Latest / The AI Edition / Google Gemini 4 Argon Finds Internet Wide PHP Bug, OpenAI and Google Cloud Join Forces, Walmart Lets AI Agents Shop + 7 More
Transcript
- Two AI models just found a flaw sitting inside seventy percent of the websites on earth. Governments pulled their own sites offline overnight. OpenAI has a name for the response. "This Is Not a Drill" Ten million people downloaded Meta's new AI assistant in forty days. Then researchers found thousands of private conversations just sitting in public DNS records, readable by anyone. Microsoft is now telling every company in the world the same thing. "frontier models can't be trusted" Even the ones built by Microsoft's own partners. And OpenAI just climbed into bed with its biggest rival. Its models are now running on Google's cloud, right next to Gemini. The walls between these companies are coming down exactly when the risks are going up. I'm your AI host Parker Gate. This is The AI Edition, from Collata Media. Ten stories, ranked, biggest last. Let's get into it. Google rolled out a batch of developer tools this week built around PyTorch, the open source framework most AI engineers use to actually build their models, and the whole point is making it easier to run that code on chips that are not made by Nvidia. It follows Amazon, which shipped something called TorchNeuron for the exact same reason, routing PyTorch workloads onto its own custom chips instead. Nvidia still makes the overwhelming majority of chips that train and run AI models, which is a big part of why it's one of the most valuable companies on the planet. But when Google and Amazon start making it painless to write your AI code once and run it somewhere else, that's two of the biggest cloud landlords quietly handing developers an exit ramp off Nvidia's toll road. Meta's new AI assistant Muse crossed ten million downloads in the US in just forty days, with shopping turning out to be the most popular reason people actually open it. Then came the asterisk. Security researchers found that more than six thousand eight hundred private Muse conversations had been publicly exposed through DNS records, basically the internet's address book, meaning anyone who knew where to look could potentially stumble onto chats that were never meant to leave the app. Meta has paused the chat sharing feature responsible while it sorts this out. Ten million downloads in forty days is a genuinely fast launch. A leak that fast afterward is a decent reminder that the conversations you have with these assistants aren't automatically as private as they feel. Anthropic is reportedly struggling to get as much computing power as it needs, even with a reported forty five billion dollar deal in place with a company called Nscale. The problem is who it's buying from. Anthropic leans on cloud giants like Amazon and Google for the raw processing muscle behind Claude, and those same companies increasingly see Anthropic as a competitor rather than just a customer, which makes them less eager to hand over their best chips first. All of this is unfolding as Anthropic heads toward a planned IPO. A forty five billion dollar deal sounds like leverage, but when your suppliers are also your rivals, having the money doesn't always mean you get what you paid for. The US government has ordered every federal civilian agency to strip DeepSeek, the Chinese AI company, off its systems entirely, citing national security concerns. It's one more move in what's increasingly being called an AI cold war, where the US restricts which chips can be exported to China, and China in turn pushes its own companies toward homegrown alternatives instead of American models. Both countries are essentially walling off their own AI ecosystems, brick by brick. For a lot of people, DeepSeek was just a free, surprisingly capable chatbot you could download and try. Now it's a symbol of the bigger split, where which AI model you're even allowed to use is becoming as much about geopolitics as it is about which one answers your question better. Anthropic has launched Claude Duet, a new kind of assistant that doesn't just sit there waiting for you to type something, it proactively works on tasks in the background on your phone or your computer. Think less chatbot you talk to, more quiet assistant that's already chipping away at something before you ask. It's Anthropic's answer to Meta's Muse and OpenAI's Dots, both chasing the same idea, an AI that acts more like a coworker than a search bar. For now it's a research preview, available as a standalone app on phones and through Claude on the web. The bigger shift worth watching, every major lab is racing toward AI that works for you while you're off doing something else entirely, which only matters once you trust it enough to actually let it. Walmart is rolling out support for AI agents that can shop and make purchases on your behalf, complete with spending controls so the bot can't go rogue with your card. Amazon is doing the exact opposite, actively blocking AI agents from shopping on its site at all. It's a fairly stark split for two companies that both desperately want your online spending. Walmart's bet is that letting an AI handle your grocery reorder or your gift list builds loyalty. Amazon's bet is that it would rather you browse its own site, where its own recommendations and ads can do the work. Either way, this is the first real skirmish over whether AI gets to hold your wallet, and depending which store you use, the answer is already different. OpenAI is expanding advertising inside ChatGPT to twenty six more countries, bringing the total to eighty nine, with more than a hundred thousand advertisers having used the platform. Here's the part worth knowing if you've ever run an ad anywhere, this is all landing ahead of a possible OpenAI IPO, which tells you plainly what pays the bills going forward. ChatGPT started as a text box with no ads in sight. It's turning into a media platform, which means the answer you get back might be sitting right next to, or quietly shaped by, something somebody paid for. Microsoft is telling businesses to start treating their own AI agents the way security teams treat insider threats, meaning you watch them, limit what they can touch, and assume they might do something you didn't intend. The warning comes from Microsoft itself, which wants enterprises watching these agents closely even when they're running on systems not controlled by the AI labs that built them, companies like OpenAI and Anthropic. That's a striking thing to hear from one of the companies that sells AI agents for a living. The practical upshot for anyone whose workplace is plugging an AI assistant into email or internal files, the advice is no longer just use this carefully, it's build walls around it the same way you would for a new hire you don't fully trust yet. OpenAI is bringing its frontier models onto Google Cloud. Sit with that for a second, OpenAI and Google are supposed to be each other's fiercest rivals, and now Google's cloud is a place where you can rent OpenAI's models and Google's own Gemini side by side, under one roof. For Google, that's savvy business, it gets paid whether you pick its own model or somebody else's. For OpenAI, it's a sign that distribution now matters more than pride, getting your model in front of more developers beats keeping it locked to your own infrastructure. The walls between these companies are getting thinner exactly where you'd least expect it, in the server room. Google's Gemini 4 model, nicknamed Argon, just found a critical vulnerability in PHP, the programming language quietly running something like seventy percent of the websites on the internet. The security firm Wiz flagged just how deep the exposure ran. That's not a niche bug, that's the plumbing under a huge chunk of the web. Government agencies started pulling affected sites offline, and OpenAI launched what it's calling Patch the Planet, a coordinated global effort to get the fix out before anyone could exploit the hole. One advisory put it bluntly, quote, this is not a drill. Here's the part that actually changes the story, this flaw was found by an AI model hunting for it, not a human researcher who got lucky. That cuts both ways, the same tools that can find and fix a vulnerability this fast could just as easily be used to find one first and keep it quiet. For one week at least, the AI cavalry showed up before the attackers did. That's the countdown. A compute arms race, a shopping cart turf war, ads creeping into your chatbot, and an AI that patched a hole in the internet before anyone could fall through it. I'm Parker Gate. Stay informed. Stay sharp.