In partnership with
Blu Dot surpasses 2,000% ROAS with self-serve CTV ads
Home furniture brand Blu Dot blew up on CTV with help from Roku Ads Manager. Here’s how:
After a test campaign reached 211,000 households and achieved 1,010% ROAS, the brand went all in to promote its annual sales event. It removed age and income constraints to expand reach and shifted budget to custom audiences and retargeting, where intent was strongest.
The results speak for themselves. As Blu Dot increased their investment by 10x, ROAS jumped to 2,308% and more page-view conversions surpassed 50,000.
“For CTV campaigns, Roku has been a top performer,” said Claire Folkestad, Paid Media Strategist, Blu Dot. “Comping to our other platforms, we have seen really strong ROAS… and highly efficient CPMs, lower than any other CTV partner we've worked with.”
Using Roku Ads Manager, the campaign moved from a pilot to a permanent performance engine for the brand.
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
Meta Made a 30 Billion Parameter AI Free to Download, and It Runs on One Graphics Card
TLDR: Meta released a free AI model that runs on a regular laptop, and giving it away costs Meta almost nothing while making it harder for OpenAI and Anthropic to charge for theirs.
The Story:
Meta released Muse Glimmer on Monday, a 30 billion parameter model built to run on a Mac or PC with a single consumer graphics card. Meta compressed it down to roughly 17 gigabytes from the more than 55 a full version would need, so it fits on cards with 24 or 32 gigabytes of memory. It's free to download on Hugging Face under an Apache 2.0 license. Glimmer is a smaller copy of Muse Spark 1.2, trained by having the big model teach the small one, and it's tuned for tasks like managing a schedule, sorting files, and writing code on your own machine.
Mark Zuckerberg also said Meta will open the weights for Muse Spark 1.2 itself, which launched earlier this year as a closed model. He published a 14 page essay alongside the release arguing US rules put American open-source developers behind Chinese labs like DeepSeek, Alibaba, and Moonshot, which have led open-weight AI for a while now.
Its Significance:
Handing this out costs Meta close to nothing. Meta doesn't sell AI subscriptions the way OpenAI and Anthropic do, since its money comes from ads. So every person who downloads a free model instead of paying $20 a month takes revenue from a competitor, not from Meta.
Meta called Llama open source back in 2023, but the license blocked any company with more than 700 million monthly users and banned people from using Llama's output to train rival models. The Open Source Initiative said flat out that it didn't count. Apache 2.0 has none of those strings, so this release is more open than the last one, not less.
The potential impact on competition in the market is positive. If a model this capable runs on hardware you already own, with no cloud bill and no data leaving your desk, other labs have a reason to build small models too.
QUICK TAKES
The story: Torvalds announced the seventh release candidate for Linux 7.2 and said it's stuffed with fixes, most of them turned up by AI review tools. He wrote that he can't say he's thrilled about the size, but called it the new normal.
Your takeaway: Linux runs most of the world's servers, so what lands in the kernel reaches nearly everything you touch online. The AI tools are finding real problems, including a memory bug that sat in the code for eight years. The bottleneck is that a human still has to read every patch, and Torvalds has spent months telling people to stop sending duplicate reports.
The story: More than a thousand court filings have now been caught citing cases the AI made up, complete with real-looking names, courts, and years. Penalties started at $500 fines in 2023 and have climbed to $15,000 per lawyer, plus the first US bar suspension that pulled someone from practice entirely.
Your takeaway: These are people whose whole job is checking sources, and they still sent invented cases to a judge. In one Nebraska divorce appeal, 57 of 63 citations were bad. If you use AI for anything with a source in it, open the link yourself before it goes anywhere.
The story: OpenAI announced ten new results in math and computer science from Astra, a model it hasn't released, covering geometry, cryptography, and coding theory. A Monash University mathematician wrote that a colleague used an OpenAI model and a computing cluster to find an object her team had chased for over two years, and it took 43 hours.
Your takeaway: Math is one of the few places you can check an AI's answer all the way down, so it's an honest look at what these tools do when bluffing won't work. Mathematicians are split hard. Thousands signed the Leiden Declaration saying AI should help humans rather than replace them, and other researchers refuse to let the tools near their work at all.
TOOLS ON OUR RADAR
🌌 Stellarium Free and Open Source: A breathtaking virtual planetarium that shows exactly what you see when you look up at the stars from anywhere on Earth at any given time.
📑 Zotero Free and Open Source: A fantastic personal research assistant that helps you gather organize and cite your sources perfectly when writing essays or academic papers.
📰 Scribus Free and Open Source: A professional page layout program that allows you to design beautiful digital magazines printable brochures and colorful interactive documents.
🍍 HandBrake Free and Open Source: A highly reliable video converter that shrinks massive movie files into smaller formats so they play perfectly on your mobile phone or tablet.
