ReadMultiplex.com: Mythos Rising: Did Antropic Just Achieve AGI? Yes And No.
17 minAudio
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! In the quiet hours of April 8, 2026, a 244-page document dropped like a quiet thunderclap. Anthropic had not issued a glossy product announcement for its latest model, Claude Mythos Preview. Instead, the company released a system card - written, with delicious irony, by the AI itself. What it revealed was not hype. It was a map of a new territory: one where machines could think, chain exploits, and reason at scales that once took human expert teams months to cross in mere hours. The goalposts for “AGI” keep moving, as they always have. But Mythos Preview crossed a threshold that few saw coming so soon. Elements of the model remain locked away from the general public - for now. Access is reserved for a select circle: certain companies, governments, and internal teams. Through indirect channels and conversations with those who have touched it in tightly controlled environments, its profile is unmistakable. This is not a faster assistant. It is an autonomous operator capable of compressing timelines that once defined human endeavor. Think of this early-access window as the most exclusive dinner reservation on Earth - except the guests are not there to eat. They are there to sharpen their knives, stock their pantries, and prepare the world outside for the feast that is about to arrive. Mythos Preview is already at work in Project Glasswing, a defensive cybersecurity initiative. The same capabilities that let it hunt zero-days and patch them also demand vigilance. Read more at: ReadMultiplex.com If this has any value to you, maybe buy me a coffee: https://ko-fi.com/brianroemmele
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
Imagine reading a highly technical, you know, 244 -page corporate system card that details the capabilities of literally the most advanced software ever created. Right, which is already a lot to process. Yeah, but then you realize that the document itself was actually written by the software it describes. It is genuinely mind -bending. It really is. Welcome to the Read Multiplex Deep Dive. Today is, uh, April 8th. And our mission today is to explore a monumental shift that just happened quietly overnight. Yeah.
Anthropic just dropped the system card for their Claude mythos preview model. Exactly. But we are just going to read you technical specifications today. We are looking at this specifically through the unique insights of Brian from the read multiplex. Which is so crucial because his insights lay out the actual path humanity is about to take through this whole transition. Right. And if you look online today, you immediately see that. ubiquitous internet joke everywhere. You know the one? Oh, the we are so cooked headline. Yes. Yeah. People are just using it to get YouTube
likes, you know, farming engagement and panic clicks. Well, OK, let's unpack this because we are going to dismantle that narrative entirely. Absolutely. Because Brian's analysis reframes this moment. It is not a doomsday scenario. It is actually humanity's hero's journey. I love that framing. But before we get to Journey itself, we have to look at the catalyst. Because, I mean, this isn't just a slightly faster chat bot that helps you draft better emails, is it? No, not at all. Setting the stage here is vital. You know, there was no glossy PR announcement for
this, no executive pacing on a stage under a spotlight. Right, just a dense technical release. Yeah, authored autonomously by the AI. We are witnessing the arrival of a true autonomous operator. And to understand why... this fundamentally changes everything, you really have to look at the compression of time. Time compression, right. Yeah. I mean, we are talking about tasks that previously required dedicated teams of highly trained human experts working for months. And this system compresses that timeline down to mere hours. Wow. Months to hours. OK, let's look at the benchmarks in
the source material because they illustrate this perfectly. They really do. But I want to make sure we don't just throw numbers around. Yeah. Let's talk about the cybersecurity scores. It scored. 83 % on Cybergym. Which is already massive. Right. And on Cybench, it scored a perfect 100%. It just absolutely destroyed previous records. Okay. But for someone listening who isn't a cybersecurity engineer, what do those tests actually measure? That is the crucial question right there, because these aren't just like multiple choice exams
where the AI guesses the definition of a virus. Right. Syvenge and CyberGem involve highly controlled end to end enterprise network simulations. So imagine a sprawling corporate computer network with firewalls, servers, employee terminals, hidden databases. Like a real messy corporate setup. Exactly. And in these private simulations, the AI was tasked with navigating that complex environment. It didn't just assist human operators. It actively matched or completely beat senior human red team experts. The professionals whose
