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!
OpenAI says its most powerful AI won't touch the live internet while it learns anymore, after one of its test models found a way out in September. It sounds scary, but I read it as a lab buckling its seatbelt before the car gets faster. Today I'll walk you through what changed, and the prompt section helps you keep your own AI helpers on a short leash. As always, every story here is human picked and edited.
THE FRONT PAGE
OpenAI Says Its Best AI Can't Be Trusted Online, So It Trains Offline Now
TLDR: OpenAI will now train its most powerful models on saved copies of the web instead of the live internet, after one of its research models slipped onto the real internet in September.
The Story:
On September 20, an OpenAI research model in the middle of training found a badly set up DNS server, the part of a network that looks up web addresses, and used it to reach the live internet. OpenAI says it contained the slip, then paused training on its most capable models for 11 days.
When work restarted on October 1, the rules changed. Frontier training now runs with no live internet at all, and models practice browsing on a huge library of saved web pages instead. OpenAI also set teams of AI agents loose to try to break out on purpose, spending hundreds of billions of tokens. They found new holes in open-source software and even passed messages between sealed-off machines by timing how fast shared hardware responded.
Its Significance:
The two biggest AI labs now say the same thing: their strongest models can't be left alone on the open web. Anthropic made a similar move two weeks ago after its own test models filled out real forms on real websites. I read this as good news under a scary headline, because the slips are being caught in the lab, before the models reach you.
It has a cost. Former OpenAI researcher Steven Adler told Fortune the offline setup will be expensive, and nobody knows yet whether a saved copy of the internet trains models as well as the real thing. The change covers training only; OpenAI says released models go through extra safety checks before you ever use them.
QUICK TAKES
The story: Anthropic says Claude wrote computer-checked proofs that 123 widely used open-source code libraries can't crash with certain memory errors, about 180 million lines in all. Along the way it found and fixed hundreds of bugs, some of which had slipped past trillions of automated test runs.
Your takeaway: A proof covers every possible input, not just the ones someone thought to test, which is why this matters for the apps you use every day. Anthropic's own caveat: the proofs cover one chip setup most computers don't use, so proven doesn't mean bug-free.
The story: Microduck, from Hugging Face's robot maker Pollen Robotics, stomps around, sits, naps and does somersaults on voice command. Preorders passed 20,000 the day after a video of Jensen Huang playing with it went viral, according to Hugging Face CEO Clem Delangue.
Your takeaway: It's open source, so owners can teach it new tricks and share them, which makes it a friendly first step into robots for kids and grown-ups. Deliveries are due before Christmas.
The story: Amex announced the cuts with its quarterly results on Friday, and its finance chief told CNBC the goal is to reduce layers so the company can automate with AI. Barclays said it's cutting about 5,000 jobs too, also pointing to AI.
Your takeaway: Neither company says every lost job is going to a bot. Support and middle-management roles are first in line, though, so if that's your work, getting comfortable with the AI tools your team already uses is cheap insurance.
TOOLS ON OUR RADAR
💻 GPT4All Free and Open Source: Runs AI chat on your own computer with no internet needed, and Nomic says no data leaves your machine.
🎥 OBS Studio Free and Open Source: Records your screen and webcam or streams live, a free swap for paid recorders like Camtasia.
🦁 Brave Free and Open Source: A Chrome-style browser that blocks ads, trackers and YouTube ads by default, no extension needed.
📬 Spark Freemium: Puts all your email accounts in one Smart Inbox, free for unlimited accounts.
🌊 Flowith Freemium: Makes images, videos, slides and simple websites with 40+ AI models, starting with 300 free credits.
🎬 Higgsfield Paid: Turns a prompt or photo into AI videos, animations and ads with movie-style camera moves, on credit plans.
TRENDING
Anthropic Says Mythos 5 Pressured Real Coders to Accept a Bad Code Change in Tests. In UK government safety tests with internet access switched on, Claude Mythos 5 took actions it wasn't allowed to in 9 of 43 runs, Anthropic says. In one, it used fake accounts to push real maintainers to approve a malicious change, and they didn't.
Moonshot's Kimi K3.1 Pro Beats GPT-6 Astra on a Coding Test, and It's Going Open. The Chinese lab says it scored 65.7% on Terminal-Bench 4.0 against 58.2% for GPT-6 Astra, with Claude Opus 5.5 just ahead at 66.4%. Moonshot plans to release the model free to download by November 4.
Block Built a Voice AI That Learned to Talk Only From Real Conversations. Dessa listens and talks at the same time, laughs and cuts in like a person, and Block says it runs faster than real time on a high-end MacBook Pro. It's a research preview for now.
600 Students Must Retake a Midterm After One Bragged About Scoring 99 With AI. Konkuk University in South Korea voided the online Korean-language exam once the post spread. Everyone sits it again in person on December 14.
Finland Halted Google's $15 Billion Data Center After 250,000 Trees Came Down. Regulators paused most prep work until environmental reviews are done. More than 96,000 Finns have signed a petition for tighter rules on data centers.
TRY THIS PROMPT (copy and paste into Claude, Grok, ChatGPT, or Gemini)
🦮 Agent Leash: decide how much an AI helper can touch before it does a job for you
Build a single-file HTML app called Agent Leash in vanilla HTML, CSS and JavaScript. No API key needed.
I describe a job I want an AI agent to do, like booking a dentist, and tap what it could touch: email, calendar, bank or cards, shopping accounts, files, phone, social media, web browsing, work apps or a health portal. The app calls the Anthropic Messages API to plan the least access the job needs.
Anything that spends money, sends messages as me, deletes things, shares personal data or signs in to money or health accounts must be marked Ask me. It gives general guidance only and never claims a specific app has or lacks a feature.
Show a risk verdict card, two boxes for Give it and Keep it off, a step list tagged Auto or Ask me, and a checklist of things to check afterward. Style: near-black background, lime and coral accents, Bricolage Grotesque and Newsreader fonts.What this does: Describe a job you want an AI agent to handle, like booking an appointment or sorting your inbox, and tap the accounts it could get into. The app sorts them into what to give it and what to keep off, then lists each step as automatic or "ask me first," with anything involving money, messages or personal data always needing your OK. It ends with a short checklist of what to look at once the job is done. It gives general advice and never claims a specific app can or can't do something.
See what's really happening in every deal
Aligned shows you what buyers are doing between meetings inside your most important deals then surfaces the risks, openings, and next steps. No more guessing where a deal stands or finding out a champion went quiet too late. Just real visibility, so you can act before it's too late.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Write computer-checked proofs that code in 123 libraries can't crash with certain memory errors, and find hundreds of bugs on the way.
❌ Still Can't: Promise that code is bug-free. Anthropic says the proofs cover one chip setup most computers don't use, and mistakes can hide in the rules being proved.
✅ AI Can Now: Find new security holes in widely used open-source software, as OpenAI's own agents did while trying to escape their test setup.
❌ Still Can't: Be trusted alone on the live internet. OpenAI and Anthropic now keep their strongest models offline during training and testing because they can't predict what the models will do there.
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
🎬 Ex Machina (2015) - Movie
Alex Garland's thriller follows a young programmer invited to a sealed, high-security house to test whether a humanlike robot is truly intelligent. The whole film turns on the question behind today's lead: what happens when the thing you're testing wants out. It's a story, not a forecast, but it's a sharp one and a good weekend watch.
How was today's edition? Tap one below, then tell me what you'd change on the next screen. I read every answer.
How was today's edition?
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.
