people to test it. If you want an invite, reply to this email and I'll get you in.
Thank you for reading. We're all beginners in something. 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.