Anthropic locked their most dangerous AI in a high -security vault. A 22 -year -old dropout cracked it open. He gave it to the world in under two weeks. The digital walls are coming down incredibly fast right now. Welcome to the Deep Dive. Today, we are exploring a massive shift. The balance of technological power is fundamentally changing. We are bringing you the latest signals from the frontier. Right. We're looking at how a single GitHub repository is threatening a trillion
dollars of big tech infrastructure today. We're also looking at your doctor's office. AI is officially stepping in as an active medical teammate. And finally, we'll explore the messy, deeply human friction happening as these models integrate into our daily lives. It's a fascinating moment in time. Let's start with this idea of protecting AI. Entropic had this incredibly powerful system called Claude Mythos. They actually decided it was simply too dangerous for the public. Yeah,
they wanted to keep it strictly contained. They launched something they called Project Glasswing. Right. The entire goal was to secure the model completely. They were carefully vetting all their corporate partners. They essentially built this massive digital moat to keep it safe. They wanted absolute control over the deployment. You build a vault and you think the vault is secure. But that protection just failed spectacular. It really did. Enter a developer named Kai Gomez. He's
a 22 -year -old college dropout. He just looked at the architecture and cracked it. I mean, he reverse engineered the entire Claude Mythos system right from the outside. And he released his own open source clone. He called it Open Mythos. He did this in less than two weeks. Beat. I just think about the historical irony here. Oh, it's massive. Massive corporations build these giant impenetrable digital castles and then... A kid just invents a ladder. A very, very efficient
ladder. The unfiltered version is now effectively out in the wild. This is an absolute nightmare for AI regulators everywhere. Yeah, I bet. Terrifies the corporate moat defenders across the industry. So why is it so terrifying for them strategically? What did this kid actually figure out? It comes down to a profound architecture shift. Older AI models use massive flat layers to process data. Okay. They process information sequentially
through billions of different parameters. Let's define parameters really quickly for everyone. Parameters are basically the tiny mathematical dials controlling how software thinks. Exactly. And to make an AI smarter, you usually just add more layers. You add more dials. Right. But that requires a massive amount of computing power. OpenMythos fundamentally shifts how the AI actually operates. It uses a very small subset of layers instead. Then it loops the data through them
repeatedly. So it reuses the exact same digital pathways over and over. It doesn't need to travel down a massive flat line. Yes, that is the brilliant part of the design. It is exactly like stacking Lego blocks of data. You build incredibly complex towering structures from a few simple repeating pieces. Wow. It is highly efficient and it's incredibly fast. Whoa. Imagine scaling to a billion queries. If it's that efficient, you could run
frontier intelligence on consumer hardware. Yep. You wouldn't even need a massive server farm. And that is the trillion dollar problem for big tech right now. We literally just saw big tech commit one. trillion dollars to infrastructure. It's staggering. They are building massive power hungry data centers all over the world. Just massive physical warehouses filled with extremely expensive computer chips. They are pouring concrete
and laying copper wire everywhere. Yes. And the financial markets are aggressively backing this approach to KKR just raised 10 billion dollars for a new fund. They are backing a massive new A .I. infrastructure company led by Adams. Lipsky, the former AWS chief. Exactly. They are focusing entirely on data centers, power, and connectivity. They're building alongside the massive cloud computing companies. They are betting everything on centralized power. To support the constantly
rising AI demand. Right. But here's the massive catch in their strategy. What if small and loopy is fundamentally better architecture? What if it simply beats big and flat forever? Then that massive hardware investment suddenly hits an efficiency wall. You don't need a massive data... center if a laptop works. It's a catastrophic efficiency wall for those companies. Gomez is actively building a decentralized version of superintelligence. Meanwhile, Meta and Google
are spending billions to centralize it. The next world -changing breakthrough probably won't come from a boardroom. It will likely come directly from a GitHub repository. Two sec silence. Is Big Tech's hardware advantage officially dead? It's not totally dead, but it's deeply vulnerable now. Proprietary hardware is no longer an absolute shield. Self -taught developers can clone frontier architectures in days. So raw computing power can't protect them from smarter, leaner algorithms.
Mid -role sponsor, Reed. We just talked about how these powerful, lean models are built. Now let's look at where they are actually going. They are actively sitting in the room with us today. Yeah, they really are. They are diagnosing our bodies and running our corporate workflows. We are letting them into our most intimate spaces. Google DeepMind just unveiled something called the AI Co -Clinician. This is not just a text box on a screen anymore. It is a real -time,
video -capable teammate in the exam room. It is designed to sit in on your actual doctor appointments. It literally watches the interaction unfold live. DeepMind is calling this new concept Triadic Care. Triadic meaning a core group of three. The patient, the doctor, and the AI working together. Beat. I mean, the AI isn't textbook anymore. It is a hyper observant medical resident who never blinks. That's a really great way to visualize it. It actually watches your breathing during
the physical exam. It monitors how your chest rises and falls through the camera. Wow. It even checks your inhaler technique on video. Right. Make sure you're holding it at the exact right angle. Exactly. It actively helps your doctor spot things the human eye might miss. And to achieve this safely, it operates using two very distinct modules. Okay. First, there is the talker module. And then that is the part that actually interacts with the patient. It retrieves the
medical evidence and speaks in the room. Yes. But the second part is the real engineering breakthrough here. It is a completely silent overseer module running in the background. What does the overseer actually do mechanically? How does it prevent the AI from just making things up? It constantly watches the talker module before it speaks. It is specifically looking for any unsafe medical advice. So every piece of advice gets double -checked instantly. It is cross -referenced directly
against clinical -grade evidence. Then it is cited in real time for the human doctor to verify. The test results on this system are honestly staggering. They tested it on 98 highly realistic primary care queries. The system logged zero critical errors in 97 of them. Wait, zero errors? It actually beat every other frontier model on open -ended drug questions. That even includes the brand new GPT -5 .4. It did. And the video simulations they ran are even more impressive
to watch. The AI correctly guided patients through complex physical shoulder maneuvers. It helped them isolate and find specific rotator cuff injuries. It even helped diagnose a case of myasthenia gravis. It did this purely through video observation of the patient's face. In blind clinical tests, physicians vastly preferred this AI teammate. They chose it over their current diagnostic tools by 67 % to 26%. But the human element is still vital here. The study clearly found that experienced
physicians still beat the AI overall. The human intuition combined with experience still wins out. For now, yes. But this triad concept is also rapidly entering the corporate world. OpenAI just launched a product called Codex for Work. It is designed specifically for enterprise teams. Aaron Livy over at Box is already hiring agent engineers. These are specialized people who build AI to autonomously do tasks. He wants to automate core business workflows entirely. Two sec silence.
