Latest / AI Fire Daily / 🎙️ EP 260: Meet OpenMythos, The 22-Year-Old Who Leveled the AI Playing Field
Transcript
- 0:00Anthropic locked their most dangerous AI in a
- 0:03high -security vault. A 22 -year -old dropout
- 0:07cracked it open. He gave it to the world in under
- 0:09two weeks. The digital walls are coming down
- 0:12incredibly fast right now. Welcome to the Deep
- 0:15Dive. Today, we are exploring a massive shift.
- 0:18The balance of technological power is fundamentally
- 0:21changing. We are bringing you the latest signals
- 0:23from the frontier. Right. We're looking at how
- 0:26a single GitHub repository is threatening a trillion
- 0:29dollars of big tech infrastructure today. We're
- 0:33also looking at your doctor's office. AI is officially
- 0:36stepping in as an active medical teammate. And
- 0:38finally, we'll explore the messy, deeply human
- 0:41friction happening as these models integrate
- 0:44into our daily lives. It's a fascinating moment
- 0:46in time. Let's start with this idea of protecting
- 0:49AI. Entropic had this incredibly powerful system
- 0:52called Claude Mythos. They actually decided it
- 0:55was simply too dangerous for the public. Yeah,
- 0:56they wanted to keep it strictly contained. They
- 0:58launched something they called Project Glasswing.
- 1:00Right. The entire goal was to secure the model
- 1:03completely. They were carefully vetting all their
- 1:06corporate partners. They essentially built this
- 1:08massive digital moat to keep it safe. They wanted
- 1:12absolute control over the deployment. You build
- 1:15a vault and you think the vault is secure. But
- 1:18that protection just failed spectacular. It really
- 1:21did. Enter a developer named Kai Gomez. He's
- 1:24a 22 -year -old college dropout. He just looked
- 1:27at the architecture and cracked it. I mean, he
- 1:29reverse engineered the entire Claude Mythos system
- 1:31right from the outside. And he released his own
- 1:34open source clone. He called it Open Mythos.
- 1:36He did this in less than two weeks. Beat. I just
- 1:39think about the historical irony here. Oh, it's
- 1:41massive. Massive corporations build these giant
- 1:44impenetrable digital castles and then... A kid
- 1:48just invents a ladder. A very, very efficient
- 1:50ladder. The unfiltered version is now effectively
- 1:52out in the wild. This is an absolute nightmare
- 1:55for AI regulators everywhere. Yeah, I bet. Terrifies
- 1:58the corporate moat defenders across the industry.
- 2:01So why is it so terrifying for them strategically?
- 2:03What did this kid actually figure out? It comes
- 2:06down to a profound architecture shift. Older
- 2:09AI models use massive flat layers to process
- 2:13data. Okay. They process information sequentially
- 2:16through billions of different parameters. Let's
- 2:19define parameters really quickly for everyone.
- 2:21Parameters are basically the tiny mathematical
- 2:23dials controlling how software thinks. Exactly.
- 2:26And to make an AI smarter, you usually just add
- 2:29more layers. You add more dials. Right. But that
- 2:32requires a massive amount of computing power.
- 2:36OpenMythos fundamentally shifts how the AI actually
- 2:38operates. It uses a very small subset of layers
- 2:41instead. Then it loops the data through them
- 2:44repeatedly. So it reuses the exact same digital
- 2:46pathways over and over. It doesn't need to travel
- 2:49down a massive flat line. Yes, that is the brilliant
- 2:52part of the design. It is exactly like stacking
- 2:55Lego blocks of data. You build incredibly complex
- 2:58towering structures from a few simple repeating
- 3:00pieces. Wow. It is highly efficient and it's
- 3:02incredibly fast. Whoa. Imagine scaling to a billion
- 3:05queries. If it's that efficient, you could run
- 3:07frontier intelligence on consumer hardware. Yep.
- 3:09You wouldn't even need a massive server farm.