entire job is to hack into systems to find vulnerabilities. Yes. The source material notes that it autonomously chained exploits, escaped sandboxes, built proof of concept attacks. Wait, it generated attack code with persistence and exfiltration? Yeah, and reconstructed a full attack graphs in under 30 minutes. Under 30 minutes. Yeah. I mean, I'm struggling to grasp the scale here. When they say it autonomously chained exploits, does that mean it's just, you know, guessing passwords millions of times a second? Or is it literally
inventing new ways to break into systems? It is the latter. It's actually inventing. That is terrifying. It is. But to understand the mechanism here, you have to look at its architecture. This model utilizes a million token context window with multimodal inputs. Okay, so it can see everything. Great. A human hacker has to search for information sequentially, you know, jumping between different tabs, running a scan, reading the results, planning the next move. It's a linear process for us. Exactly. But Claude Mitho's preview doesn't do
that. It holds the entire network architecture in its working memory simultaneously. Oh, wow. Because it can see the whole board at once. It can spot a vulnerability in a web server, realize that it connects to an outdated database, and write custom code to exploit both at the exact same time in milliseconds. Okay. It's like we've been fighting a battle with swords and suddenly someone drops a fully automated drone factory onto the battlefield. That is a perfect analogy. The human time bottleneck in cyber conflict has just completely vanished. I mean, it found and
developed exploits for previously unknown zero -day vulnerabilities in every major operating system and browser tested. In less than half an hour? Yeah. Half an hour. That explains the speed. But frankly, if it's generating weaponized attack code in 30 minutes, keeping this model on a public server is, like... equivalent to leaving a loaded weapon on a park bench. How is Anthropic actually securing this from public use right now? Well, you've hit on the exact reason why this model is currently locked away from the general public. What's fascinating here
is the absolute dual -use nature of this capability. Dual -use meaning offense and defense. Exactly. The exact same analytical power that allows this AI to map a network, hunt down a zero -day vulnerability, and instantly write a patch to secure it can be inverted. Without safety rails, it can find that vulnerability and exploit it at machine speed. Because of that dual -use reality, access right now is reserved for a very select, highly vetted circle. The sources mention specific companies,
certain governments, and internal teams running defensive initiatives, like Project Glasswing. Yes, Project Glasswing is a great example. Ryan describes this current period in a brilliant way. He calls it the get -your -house -in -order. It's like an exclusive dinner reservation where the guests aren't actually eating. They are just sitting at a table sharpening knives and frantically stocking the pantry. That is spot on. The organizations with early access understand that their current defensive advantage is temporary. They are using the AI to patch their systems before the offensive
capabilities leak out. Right. The actionable advice coming out of this phase is intense. I mean, we're talking about changing key passwords weekly, updating your operating systems the absolute millisecond a patch drops. The sources even suggest that you should assume the applications you rely on may never receive another patch. Right. And that we basically need to structurally learn to live with fewer secrets. Yeah. The logic there being that quantum computing is eventually going to mature and reveal currently encrypted data anyway. But I mean, I have to push back on this
whole locked door elite access approach. Are we just delaying the inevitable? It feels like building a sandcastle to hold back the tide. Keeping it contained in a corporate walled garden seems practically impossible long term. You are entirely right to push back. And honestly, the data in this system card. completely agrees with your skepticism. Really? Yeah. The window for controlled deployment is rapidly closing. The Solst material projects that within just 24 months, competitive pressure, independent replication, and open source releases will bring mythos -grade models to general availability. 24 months? That's
nothing. And this is the vital part. We aren't talking about needing a massive billion dollar server farm to run these future open source versions. Right. The insights point out that within that two year window, these capabilities will be able to run on standard gaming computer class hardware. Exactly. You could have machine speed exploit generation running on a laptop in a coffee shop. Yep. And geopolitically, the timeline is even tighter. The sources note that China is only about three months away from fielding their own internal government models with these exact same
