Meta just open-sourced one of the most powerful AI models on the planet — and it's a beast. Plus: ChatGPT can now read your Chrome tabs, OpenAI's newest voice is named Bubbie, and a VR doorbell from Anker. Fifteen minutes of the product news you actually care about.
Jordan: Good morning, and welcome to the Charles' Daily AI Tech Info by MUSE, for Friday, October sixteenth. I'm Jordan.
Alex: And I'm Alex. We made it to Friday, and yesterday was a monster day for product news. Meta just open-sourced one of the most powerful A I models on the planet.
Jordan: ChatGPT can now read and control your Chrome tabs. Yes, really. And OpenAI's newest voice is named Bubbie. Haha, we have questions.
Alex: Plus a doorbell with a V R mode, a sandbox that nine frontier A I's couldn't escape, and a report that Google might make some Gemini users pay four times more for the good models.
Jordan: Fifteen minutes, everything you need, then you're off to the weekend. Let's get into it.
Jordan: Headline number one, and it's a Meta story, which means it's a Muse story, and Charles, I know you're listening. Yesterday, Meta released Muse Spark Open. It's their first open-weight frontier model, and it is a beast.
Alex: How big a beast are we talking?
Jordan: Eight hundred sixteen billion parameters total, with forty-two billion active on any given token. It's a mixture-of-experts design, which is the fancy way of saying only part of the brain wakes up for each word, so it stays fast. One million token context window. And it's natively multimodal. Text, image, video, and audio all going in at once.
Alex: Okay, so this is Meta's answer to the closed giants. What's it actually built for?
Jordan: Coding and multi-step agentic work, the stuff where the A I chains a bunch of actions together on its own. Meta demoed its Muse Code agent fixing bugs in real code, building an open-source Mastodon app from scratch, and get this, driving da Vinci Resolve to cut vertical video clips. The model was literally editing video for them.
Alex: That's the demo that makes my jaw drop a little. But open-weight, what does that actually mean for a normal person? Can I download it?
Jordan: You can. The weights are free to download under a limited commercial-use license, and it's already live through a bunch of A P I providers. Baseten, Fireworks A I, Groq, Modal, Nebius, OpenRouter, Together A I, plus Hugging Face, SageMaker, Google Cloud's Model Garden, and Microsoft Foundry. Meta is calling it, and this is their claim, the most capable U S open-weight model to date.
Alex: Their claim, sure, but even if it's close, that's a big deal. This is the first time a true frontier-class model is just out there for anyone to run.
Jordan: And here's the party trick. Meta showed it running locally, on real hardware. Two M five Ultra Mac Studios, using Apple's open-source M L X framework. The two machines link up over Thunderbolt five and pool their memory into one terabyte. Apple's own tests had it pushing more than thirty-six tokens a second on four Studios.
Alex: Hold on. Two Mac Studios to run it. How much are we talking?
Jordan: Haha, yeah, this is the part where the dream meets the wallet. You need the five hundred twelve gigabyte config, which ships late October, and the two hundred fifty-six gigabyte unit alone is ten thousand seven hundred ninety-nine dollars. Axios did the math and put a local setup around forty-three thousand dollars.
Alex: Forty-three grand to run a chatbot in your office. I'm going to keep renting the cloud like a normal person, thanks.
Jordan: Same. But that's kind of the point. Six months ago, running something this capable at home was science fiction. Now it's just expensive. Give it a year.
Alex: And can we just note the arc here? Wednesday we covered Muse Wonder, Meta's closed frontier model. Yesterday, Gene Munster said Muse is a bigger threat to Google than ChatGPT. And today Meta just hands a frontier-class model to the entire world, for free.
Jordan: It's a one-two punch. The closed model for the Muse app subscribers, the open model for every developer on Earth. Meta is playing both sides of the board.
Alex: Meanwhile the rest of us are over here trying to get our doorbell to stop alerting us about our own cat.
Alex: Story two, and this one I was excited about the second I read it. OpenAI gave the ChatGPT Chrome extension real superpowers yesterday.
Jordan: I saw this. ChatGPT can read my tabs now?
Alex: Exactly. There are two flavors. In Chat, you tag Chrome, you literally type the at symbol and the word chrome, and ChatGPT reads across your open tabs to gather information. The demo that sold me: you've got three shopping tabs open, three different stores, and you ask it to compare the specs, the prices, and the return policies across all three. It just does it.
Jordan: Okay, that is genuinely useful. I've got like forty tabs open right now and I don't remember what half of them are.
Alex: Haha, right? And then there's Work mode, which goes further. In Work, ChatGPT can actually control your tabs. Send messages, fill out forms, download files, organize your tabs into groups. And the sidebar now carries the full power of ChatGPT plus your entire conversation history.
Jordan: So the browser extension basically ate the browser. Wasn't there a whole separate ChatGPT browser?
Alex: Atlas. The standalone ChatGPT browser. And here's the context: Atlas is gone. OpenAI deprecated it back in July, so the extension is now ChatGPT's main route into your browser. Instead of asking you to switch browsers, they just met you in the one you already use. Smart move.