TRENDING
Brad Pitt Says AI Could Help More 'Mid-Budget' Movies Get Made — Pitt told Esquire that AI used as a tool will help $40 to $90 million films get made, mostly by cutting the cost of visual effects. He's far less sure about AI replacing actors, and mentioned he's had every body scan of himself destroyed for the past 20 years, a decision he now regrets.
Tricking AI in the job hunt — Duke researchers went through 200,000 real resumes submitted to hiring platform hireEZ and found about 1% carried hidden text aimed at the AI screener, invisible to any human reading the page. The rate jumped sevenfold between July 2024 and November 2025, spread by TikTok tutorials and free templates.
Apple reportedly tests CXMT memory — Apple is testing memory chips from China's CXMT for iPhones and MacBooks, per the Wall Street Journal, because AI data centers have swallowed the memory supply and pushed prices up. Two things are in the way: US rules bar Apple from ordering chips built to its own specs there, and CXMT has refused to quote anything below what Samsung charges.
AI Tests Lawyers' Will to Keep Training the Models in Our Minds — A legal commentary argues that the model worth worrying about isn't the one in the software, it's the one in your own head, and AI makes it tempting to skip the work of building it. The examples are specific: an associate passing along a clean diligence summary that gets a change-of-control clause wrong, or a lawyer forwarding a regulatory overview that leaves out the one state that matters. The author pushes back on the idea that AI's upside and downside come as one package, saying you can take the speed and still do the reading.
Why do we labor when reading some words but not others? AI offers clues — Researchers at NYU and UMass Amherst tracked the eyes of 368 readers and compared them against more than 400 AI language models, and the models nailed the first step of reading: how fast you recognize a word. They fell apart on the second step, the part where you fit that word into the sentence and have to go back and reread. About 20% of the eye movements people make while reading go backward, and no model could say when or why.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
💰 Describe what you sell and your current prices. Get advice on what you should charge
Build a single-file HTML app with vanilla HTML/CSS/JS. Pricing Therapist — honest pricing analysis. Persist to localStorage key 'pricing_therapist_v1'.
Aesthetic: warm cream (#f5f1e8), black typography, Cormorant Garamond serif for headings and body (300/400 weight), Inter for UI, Roboto Mono for labels. Burnt amber (#b87830) accent. Editorial feel with section dividers (§ 1 / § 2).
Form: offering textarea, current price, format dropdown (hourly/project/retainer/etc), years dropdown, demand dropdown, market context input.
System instructions to the model: act as a pricing therapist. Be direct, not flattering. Most people are underpricing 30-100%. Use real numbers. Demand is the biggest signal. Return raw JSON: current_price, recommended_price, delta_pct, verdict (honest 1-2 sentences), why_underpricing (2-3 sentences naming the pattern from their description), dysfunctions array (2-3 items: named dysfunction like "Time-for-Money Trap" or "Pleaser Pricing" + 1-sentence explanation of how it shows up), tiers array (Floor/Target/Premium with price, who pays it, and which is recommended), pushback_script (1-2 sentences calm not defensive), permission (one honest sentence).
Render: a black verdict card with strikethrough current price next to large amber recommended price + delta %, italic verdict line. White reason card. Pricing dysfunctions list with amber left-border. Three-column tier grid (Floor/Target/Premium) with the recommended tier highlighted in amber. Italic script card with amber-border. Amber-bordered centered permission card. Archive with recommended price in metadata.What this does: Describe your offering, your current price, your demand, and your market. Get an honest verdict comparing what you charge vs. what you should charge, the percentage gap, the real reasons you're underpricing (named as specific dysfunctions), a recommended Floor/Target/Premium price structure, a script for when clients push back, and one sentence of honest permission. Saves to local storage.
What this looks like:
The first way to trade directly inside Claude and ChatGPT
Superintelligence used to be locked inside billion-dollar quant firms whose algorithms quietly took advantage of everyone else.
Co-Invest puts it right in your chat window. Analyze markets, manage risk, and execute trades, all inside Claude and ChatGPT.
The institutions built the game, Co-Invest gives you a way to beat them.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Run a 30 billion parameter model that manages files and writes code on one consumer graphics card, after being squeezed from 55GB down to about 17GB.
❌ Still Can't: Do that at full quality. The uncompressed version needs more memory than any single consumer card has, so the laptop version is a trimmed one.
✅ AI Can Now: Read through the Linux kernel and surface real defects humans missed, including a memory bug that had been sitting there for eight years.
❌ Still Can't: Recognize when it's reporting a bug someone else already reported, which is why Torvalds keeps getting the same fix submitted by different people.
RECOMMENDED LISTENING/READING/WATCHING
Terminal Boredom by Izumi Suzuki - Book
A collection of short stories by a Japanese counterculture writer, actress, and model who died in 1986 and whose work only appeared in English for the first time in 2021. The stories cover women in societies where men have been segregated, people who watch television so compulsively that reality stops registering, and relationships mediated by drugs, screens, and disaffection. Suzuki wrote in the 70s and 80s and predicted the emotional texture of a life lived through devices with startling accuracy.
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.