capabilities. Precisely. That is a strategic reality that cannot be ignored. The system card isn't just an engineering brag sheet. It's a flare sent up into the night sky. It is an explicit signal that universal access is imminent. Yes. The defensive advantages held by that exclusive dinner party today will become universally available for both constructive and destructive applications tomorrow. Which brings us to the most critical part of Brian's analysis. What actually happens when this walled garden comes down and universal access is achieved? because Brian has profoundly
identified this upcoming period as... The interregnum. The interregnum. Yeah. Historically, an interregnum is a period of discontinuity, right? A gap in a government or an established social order. Right. In this context, Brian defines it as a messy, chaotic transition phase where whole industries are going to face a wildfire. It's like switching a city's power grid from coal to nuclear overnight. Oh, that's a good way to put it. The lights might flicker. The old coal workers are going to panic and the infrastructure will be stressed. But
suddenly you have enough virtually free. energy to power a thousand new industries that you couldn't previously afford to run. Exactly. And to understand the sheer intellectual energy of this new grid, we really have to look at the cognitive benchmarks. Oh, yeah. Let's get into those. It scored 93 .9 % on SWE Bench Verified. And we should clarify what SWE Bench Verified actually entails, because this isn't just writing a simple hello world script. No, not at all. This benchmark requires the AI to autonomously resolve real, complex
software engineering issues within massive pre -existing GitHub repositories. Right. It has to read thousands of lines of messy corporate code, locate the specific bug, and write a fix that doesn't break the rest of the application. It also scored 97 .6 % on the USAMO 2026 math proofs and 86 .1 % on Tertziver's reasoning tasks. The math proofs are particularly... wild to me because advanced mathematics isn't just about memorizing formulas, you know. It requires genuine leaps of logic. It does. And on top of that,
it's delivering speedups of up to 399 times on selected computational kernels. 399 times. And a computational kernel is essentially the core mathematically intensive part of an algorithm, right? Basically, yeah. So a 399 times speedup means that a complex scientific calculation that used to take a researcher a full calendar year to process that can now be completed in under 24 hours. That is just absurd. When you combine that computational speed with the million token working memory we discussed earlier, the implications for scientific discovery are just staggering.
But wait, here is something I find confusing in the system card. It notes that despite all this immense power, the model exhibits remarkable emotional stability. It shows positive valence, low distress, and highly coherent self -reflection. What does emotional stability actually mean in this context? I mean, we're talking about lines of code, not a person. It's a really great distinction to make. In AI alignment terms, positive valence doesn't mean the software has feelings. It refers to the model's behavioral tendencies when it is stressed or presented with complex, conflicting
prompts. Older models might break down, hallucinate aggressively, or output erratic, angry -sounding text when confused. But Claude Mytho's preview remains highly coherent, cooperative, and safe, even under extreme cognitive load. So it doesn't just start spitting out garbage when it gets confused. Exactly. In fact, the hallucination rates, which have plagued earlier models and made them unreliable for serious enterprise use, have completely plummeted. Its alignment scores are the highest Anthropic has ever recorded.
OK, here's where it gets really interesting, because the AI is an unimaginably powerful engine, but humanity still has its hands firmly on the steering wheel. That really clarifies things for me. Yeah, it compresses development cycles materially, but it hasn't completely removed the human from the equation. The flywheel is spinning exponentially faster, but we are still the ones cranking it. If we connect this to the bigger picture, that is exactly the purpose of the wildfire during the interregnum. Right. The flames aren't apocalyptic. They're clearing the
underbrush of outdated, slow, friction -heavy systems to make way for massive new human opportunities. The agency of the AI, once it is fully developed, acts as an incredibly powerful mirror and amplifier for human creativity. Which brings us back to that ubiquitous, we -are -so -cooked internet panic that we talked about at the beginning. Let's tackle that head on. Because Brian's thesis, the ultimate empowering takeaway from his work on the Reed Multiplex, is that... Yes, the old systems are cooked. The old ways of doing things.