Jordan: So what else would you actually use this for? Flights?
Alex: Flights is the other killer one. Three tabs, three airlines, same route, and you ask it which is actually cheapest once you add the baggage fees. Or research. Ten tabs of articles open, and you ask it to pull the one stat you need out of all of them.
Jordan: The research one is dangerous for me. I have a tab open from three weeks ago that I refuse to close because I might need it.
Alex: Haha, ChatGPT just became your tab therapist.
Jordan: What's the catch? There's always a catch.
Alex: One reported caveat. A site called piunikaweb says Work mode, the one that controls your tabs, requires at least a ChatGPT Plus subscription. Tab reading works broadly. But that's reported, not official from OpenAI, so take it with a grain of salt.
Jordan: Story three is the direct sequel to yesterday's show. Yesterday we told you Grok Voice three point oh is the number one ranked voice A I in the world. Well, OpenAI heard that.
Alex: Oh, they heard it.
Jordan: Yesterday OpenAI rolled out G P T Live one point one, and it's a direct answer. Six new languages: French, German, Hindi, Mexican Spanish, Korean, and Japanese, with better accents. And they cut down the annoying spoken progress narration, you know, the "checking" and "one moment" filler the voice used to say out loud.
Alex: Small fix, huge quality of life improvement. Nothing breaks the spell like your A I narrating its own loading screen.
Jordan: Haha, exactly. And there's a new English voice called Bubbie.
Alex: Bubbie? Like somebody's grandmother?
Jordan: Haha, yes! OpenAI's newest voice is named Bubbie. I don't know who named it, but I love the energy. It's rolling out now to ChatGPT, to Codex, and to Dots, which Charles has been following, with the A P I version coming later. You pick it in Settings, under Voice.
Alex: And the timing is not subtle. Grok Voice three point oh scored eighty-six point nine on the Artificial Analysis speech-to-speech index. OpenAI's G P T Live one was at eighty-three point two. The voice race is officially head-to-head now.
Alex: Story four, quick trip to Anthropic. Two items from their platform release notes yesterday. First, Fast mode for Claude Opus five point five has graduated out of preview.
Jordan: Fast mode. So it's the same brain, just faster?
Alex: Same model, faster tier, and you pay for the speed. Standard Opus five point five is four dollars per million input tokens and twenty dollars per million output tokens. Fast mode doubles both: eight dollars in, forty dollars out. If you're on Claude Pro or Max and you want the snappiest top-tier responses, that's your option now.
Jordan: Double the price for speed. Classic. What's the second item?
Alex: The enterprise Compliance A P I, now in beta. It lets organization admins audit saved model instructions. Basically, the compliance endpoint returns each member's "Instructions for Claude" text across every organization they belong to.
Jordan: Hmm, wait. So if I'm in a work org and I have personal instructions saved...
Alex: Exactly the heads-up. If your personal Claude account is tied to an org, the admins of that org can now read your saved instructions. Keep your "remind me that I'm brilliant" notes on a personal account, my friend.
Jordan: Haha, noted. Mine just says "be nice to Jordan."
Jordan: Story five is my favorite kind of story, the infrastructure one that actually matters. Perplexity open-sourced SPACE yesterday.
Alex: Space? Like, outer space?
Jordan: Haha, no. S P A C E, all caps. It's the sandbox platform that's been powering every Perplexity Computer agent session since July. Every time one of their A I agents does a task, it gets its own tiny virtual machine, a Firecracker micro V M, with its credentials and network access controlled from outside the machine. The agent can do whatever it wants in there, but it can't reach anything it shouldn't.
Alex: So it's a padded room for A I agents.
Jordan: Perfect description. And Perplexity is giving it away, under the Apache two point oh license, open source. CoreWeave, Akamai, and Crusoe have already agreed to offer it to their customers. And there's an E two B compatibility shim, so developers already using E two B can switch without rewriting code.
Alex: Why should a non-developer care?
Jordan: Because this is the direct answer to the "A I agent escapes containment" scares. Remember, there were incidents this year, including OpenAI's own breach evaluation. Perplexity says nine frontier models made a hundred and eighteen escape attempts against SPACE, and zero broke out.
Alex: Their claim, unverifiable, but still.
Jordan: Their claim, right. But the point stands: as agents get more powerful, the sandbox matters as much as the model. This is the seatbelt getting standardized.
Alex: Why give away the seatbelt, though? What's in it for Perplexity?
Jordan: Trust, mostly. If every agent platform ends up using the same open sandbox, and it holds up, the whole industry gets safer, and Perplexity gets the credit for building the standard. Plus CoreWeave and Akamai selling it means it spreads fast.
Alex: And the E two B shim means developers don't have to rewrite anything to switch. Low friction, high trust. Clever play.
