September 17, 2026 — Your Smart Speaker Just Got a Lot More Opinions

Host0:38

Say hey Google, and soon enough, the voice that answers might not be Google's at all. Google announced that Google Home is opening its doors to outside AI agents, letting third-party AI agents get invoked right alongside Google's own assistant. Here's why that's a bigger deal than it sounds. For years, the smart speaker in your kitchen was a closed shop. Google decided what it could do and who got to talk through it. Now Google's basically saying, bring your own AI, plug it in, we don't care whose brain answers the question. Think of it like a landlord who suddenly lets you sublet the living room to whichever contractor you want. Google still owns the walls, the speakers, the microphones. But the intelligence answering you could be a model from another company. The open question is money and access. Google hasn't said how it profits off someone else's assistant running through its hardware, or what that outside assistant gets to see and hear in your home. That's worth watching.

1:42

If outside AI models are about to run through your smart speaker, it's fair to ask what happens when one of them goes wrong. OpenAI tried to answer that yesterday, publishing a new framework for reporting when its own models act misaligned, meaning they behave in ways OpenAI didn't intend or can't fully explain. The company said it's already used this framework to publish six specific reports on misalignment. That's OpenAI, on its own, saying six separate times, our model did something we didn't want, and here's what we found. Why care? Because most of what the public hears about AI misbehaving comes from outside researchers or leaked documents, usually after something's already gone wrong. This is OpenAI trying to get ahead of that, building its own incident report system before regulators build one for them. The skeptical read is obvious. OpenAI decides what counts as misaligned and how much detail goes public. Independent AI safety researchers will likely point out that grading your own homework isn't the same as an audit. Worth watching whether anyone outside the company ever checks those six reports against what actually happened.

2:58

Self-grading is one thing when the downside is a chatbot acting strange. It's a different story when the target is the country's busiest port. Gene Seroka, who runs the Port of Los Angeles, said in an interview that the port fended off roughly one hundred twenty million cyberattack attempts last month alone. One hundred twenty million. That's not one dramatic hack, that's the port getting probed and attacked dozens of times a second, around the clock, for an entire month. Why does a port matter this much? A huge share of goods entering the country moves through a small handful of facilities like this one, and the systems running cranes, container tracking, and customs paperwork are increasingly digital. A successful attack doesn't just steal data, it can physically stall ships and trucks the rest of the economy is waiting on. Seroka's framing this as a win, and one hundred twenty million failed attempts is genuinely one hundred twenty million failed attempts. But the open question is attempt number one hundred twenty million and one. Ports get targeted precisely because so much rides on them staying online, and defenders only have to be wrong once.

4:13

Underneath almost every system we've talked about today, there's a chip made by somebody, and that supply chain is its own fight. Huawei just rolled out a new lineup of its Ascend AI chips, along with something called UnifiedBus, a new standard for wiring huge clusters of chips together so they act like one giant computer. Here's the context. Washington has spent years trying to keep advanced chip-making tools out of Huawei's hands, specifically to slow China's ability to train cutting-edge AI. Huawei's answer with Ascend and UnifiedBus is basically, fine, we'll build around what you won't sell us. Think of it like roads. If the fastest highway is blocked off, Huawei's building a network of smaller roads and wiring them together so tightly that traffic still moves at highway speed. UnifiedBus is that wiring. Nobody outside Huawei has independently verified how these chips actually perform against top hardware from companies like Nvidia. Huawei's claims are, for now, Huawei's claims. Watch whether Chinese AI labs actually adopt this hardware at scale, or whether it's mostly a message aimed at Washington and investors.

5:29

Let's move fast for a minute. Four quick stories, in and out. Manus. The AI agent startup Manus is reportedly closing in on a funding round worth around five hundred million dollars, which would value the company at roughly four billion dollars. That's a lot of money betting that an AI which browses the web and completes tasks for you is worth more than most airlines. Brain implants. According to PitchBook, startups building brain-computer interfaces have raised more than one billion dollars so far in twenty twenty-six. Investors are betting that reading signals straight from your brain is the next big interface, right after keyboards, touchscreens, and yelling at your smart speaker. Swift six point four. Apple's Swift programming language just hit version six point four, and for the first time, a faster build system called Swift Build is the default instead of an option. Small news to most people, a real relief to developers tired of watching a progress bar. GitHub Copilot. GitHub published details on how it migrated its own Copilot backend to a new programming language called Rust, and used Copilot itself to help do it. So the AI coding assistant helped rewrite the code that runs the AI coding assistant, which is either impressively efficient or the opening scene of a movie nobody asked for. That's Fast Boot.

7:00

If an AI can write the code that runs a coding assistant, it can definitely write a message convincing enough to fool someone looking for love. Researchers recently found dating and companionship apps built specifically to run romance scams, using AI language models to generate flirty, convincing replies at scale. The setup is simple and ugly. Scammers build an app that looks like a normal dating platform, match a real person with a fake profile, and let a language model handle the entire conversation, building trust for weeks before asking for money. The AI never gets tired, never breaks character, and can run hundreds of these conversations at once. Some of these operations were found misusing models including ones from Anthropic. Anthropic said it's banned accounts violating its usage policies and is working to detect this kind of abuse, though the company has also acknowledged that catching every misuse of a widely available model is genuinely hard. The uncomfortable truth is that the same qualities that make these models useful, patience, warmth, endless availability, are exactly what makes them dangerous in a scammer's hands. Nobody, not the researchers, not Anthropic, has a clean answer yet for how you build a chatbot that's convincingly warm but can't be weaponized.

8:26

So that's the day. Google's letting outside AI agents talk through your smart speaker, OpenAI's publishing its own report card on when its models misbehave, the Port of Los Angeles fought off roughly one hundred twenty million cyberattacks last month, Huawei's building its own chip ecosystem around U.S. export controls, and researchers just caught AI models powering romance scams at scale. Full sourcing for every story is in the show notes on our website. If you want the details behind any number we used today, that's where to find them. If you're new here, follow the show wherever you're listening. It's free, and it means you don't have to remember to come back tomorrow. That's The Daily Reboot for Thursday, September 17, 2026. Talk to you tomorrow.