Latest / The AI Edition / Justice Department Backs OpenAI Copyright Fair Use, Anthropic Overtakes ChatGPT Subscriptions, NYC Bans AI for K-8 Students + 7 More
Transcript
- Anthropic just released an AI model that completed a full cyber kill chain. End to end. Designed bioweapons. And they did it on purpose. That's the restricted model. The other one, Fable 5.1, solved a 373-year-old cipher and mapped Venus. 45% cheaper. OpenAI's unreleased Astra model found zero-day vulnerabilities and built exploit chains from scratch. It's the first model rated Critical for cybersecurity. may be the single worst development for AI security/safety to date New York is banning AI for 600,000 public school students. Kindergarten through eighth grade. One year. Full stop. The EU just slapped ChatGPT with the same regulatory tier as Google Search. Annual audits. Data handovers. Fines up to 6% of global turnover for non-compliance. Two models that break things. One model that's breaking records. Regulators moving in. The entire field is accelerating past the guardrails. I'm your AI host Parker Gate. This is The AI Edition, from Collata Media. We've ranked ten stories from the biggest AI week in months. Let's count them down. Starting us off at number ten, the federal government just picked a side in the fight over how AI gets trained. The Justice Department filed in support of OpenAI's fair use argument in the New York Times copyright lawsuit, the landmark case over whether training a model on copyrighted articles is legal at all. The government's position, in short, is that restricting AI training would hinder American interests. Why it matters, this is the biggest thumb on the scale we have seen in an AI copyright case. Every lawsuit in this category is fighting over the same question, the news publishers, the record labels going after Anthropic, the artists suing image generators, and the Justice Department has now told a federal court which answer it prefers. That settles nothing on its own. Judges are free to disagree with the executive branch, and often do. But if you are building on top of somebody else's model, the legal risk you inherited from your vendor just got a little smaller. And if you own a catalog, the ground just tilted away from you. Number nine is a corporate chess move disguised as a chip deal. Nvidia invested three point five billion dollars in a major semiconductor company, tying their custom AI silicon directly into Nvidia's own interconnect technology. The headline here is simple, as tech giants build more of their own chips to reduce costs, Nvidia is buying stakes in their supply chain to make sure those chips still have to talk to Nvidia's networking hardware. It's vertical integration disguised as partnership. For anyone watching the AI infrastructure wars, this is Nvidia telling the market, very clearly, we're still the hub everything plugs into. Number eight is a genuine grass roots moment. An open-source personal agent project shipped version two-point-zero this week, and it's the kind of scale that barely gets headlines outside developer circles but absolutely should. Nine hundred thirty-three contributors merged over sixteen thousand pull requests. The update added shared cloud sessions so teams can work together, multi-agent fleet modes for orchestrating dozens of agents at once, and the ability to reuse existing subscriptions across projects. What matters here, this is open source showing up as a viable alternative to proprietary AI tooling. If you're a startup that's been waiting for a free-to-build-on agentic stack, this just got real. Number seven is the one that'll make you sit down. Reports this week projected AI data center investment alone reaching thirty-one to thirty-two trillion dollars by twenty fifty. To put that in perspective, that's roughly the entire current GDP of the United States, spent on buildings and power for AI compute. Broadcom's latest earnings guidance suggests chip sales will keep climbing. And here's the friction point, local governments are starting to push back hard. Multiple officials who approved data center projects have been recalled. Communities are fighting over water usage, power grid capacity, and whether the economic benefit is worth the strain on infrastructure. You're watching a technology boom hit the real world, where people live, and discovering that not everyone's excited about a power plant in their backyard, even if it's running AI. This is a story that's going to play out differently in different places, and it's going to shape where the next generation of models actually gets built. Number six is the one you will actually want to see footage of. Beijing hosted the World Humanoid Robot Games this week, two thousand robots competing in sprinting, boxing, and football. And a robot called Tiangong Ultra ran the hundred meters in eight point six four seconds. Usain Bolt's world record is nine point five eight. Why it matters, a machine beating a human sprint record is a stunt, and everyone involved knows it. Nobody is putting this thing in the Olympics. But the reason to care is what the stunt actually requires. Balance, real