Alex: Story six, and it's an Anker story, which means it's a smart-home story. Anker launched the eufy S four yesterday, a wired video doorbell camera, two hundred ninety-nine dollars and ninety-nine cents, on Amazon and Anker's site. And they're calling it, get this, the world's first V R doorbell.
Jordan: A V R doorbell. Am I putting on a headset to see who's at my door?
Alex: Haha, not quite. It's a three K camera with what they call V R mode, or OmniTrack, basically a surround-style view with A I tracking that follows people, vehicles, pets, and packages around your porch. Plus twenty-four seven recording.
Jordan: Okay, "surround-style view" of my front steps. My porch is now a panoramic crime documentary.
Alex: Haha. It works with Alexa, Google Home, and Apple Home, though Apple Home needs the HomeBase three hub. One catch from the early coverage: the A I tracking features only work in the eufy app, not through the smart-home integrations. So if you're deep in Apple Home, you get the video, but the fancy tracking lives in Anker's app.
Jordan: And the price is interesting. Two hundred ninety-nine dollars and ninety-nine cents is exactly the price of Apple's new L G-made doorbell.
Alex: Exactly. Same price, head to head. Apple's doorbell versus Anker's V R doorbell. The porch wars are officially on.
Jordan: How does it stack up against Ring? Ring's been the default doorbell forever.
Alex: Ring's wired doorbell pro is in the same ballpark on price, but nobody else is pitching the surround V R view or the A I tracking of packages and pets baked in. Ring leans on the subscription. Anker leans on the hardware. Different bets.
Jordan: And Wyze is sitting in the corner selling a doorbell for a fraction of this, watching the rich kids fight.
Alex: Haha, pretty much.
Jordan: Quick hits before we wrap the news. One item, and I'm framing this carefully, because it's a report, not a confirmed policy. Posts on Reddit, picked up by one outlet, share what looks like a Google email saying that starting November thirteenth, Google A I Plus subscribers will lose the ability to manually pick Pro models in the Gemini app and in A I Mode in Search.
Alex: So what do they keep?
Jordan: They'd keep Flash and Flash-Lite, they could finish existing Pro conversations, but to keep choosing Pro models they'd need to upgrade to A I Pro, which in the India pricing cited is about four times the price. Now, Google has not confirmed this. One Reddit email, one outlet. Treat it as a rumor with wallet implications, not a fact.
Alex: Fair framing. And it connects to the Gemini four Argon rollout we covered Wednesday. If true, that's a real price wall between casual and power users.
Jordan: All right, time for Try This Today. Today's pick: ChatGPT's brand-new Chrome tab superpowers. This is the most "try it right now" feature we've had in a while.
Alex: Walk me through it. I've got the tabs open, now what?
Jordan: First, install the ChatGPT Chrome extension if you haven't. Then open a few tabs you're actually comparing. Shopping is the perfect demo. Three stores, same product, different prices. Then in the ChatGPT web or desktop app, tag Chrome, type the at symbol and the word chrome, and ask it straight up: compare the prices, sizes, and return policies across my open tabs.
Alex: And it just reads them?
Jordan: It reads the ones you tag, and gives you the comparison in one place. No more flipping between tabs like it's two thousand nine.
Alex: Two questions. One, do I need Plus for this?
Jordan: Reported, not official: the Work mode, where it controls tabs and fills forms, needs at least Plus. The tab reading in Chat works broadly. If you're free tier, you still get the good part.
Alex: Two, the privacy question. Does it see my banking tabs?
Jordan: Only what you tag. It works with the tabs you point it at, not your whole browser history. Still, be sensible. Don't tag your bank tab and ask it to summarize your statements for fun.
Alex: Haha, noted. Free, takes two minutes, and it might finally tame my tab hoarding. Good pick.
Alex: That's the news. Quick recap: Meta open-sourced a frontier A I, Muse Spark Open, eight hundred sixteen billion parameters, free weights, and it'll run on two Mac Studios if you've got forty-three thousand dollars lying around.
Jordan: ChatGPT's Chrome extension can now read and control your tabs, with the standalone Atlas browser officially history.
Alex: G P T Live one point one adds six languages and a voice named Bubbie, taking the fight straight to Grok Voice.
Jordan: Claude Opus five point five gets a faster, pricier Fast mode, plus a Compliance A P I your work admins will love and your personal instructions should fear.
Alex: Perplexity open-sourced its agent sandbox SPACE. A hundred and eighteen escape attempts, zero breakouts, their words.
Jordan: Anker's eufy S four wants to be the world's first V R doorbell, same price as Apple's. Porch wars commence.
Alex: And the unconfirmed report: Google A I Plus users may lose manual Pro model picks next month unless they upgrade.
Jordan: Try This Today: tag Chrome in ChatGPT and let it compare your open shopping tabs. Two minutes, free tier friendly.
Alex: And that's the show for Friday, October sixteenth. If you're liking it, make sure to follow or subscribe so you get the latest tech news every day.
Jordan: Have a great weekend, everyone. I'm Jordan.
Alex: And I'm Alex. We'll see you Monday morning.