- 3:11And that is the trillion dollar problem for big
- 3:14tech right now. We literally just saw big tech
- 3:17commit one. trillion dollars to infrastructure.
- 3:20It's staggering. They are building massive power
- 3:23hungry data centers all over the world. Just
- 3:26massive physical warehouses filled with extremely
- 3:30expensive computer chips. They are pouring concrete
- 3:33and laying copper wire everywhere. Yes. And the
- 3:36financial markets are aggressively backing this
- 3:38approach to KKR just raised 10 billion dollars
- 3:42for a new fund. They are backing a massive new
- 3:45A .I. infrastructure company led by Adams. Lipsky,
- 3:48the former AWS chief. Exactly. They are focusing
- 3:51entirely on data centers, power, and connectivity.
- 3:54They're building alongside the massive cloud
- 3:56computing companies. They are betting everything
- 3:58on centralized power. To support the constantly
- 4:01rising AI demand. Right. But here's the massive
- 4:04catch in their strategy. What if small and loopy
- 4:07is fundamentally better architecture? What if
- 4:10it simply beats big and flat forever? Then that
- 4:13massive hardware investment suddenly hits an
- 4:15efficiency wall. You don't need a massive data...
- 4:17center if a laptop works. It's a catastrophic
- 4:19efficiency wall for those companies. Gomez is
- 4:22actively building a decentralized version of
- 4:25superintelligence. Meanwhile, Meta and Google
- 4:28are spending billions to centralize it. The next
- 4:31world -changing breakthrough probably won't come
- 4:33from a boardroom. It will likely come directly
- 4:36from a GitHub repository. Two sec silence. Is
- 4:39Big Tech's hardware advantage officially dead?
- 4:42It's not totally dead, but it's deeply vulnerable
- 4:45now. Proprietary hardware is no longer an absolute
- 4:48shield. Self -taught developers can clone frontier
- 4:51architectures in days. So raw computing power
- 4:54can't protect them from smarter, leaner algorithms.
- 4:57Mid -role sponsor, Reed. We just talked about
- 5:00how these powerful, lean models are built. Now
- 5:02let's look at where they are actually going.
- 5:03They are actively sitting in the room with us
- 5:05today. Yeah, they really are. They are diagnosing
- 5:07our bodies and running our corporate workflows.
- 5:09We are letting them into our most intimate spaces.
- 5:11Google DeepMind just unveiled something called
- 5:14the AI Co -Clinician. This is not just a text
- 5:17box on a screen anymore. It is a real -time,
- 5:19video -capable teammate in the exam room. It
- 5:22is designed to sit in on your actual doctor appointments.
- 5:25It literally watches the interaction unfold live.
- 5:29DeepMind is calling this new concept Triadic
- 5:31Care. Triadic meaning a core group of three.
- 5:35The patient, the doctor, and the AI working together.
- 5:39Beat. I mean, the AI isn't textbook anymore.
- 5:42It is a hyper observant medical resident who
- 5:45never blinks. That's a really great way to visualize
- 5:47it. It actually watches your breathing during
- 5:50the physical exam. It monitors how your chest
- 5:53rises and falls through the camera. Wow. It even
- 5:56checks your inhaler technique on video. Right.
- 5:59Make sure you're holding it at the exact right
- 6:00angle. Exactly. It actively helps your doctor
- 6:03spot things the human eye might miss. And to
- 6:05achieve this safely, it operates using two very
- 6:08distinct modules. Okay. First, there is the talker
- 6:11module. And then that is the part that actually
- 6:13interacts with the patient. It retrieves the
- 6:14medical evidence and speaks in the room. Yes.
- 6:16But the second part is the real engineering breakthrough
- 6:19here. It is a completely silent overseer module
- 6:21running in the background. What does the overseer
- 6:24actually do mechanically? How does it prevent
- 6:27the AI from just making things up? It constantly
- 6:30watches the talker module before it speaks. It
- 6:33is specifically looking for any unsafe medical
- 6:36advice. So every piece of advice gets double
- 6:38-checked instantly. It is cross -referenced directly
- 6:41against clinical -grade evidence. Then it is
- 6:43cited in real time for the human doctor to verify.