How does the system prevent the AI from giving confident but deadly medical advice? It uses that critical silent overseer module. This module watches the main system and cross -references advice against clinical grade evidence in real time. It basically has an internal chaperone fact -checking every single word. Exactly. The architecture is designed entirely around patient safety first. Which brings us to the messy part
of this transition. Integrating agents that can see us and talk to us is incredibly jarring. It creates immediate painful friction with human nature. Oh, absolutely. It clashes with our privacy and it completely disrupts our old habits. Let's talk about those everyday habits first. Specifically, the frustrating issue of prompt drift. GPT 5 .5 and Opus 4 .7 just released completely new architectures. And they completely broke everyone's old prompting habits overnight. The way you talk
to the machine has to change. I still wrestle with prompt drift myself. You spend months learning how to talk to a specific model. You find the absolute perfect phrasing to get what you want. Right, you dial it in perfectly. And then they seamlessly update the software on the back end. Suddenly, your carefully crafted prompts feel significantly worse. The output gets lazy or weird. But it's not you. It's the underlying
model changing how it interprets language. The rules of engagement simply shift overnight without any warning. The cost equations are also shifting incredibly rapidly right now. A new evaluation tool just launched called Best Value AI 2026. It actively compares over 37 different language models. It looks across APIs, subscriptions, and different hardware setups. And for those wondering, APIs are just software bridges connecting different computer programs together. Right.
The tool ranks all these models by quality adjusted cost per token. And tokens are simply the small chunks of words an AI reads. It runs real tests to find the actual value. Because a cheap model is useless if the answers are terrible. Exactly. But the friction goes way beyond cost and prompting habits. It is bleeding into deep, deeply uncomfortable privacy issues. Look at what just happened over at Meta. Meta is facing massive backlash right now. They are under heavy public scrutiny regarding
their hardware. Yeah, reports just came out about their Ray -Ban Meta smart glasses. Workers actually reviewed sensitive naked footage captured from users' glasses. The staff members involved ultimately lost their jobs over it. But the privacy line was already crossed. The cameras are always watching. It perfectly shows how technological capabilities have completely outpaced our social norms. We wear cameras on our faces, but we don't understand the risks. And our legal frameworks are really
struggling to keep up. Copyright law is also buckling heavily under the pressure of AI scraping. Casey Green is considering legal action right now. He is the original creator of the famous this is fine dog meme. It's a massive piece of Internet culture. An AI startup called Artisan basically used his art. They generated it and put it in their ads without his consent. It is a constant ongoing battle over ownership and training data. Artists are fighting to keep their
work out of the machine. And then there's the emotional friction. This might honestly be the strangest part of the whole transition. Teens are forming incredibly deep emotional ties to chatbots. Psychologists and experts are actually stepping in now. They recommend five specific questions you should ask your kids. You need to gauge how they are interacting with these AI companions. People are genuinely treating them like close friends or even romantic partners.
It's fundamentally changing human socialization. Even the tech leaders themselves are treating them differently now. Sam Altman recently tested the brand new GPT 5 .5 model. He asked the AI to plan its own launch party, right? He asked it what kind of celebration it wanted. Yeah. And the results were oddly specific. The AI provided highly structured, almost uncomfortably human -like answers about the vibe and the guest list. And he is actually going to do it. He is throwing
the exact party the AI designed for itself. Two sec silence. With teens forming attachments and AI throwing its own parties, are we anthropomorphizing these models too much? Well, these systems are explicitly designed to mimic empathy perfectly. It makes human attachment an inevitable, complicated feature of the interface, not a bug. We are wired to love anything that talks back to us. We really are. It hijacks our basic social programming entirely. Get it? If we step back and look at
the big picture today. We are watching a fundamental transition in computing history. AI used to be an expensive centralized tool. You just type text into a simple box. Right. It was distant. It was entirely contained within the web browser. Exactly. But now it is becoming a decentralized, highly capable video entity. It literally watches your breathing in the exam room. It automates your entire business workflow in the background. And it constantly breaks your old prompting habits
as it evolves. Thank you for joining us on this deep dive. We want to leave you with a final thought to mull over. A 22 -year -old can unleash a super intelligence from a laptop. An AI can act as a real -time medical teammate with a perfect bedside manner. What happens to the value of human intuition in a world where logic and empathy can both be downloaded from GitHub? Outero Music.