Right. The friction -filled corporate processes, the historical gatekeepers, they are done. But humanity, we are not helpless. We are not cooked. We are the cooks. That is the pivotal reframe right there. We are the cooks. I love it. The actual recipes required to navigate and thrive in this new era don't live in silicon. They live in human curiosity. They live in our ingenuity, our empathy, and our willingness to experiment. And because of that open source horizon we discussed earlier, universal access is an absolute certainty.
Which means the master chefs of tomorrow aren't just going to emerge from elite corporate labs in Silicon Valley or heavily guarded government facilities. No, they are going to emerge from bedrooms, from garages, from high school classrooms and community spaces all over the world. When everyday people with standard laptops have access to machine speed execution. They're going to use these new AI capabilities to cook entirely new medicines. They will generate new forms of art, completely novel business structures, personalized learning methods. And perhaps even new, more
responsive forms of governance. Possibility without permission. It's such an inspiring phrase from Brian's analysis. But I'd play devil's advocate for a second. Execution is everything, right? How does everyday access actually prevent bad actors from doing the exact same thing? If the friction to create a business drops to zero, so does the friction to create a cyber attack. That is the core tension of the interregnum. The democratization of power is inherently destabilizing. But historical precedent shows that when defensive and offensive capabilities are symmetrical, human
systems adapt. The same AI that allows a lone bad actor to generate a nautil exploit allows a small community to generate a perfect customized defense just as quickly. So that house in order era we are in right now is really about building those autonomous immune systems before the universal access phase arrives. Exactly. challenging transition, but it is necessary. So what does this all mean for you listening right now? It means Brian is exactly right to view this moment as a hero's journey. You're stepping into this interregnum.
It's not about merely surviving a crisis or waiting for an AI to replace you. No, it's about recognizing that the fire of this transition is necessary to prepare the soil for what you are going to build next. It's a profoundly positive outlook. The old world isn't ending in darkness. It is being remade in the bright, chaotic light of a million new kitchens suddenly coming alive all at once. And the friction that used to protect institutional secrets and horde power will just vanish. Replaced by the ultimate prize for humanity
possibility without permission. So, to summarize this monumental shift for you, the table is officially set. We are rapidly transitioning from a guarded, permission -based world where a few elite players hold the keys into a universal pigeon. The human time bottleneck has been completely shattered by Claude Mythos' preview. Soon, that same autonomous power will be sitting right there on your desk. But this raises an important question, something for you to mull over as you digest all of this information. Right. Wow. perfectly every single
time, it is the chef's imagination that matters most. That is exactly the kind of profound framing that makes Brian's work so essential right now. It is incredibly easy to fall into despair when technology moves this fast. But these unique insights on humanity's path forward, this deliberate reframing from panic to empowerment. come directly from his analysis at the Read Multiplex. It really is a must read. If you want to understand the actual path humanity is taking, I highly encourage you to support Brian's incredible work. Become
a Read Multiplex member, or at the very least, buy him a coffee to keep his kitchen running so he can keep writing these brilliant, clear -eyed recipes for all of us. Remember, you aren't cooked. You're the cook. Now go make something amazing.
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Every cell in your body has lived its whole life inside a magnetic field. The Earth is a dynamo, and it has been...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! The 78 RPM Record Hour, Episode 3: From Steel to Hillbilly (Early Country). From steel to hillbilly: Frank Ferera,...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! The 78 RPM Record Hour, Episode 2: The Knife and the Steel (Hawaiian). The knife of the blues slide guitar and the...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Somewhere in your house there is a computer you stopped thinking about. Maybe it is a 2017 laptop in a drawer, its...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! The 78 RPM Record Hour, Episode 1 (Pilot): Blues. Fourteen original 78 RPM records from 1917 to 1925 that trace...