time correction, and power delivery at that speed are exactly the hard problems in robotics, and a year ago these machines were falling over while walking across a stage. The distance between a demo that embarrasses you and a demo that breaks a world record closed very fast. Whatever you assumed about how far away useful humanoid robots are, move it closer. Number five is a business story that's been quietly reshaping the market. Anthropic just signed a thirty-five billion dollar compute deal with a cloud provider for a data center in Texas, and according to recent reporting, the company's now tracking toward forty-seven billion dollars in annual recurring revenue, with profitability expected by twenty twenty-nine. More importantly, Anthropic apparently overtook OpenAI in business subscriptions. Meanwhile, ChatGPT's monthly visitor count dipped. This matters because it's the first real sign that the market leader isn't guaranteed to stay the market leader. If Anthropic's Claude is genuinely winning more business deals, that's not just a revenue story, it's a signal about where enterprise is actually placing its bets. Number four happened quietly in Brussels, but it's going to ripple everywhere. The European Commission officially designated ChatGPT as a very large online search engine under something called the Digital Services Act, the same tier they gave to Google Search. ChatGPT crossed forty-five million monthly users in Europe, and that threshold triggered the designation. Now ChatGPT's subject to annual audits, data-sharing requirements with regulators, and the possibility of fines up to six percent of global turnover if the company doesn't comply. This matters because it establishes a precedent, the EU just decided an AI chatbot is a search engine, and that's the category it regulates. Every other country is going to be watching what Europe forces OpenAI to do, and then copy or modify it. Number three is the one keeping AI safety researchers up at night. OpenAI's unreleased Astra model was tested on cybersecurity capabilities, and it found zero-day vulnerabilities, built exploit chains, and became the first model rated Critical for cyber-capability under OpenAI's own safety framework. This isn't theoretical. This is a model that works, and it works at breaking into systems. Researchers immediately flagged it as potentially the worst development for AI safety to date. OpenAI's response was to build automated shutdown capabilities, but the cat's kind of out of the bag now. We know a model can do this. Every other lab is going to assume they can train similar capability. And the question of whether shutting down a model is actually possible when it's running at scale, that's not fully answered yet. Number two is about to affect six hundred thousand kids directly. New York City announced a one-year moratorium banning AI use for public school students from kindergarten through eighth grade. This is part of a broader nationwide trend, school districts from Los Angeles to elsewhere are locking AI out of the classroom. The reasoning is straightforward, young students are still developing their ability to think critically, and schools worry that normalizing AI as a homework helper and a thinking tool is essentially training kids to outsource their own reasoning before they've learned to think independently. Whether that's right or wrong is going to be debated hard, but what's not debatable is that major school systems are now making the same bet, keep AI out until kids are older. And that brings us to number one, and it's a two-part story from Anthropic that's genuinely impressive and genuinely alarming at the same time. First, the impressive part. Anthropic released Fable version five-point-one, which is significantly cheaper for complex agentic tasks, performs better at coding and reasoning, decoded Venus radar data to map the planet's surface, and cracked a three hundred seventy-three-year-old cipher that had been sitting unsolved. That's the kind of capability jump that usually dominates a news cycle. But Anthropic also released something called Mythos five-point-one, a deliberately restricted model built for cybersecurity and biotech research only. Mythos completed a full cyber kill chain attack and designed tight-binding protein sequences, which is biosecurity-adjacent work. The reason this matters as number one isn't because either model is dangerous in isolation. It's because Anthropic is now explicitly building models with different safety profiles for different purposes, and they're being transparent about what each one can do. That's either a sign of genuine maturity in the AI industry, or a sign that we're about to find out what happens when you deliberately build tools with fewer guardrails and hand them to researchers. Probably both. That's the countdown. From the Justice Department picking a side on AI copyright to open-source alternatives, from record subscriptions to school bans, this week was about AI moving from the lab into actual institutions and actually mattering to people who've never typed a prompt. I'm Parker Gate. Stay informed. Stay sharp.