- 6:46The test results on this system are honestly
- 6:48staggering. They tested it on 98 highly realistic
- 6:52primary care queries. The system logged zero
- 6:55critical errors in 97 of them. Wait, zero errors?
- 6:59It actually beat every other frontier model on
- 7:01open -ended drug questions. That even includes
- 7:03the brand new GPT -5 .4. It did. And the video
- 7:07simulations they ran are even more impressive
- 7:09to watch. The AI correctly guided patients through
- 7:12complex physical shoulder maneuvers. It helped
- 7:15them isolate and find specific rotator cuff injuries.
- 7:18It even helped diagnose a case of myasthenia
- 7:20gravis. It did this purely through video observation
- 7:23of the patient's face. In blind clinical tests,
- 7:26physicians vastly preferred this AI teammate.
- 7:29They chose it over their current diagnostic tools
- 7:32by 67 % to 26%. But the human element is still
- 7:36vital here. The study clearly found that experienced
- 7:38physicians still beat the AI overall. The human
- 7:42intuition combined with experience still wins
- 7:45out. For now, yes. But this triad concept is
- 7:48also rapidly entering the corporate world. OpenAI
- 7:51just launched a product called Codex for Work.
- 7:54It is designed specifically for enterprise teams.
- 7:56Aaron Livy over at Box is already hiring agent
- 7:59engineers. These are specialized people who build
- 8:02AI to autonomously do tasks. He wants to automate
- 8:05core business workflows entirely. Two sec silence.
- 8:08How does the system prevent the AI from giving
- 8:11confident but deadly medical advice? It uses
- 8:13that critical silent overseer module. This module
- 8:16watches the main system and cross -references
- 8:19advice against clinical grade evidence in real
- 8:21time. It basically has an internal chaperone
- 8:23fact -checking every single word. Exactly. The
- 8:26architecture is designed entirely around patient
- 8:28safety first. Which brings us to the messy part
- 8:30of this transition. Integrating agents that can
- 8:34see us and talk to us is incredibly jarring.
- 8:37It creates immediate painful friction with human
- 8:40nature. Oh, absolutely. It clashes with our privacy
- 8:42and it completely disrupts our old habits. Let's
- 8:45talk about those everyday habits first. Specifically,
- 8:48the frustrating issue of prompt drift. GPT 5
- 8:52.5 and Opus 4 .7 just released completely new
- 8:55architectures. And they completely broke everyone's
- 8:58old prompting habits overnight. The way you talk
- 9:01to the machine has to change. I still wrestle
- 9:03with prompt drift myself. You spend months learning
- 9:06how to talk to a specific model. You find the
- 9:09absolute perfect phrasing to get what you want.
- 9:11Right, you dial it in perfectly. And then they
- 9:13seamlessly update the software on the back end.
- 9:15Suddenly, your carefully crafted prompts feel
- 9:18significantly worse. The output gets lazy or
- 9:21weird. But it's not you. It's the underlying
- 9:23model changing how it interprets language. The
- 9:26rules of engagement simply shift overnight without
- 9:28any warning. The cost equations are also shifting
- 9:31incredibly rapidly right now. A new evaluation
- 9:33tool just launched called Best Value AI 2026.
- 9:37It actively compares over 37 different language
- 9:40models. It looks across APIs, subscriptions,
- 9:43and different hardware setups. And for those
- 9:45wondering, APIs are just software bridges connecting
- 9:47different computer programs together. Right.
- 9:50The tool ranks all these models by quality adjusted
- 9:52cost per token. And tokens are simply the small
- 9:56chunks of words an AI reads. It runs real tests
- 9:59to find the actual value. Because a cheap model
- 10:03is useless if the answers are terrible. Exactly.
