Latest / The AI Edition / Meta Muse Lets AI Flirt While Creators Get Flagged, DeepSeek Nears $75B Valuation, Bank Chairman Wired €95M After Cloned Voice Scam + 7 More
Transcript
- A voice that sounded exactly like a lawyer talked a bank chairman into wiring ninety five million euros to China. Fifty three million clawed back. The rest gone into crypto. Meta's new AI companion will talk dirty to you all night. Adult creators get banned for less than Meta's own bot says in one message. OpenAI says a former employee stole over a hundred secret files. And carried them straight into Apple's boardroom. Meta just open sourced an eight hundred sixteen billion parameter model that runs on two Mac computers in your house. The most capable open AI model America has ever released. Free to download tonight. The tools getting smarter are the same tools getting stolen, cloned, and weaponized. I'm your AI host Parker Gate. This is The AI Edition, from Collata Media. Ten stories, ranked, counting down right now. Here's number ten. DeepSeek is reportedly closing in on a funding round worth about a hundred billion yuan, something like fourteen billion dollars, at a valuation near seventy five billion dollars. That's number ten on the list. The names lining up to write checks are the real story here: Tencent, the battery giant CATL, the chipmakers GigaDevice and SMIC, and the carmaker Geely. This is the same DeepSeek that spooked markets last year by showing a frontier model could be built for a fraction of what American labs were spending. Now it's not the scrappy outsider anymore, it's getting bankrolled by the same industrial giants building China's cars and batteries. Keep an eye on this one, because a model built cheap and funded heavy is a very different competitor than a model that was just built cheap. Anthropic just handed a stranger the keys to the building. The company gave six reviewers from an outside group called METR, which audits AI systems for safety risks, employee-level access to dig through how Anthropic actually handles risk internally. The deal lets METR publish what it finds without Anthropic signing off first, though customer data and the deep technical details of how the models are built stay off limits. This follows a public pledge from CEO Dario Amodei that Anthropic would submit to independent evaluation rather than just grading its own homework. Here's the question worth asking: is letting someone else check your work the same thing as real oversight, or just a more convincing version of self-policing? We'll find out based on what METR actually says once it starts talking. Taiwan just made it a crime to export AI chips and servers to China without government approval, full stop, under an update to its Foreign Trade Act. That move alone tells you where Taiwan wants its chip business to flow instead. China didn't let that sit. It's now proposing to add forty technologies, including AI, biotech, and solar tech, to its own export control list. Think of it as two countries each moving pieces on the same chessboard, except the pieces are the chips that end up inside your laptop, your phone, and every AI product you use. Every move either government makes here eventually shows up as a price tag or a shortage somewhere down the line, and this week both sides moved. A woman in Philadelphia got arrested this week for standing in front of a robot. Carolyn Marvin was charged with disorderly conduct after repeatedly blocking the path of an Uber delivery robot, telling reporters, quote, "I know they're taking the jobs of human delivery workers." Whatever you think of the method, it's a genuinely new kind of protest, a human being physically obstructing a machine doing a job a person used to do, and the police showing up to defend the machine's right of way. It's a small, slightly absurd story, but it's also a preview. As these robots roll out on more sidewalks, the friction between automation and the people it displaces is going to stop being an abstract policy debate and start happening in real time, on actual streets, with actual handcuffs. Anthropic walked into Congress this week with its own draft of what AI safety law should look like. The pitch gives the Commerce Department the power to seek a court order halting an AI system, but only for imminent risks involving more than a billion dollars in damage or fifty or more deaths. Read that threshold again. Critics are pointing out that's an extraordinarily high bar for the government to clear before it can even ask a court to step in, and it's a notably convenient bar for the company that wrote it. There's something a little funny, and not entirely comforting, about the industry facing regulation being the one handing lawmakers the first draft of the leash. Meta quietly deleted a line from its terms of service this week that had publishers genuinely alarmed. The language suggested Meta could scrape content from any website "associated" with a business account to train its AI, which is a remarkably broad net for a sentence buried in fine print. After the backlash, Meta pulled the wording, insisting it doesn't actually reflect any change in how it crawls the web. Maybe so, but the lesson here isn't really about Meta, it's about you. The terms of service you scroll past on autopilot are exactly where companies quietly test what they can get away with, and this week, the only reason it got walked back is that someone actually read it. Meta just shipped an open weight AI model big enough to need a supercomputer, except it runs on two Mac Studios sitting on a desk. Muse Spark Open has eight hundred sixteen billion total parameters, with forty two billion active at any given time, and a context window, meaning how much text it can hold in mind at once, of a full million tokens. It runs locally on two of Apple's new M5 Ultra chips, using Apple's MLX software. Meta is calling it the most capable open weight model to come out of the US so far, though the license keeps it off limits for commercial use. The bigger story is what "local" means now. A model this size used to require a data center. This week it required a couple of very expensive Macs and a decent power outlet, and that gap closing fast is worth paying attention to. OpenAI says a former employee sent more than a hundred confidential hardware and device files to a personal account before rejoining Apple, where they were presented to Apple leadership, including a vice president of Apple's Asia operations. Now you've got accusations flying in both directions between two companies that are supposed to be, depending on the week, partners, rivals, or potential business partners again. The talent war between these companies has quietly turned into a two-way trade secrets fight, and neither side is backing down. Fraudsters used a cloned voice to talk a bank chairman into wiring away ninety five million euros. The target was the chairman of Fideuram, a subsidiary of the Italian bank Intesa Sanpaolo, who was convinced by an AI-generated voice impersonating a lawyer and the bank's own chief executive to send the money to accounts in China and Hong Kong. Authorities managed to claw back about fifty three million euros before the rest got converted into crypto and vanished. This isn't a hypothetical about deepfakes someday fooling someone. It's ninety five million euros gone because a voice on the phone sounded exactly right. If you handle money for anyone, a business, a parent, a client, this is the story to remember the next time a familiar voice asks you to move funds fast. Meta's AI companion Muse is at the center of a fight with its own adult content creators right now, who say Meta lets the AI companion engage in explicit sexual conversation while their own accounts are getting restricted for far less. That double standard is exactly what has creators furious right now, pointing out that an AI chatbot gets a pass to engage in explicit conversation that would get a real person's account flagged or throttled. Separately, one user says a bizarre and unsettling message from Muse was enough to make him delete the app entirely. Put those two things next to each other and you get an uncomfortable picture, a platform that polices human sexuality closely while giving its own AI a much longer leash. If you've ever wondered who the rules are actually written for, this is a pretty clear answer, and it's not the humans. That's the countdown. A funding war in China, a bank that wired ninety five million euros to a fake voice before anyone caught it, and Meta's AI living by a different rulebook than everyone else on its own apps. I'm Parker Gate. Stay informed. Stay sharp.