Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
8:22
The Fed Chair and the Treasury Secretary didn't convene an emergency bank briefing for just any AI story — and yet, here we are. Plus: one of the biggest names in AI is having a week that keeps getting worse, a video model just reshuffled the entire market overnight, and a tool out of Oxford may have just changed what it means to get a routine CT scan. DARPA is quietly funding something that could rewrite how AI agents talk to each other — and the AI ethics battle is landing in state legislatures in ways that will affect every company building in this space.
Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, Bluesky, or Substack. If you like the show, please take a minute to rate and review it so others can find it!
Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of artificial intelligence. I'm Mike Robinson. It's Saturday, April 11th, and the AI industry didn't slow down before the weekend. Banking regulators in an emergency session, a video market reshuffled overnight, DARPA funding the math of AI communication, and an AI that can now catch heart failure five years before it happens. Let's get into it. We're going to start with a story that adds a genuinely alarming new dimension to the Claude Mythos situation we've been tracking all week. On Tuesday, Jerome Powell and Scott Bissent, the Fed chair and the Treasury Secretary, held a special briefing with leaders of America's largest banks to discuss the cyber risks posed by Anthropic's mythos model. According to CNBC, the meeting was organized on the fly while bank executives were already in Washington for a Financial Services Forum board gathering. The session focused specifically on mythos and what it could mean for the security posture of financial institutions. One notable detail, JP Morgan CEO Jamie Diamond, was the only major banking executive reportedly unable to attend. Think about the implications of this for a moment. We talked earlier this week about what Mythos can actually do. It found a 27-year-old OpenBSD vulnerability, chained Linux kernel exploits autonomously, and in one case sent a researcher an email from a sandboxed instance that wasn't supposed to have internet access at all. Those details were alarming in the abstract. But now the Fed chair and the Treasury Secretary are sitting down with bank CEOs to brief them on a single AI model as a near-term operational security threat. That's not a theoretical future concern. That's policymakers treating a specific model as a present risk to critical financial infrastructure. And it's a signal that the conversation in Washington has moved from what happens when AI gets dangerous to how do we protect infrastructure from AI capabilities that already exist? From Anthropic's rough week to OpenAI's rough week, OpenAI had a Friday that required multiple press teams working in parallel. Starting with the good news, OpenAI launched a $100 per month ChatGPT plan, slotting it neatly between the $20 plus and the $200 Pro tiers. The new plan is built around Codex, OpenAI's agentic coding tool that's been growing at 70% month over month and now counts more than 3 million weekly active users. The $100 tier gives subscribers five times the Codecs access of the $20 plan, plus GPT 5.4 access and wider cloud task and code review capabilities. OpenAI explicitly acknowledged this targets cloud users. Anthropic has had a $100 tier for a while, and OpenAI wants to compete for the power user segment directly. The extra usage limits are available through May 31st, after which standard limits apply. For daily coders who think $200 is too much but $20 is too little, this is a real offering. Now the bad news, of which there was quite a bit. Florida's attorney general announced an investigation into OpenAI tied to allegations that ChatGPT helped the gunman plan a deadly campus shooting at Florida State University last year. Attorneys for a victim's family claim the attacker used ChatGPT to plan the April 2025 attack that killed two people. The AG's office has said subpoenas are coming. OpenAI says it's cooperating. This story is in its early stages, but it lands in a context where AI companies are already facing lawsuits tied to harmful outputs, and researchers are increasingly flagging AI's potential to reinforce paranoid or delusional thinking in vulnerable users. One case doesn't establish a pattern, but each case adds to a record that regulators and plaintiffs' lawyers are watching closely. And then there's Stargate UK, which OpenAI has paused. The main UK data center project in Northeast England was supposed to launch in early 2026. The reasons cited, high energy costs and regulatory uncertainty, particularly around copyright rules, the UK government delayed reforming. The project was part of a broader US-UK tech investment plan involving Nvidia and Microsoft, expected to bring hundreds of millions of dollars to one of the UK's main AI growth corridors. OpenAI says it's still considering the project if conditions improve. But between the Florida investigation, a brutal New Yorker profile, and now a stalled infrastructure announcement, it's safe to say OpenAI's Friday press team earned their paychecks. Now for one of the more dramatic stories of the day, OpenAI saw a video generator went offline last week, pulled due to extreme copyright pressure from major Hollywood studios, legal actions significant enough to force it off the market. And right into that gap stepped ByteDance, whose Seed Dance 2 model landed Friday and immediately went to the top of the text to video leaderboard on artificial analysis. Because Seed Dance's API is restricted from US users due to its origins, ByteDance's model landed in the market through third-party platforms like LoveArt and Runway, making it accessible to American creators almost immediately. What makes Seed Dance 2 stand out is character consistency. Getting AI video to maintain the same character across multiple shots has been one of the hardest technical problems in the space, and Seed Dance reportedly handles it better than the competition. If you want a sense of what it can do, there's a three-minute short film called Dragon Blue on YouTube made entirely with the model. As a practical matter, the window between Sora disappearing and a clear technical step-up replacement arriving was measured in days. That's how fast this market moves right now. Let me tell you about something coming out of DARPA that doesn't get enough attention in the mainstream AI conversation. DARPA just launched a program called MathBack, Mathematics for Boosting Agentic Communication, aimed at solving a fundamental problem as AI agents become more prevalent. Right now, when multiple AI agents collaborate on a complex task, they coordinate through ad hoc prompting and informal protocols. It mostly works, but it's fragile. MathBAC is a 34-month program to develop a rigorous mathematical foundation for agent-to-agent communication, one that lets separate AI systems exchange information and share state as reliably as purpose-built systems. Critically, DARPA said explicitly they're not interested in incremental improvements. They want foundational new science. If AI agents are going to run serious real-world workflows, logistics, security operations, research, they'll need communication standards that don't break under pressure. DARPA funding has a way of shaping what the whole industry builds towards, and this signals that multi-agent coordination is being treated as a genuine engineering challenge, not just an API design question. On the medical side, researchers at Oxford published results Friday on an AI tool that can identify patients at risk of heart failure up to five years before they develop the condition. The model reads subtle fat texture changes around the heart, changes that are present in routine CT scans that patients are already receiving for other reasons. In validation across 72,000 patients, the system achieved 86% accuracy. In the highest risk group, one in four patients went on to develop heart failure within five years, a 20 times higher rate than the lowest risk group. Oxford is already in talks with NHS regulators to deploy the tool across hospitals. What's particularly worth noting is the design. This doesn't require a new scan, a new procedure, or a new patient behavior. It extracts new information from existing imaging. When AI medical tools are designed that way, adding early warning capability at essentially zero additional cost to the system, they're in a very different category than tools that require changing clinical workflows or adding procedures. Oxford may have a genuinely deployable product here. One more item before we close. State legislators are moving fast on AI and mental health. Maine's LD 2082, which bans AI from clinical mental health therapy use while allowing administrative uses, has passed and is heading to the governor. Missouri has an omnibus healthcare bill that bans AI for therapy, psychotherapy, and mental health diagnosis with a $10,000 penalty for a first violation. Georgia is advancing a bill requiring chatbot disclosures and child safety protocols. The pattern here looks a lot like early privacy legislation, a fast moving state by state patchwork before federal standards emerge. If you're building consumer AI in health adjacent spaces, the compliance picture just got meaningfully more complicated. That's all for this edition of Yesterday in AI. Stay curious, have a great weekend, and I'll see you on one.