- 10:05But the friction goes way beyond cost and prompting
- 10:08habits. It is bleeding into deep, deeply uncomfortable
- 10:11privacy issues. Look at what just happened over
- 10:14at Meta. Meta is facing massive backlash right
- 10:17now. They are under heavy public scrutiny regarding
- 10:19their hardware. Yeah, reports just came out about
- 10:21their Ray -Ban Meta smart glasses. Workers actually
- 10:24reviewed sensitive naked footage captured from
- 10:27users' glasses. The staff members involved ultimately
- 10:29lost their jobs over it. But the privacy line
- 10:32was already crossed. The cameras are always watching.
- 10:34It perfectly shows how technological capabilities
- 10:37have completely outpaced our social norms. We
- 10:40wear cameras on our faces, but we don't understand
- 10:43the risks. And our legal frameworks are really
- 10:46struggling to keep up. Copyright law is also
- 10:49buckling heavily under the pressure of AI scraping.
- 10:53Casey Green is considering legal action right
- 10:55now. He is the original creator of the famous
- 10:58this is fine dog meme. It's a massive piece of
- 11:02Internet culture. An AI startup called Artisan
- 11:04basically used his art. They generated it and
- 11:08put it in their ads without his consent. It is
- 11:10a constant ongoing battle over ownership and
- 11:13training data. Artists are fighting to keep their
- 11:16work out of the machine. And then there's the
- 11:18emotional friction. This might honestly be the
- 11:20strangest part of the whole transition. Teens
- 11:22are forming incredibly deep emotional ties to
- 11:25chatbots. Psychologists and experts are actually
- 11:27stepping in now. They recommend five specific
- 11:29questions you should ask your kids. You need
- 11:32to gauge how they are interacting with these
- 11:34AI companions. People are genuinely treating
- 11:36them like close friends or even romantic partners.
- 11:39It's fundamentally changing human socialization.
- 11:42Even the tech leaders themselves are treating
- 11:44them differently now. Sam Altman recently tested
- 11:47the brand new GPT 5 .5 model. He asked the AI
- 11:50to plan its own launch party, right? He asked
- 11:53it what kind of celebration it wanted. Yeah.
- 11:55And the results were oddly specific. The AI provided
- 11:58highly structured, almost uncomfortably human
- 12:01-like answers about the vibe and the guest list.
- 12:04And he is actually going to do it. He is throwing
- 12:06the exact party the AI designed for itself. Two
- 12:10sec silence. With teens forming attachments and
- 12:13AI throwing its own parties, are we anthropomorphizing
- 12:16these models too much? Well, these systems are
- 12:18explicitly designed to mimic empathy perfectly.
- 12:21It makes human attachment an inevitable, complicated
- 12:24feature of the interface, not a bug. We are wired
- 12:27to love anything that talks back to us. We really
- 12:29are. It hijacks our basic social programming
- 12:31entirely. Get it? If we step back and look at
- 12:34the big picture today. We are watching a fundamental
- 12:36transition in computing history. AI used to be
- 12:39an expensive centralized tool. You just type
- 12:42text into a simple box. Right. It was distant.
- 12:45It was entirely contained within the web browser.
- 12:47Exactly. But now it is becoming a decentralized,
- 12:50highly capable video entity. It literally watches
- 12:54your breathing in the exam room. It automates
- 12:56your entire business workflow in the background.
- 12:58And it constantly breaks your old prompting habits
- 13:01as it evolves. Thank you for joining us on this
- 13:03deep dive. We want to leave you with a final
- 13:05thought to mull over. A 22 -year -old can unleash
- 13:08a super intelligence from a laptop. An AI can
- 13:11act as a real -time medical teammate with a perfect
- 13:13bedside manner. What happens to the value of
- 13:16human intuition in a world where logic and empathy
- 13:18can both be downloaded from GitHub? Outero Music.