Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

ReadMultiplex.com Podcast.

ReadMultiplex.com Podcast. artwork

ReadMultiplex.com: METR: The New AI Measurement Gatekeepers and the Quiet Path to State Control of AI

Sep 14, 202657 minAudio
ShareXLinkedInFacebookWhatsAppEmail

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! For over a decade a small group have planned to become the central authority to “protect you” from AI. It sounds like a bad movie script, but understand that when you read this, all of this is taking place. So let's break down the mechanics of the public script, because once you see the pattern, I swear you cannot unsee it. It's everywhere once you look not with conspiracy but with evidence. Step one. Name an unbounded apocalyptic catastrophe. Step two. Attach a very short, very terrifying countdown clock to it. Step three, demand centralized political permission to build or continue development in that sector. And step four, this is the kicker, when the catastrophe inevitably doesn't arrive on schedule, you don't apologize. You just move the clock and ask for more money. It feels exactly like a doomsday cult. You know, like the people who gather on a mountaintop all wearing Nike sneakers on a Tuesday because the leader said a spaceship is coming to destroy the Earth. And then Wednesday morning rolls around. The sun comes up, no spaceship. But they don't disband. They don't admit they were wrong. The leader just says, our unified prayers bought us a brief window of time. The spaceship is actually coming next month, but we need stricter rules and larger donations to ensure we survive. This mountaintop analogy perfectly captures the psychological trap here. When a group's identity, and frankly their financial structure, are tied to an impending apocalypse, a failed prediction doesn't result in them re-evaluating their core belief. It results in them doubling down on the urgency. And this isn't just an AI phenomenon, this lens examines the historical base rates of these exact types of apocalyptic predictions. The clocks that would not pay.And the methodology of how these historical predictions failed is identical to how the current AI doom predictions will fail. We look at the trail of failures that destroyed generations it starts with a massive cultural touchstone from 1968, Paul Ehrlich's book, The Population Bomb. Ehrlich, a biologist, predicted that the battle to feed all of humanity was already over. He said that in the 1970s and 80s, hundreds of millions of people would starve to death regardless of any crash programs embarked upon at that time. And people just accepted this as fact. It was treated as a mathematical, scientific inevitability by the press and policymakers. We drive past the “No Nukes” anti enegery deception that sent the US on to a deeper connection to petrochemical energy. None of this is new, the new mask goes on the same corpse of dead end ideology. Read more at: ReadMultiplex.com Support this work by buying us 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 waking up, you grab your phone, and you're just sort of mindlessly scrolling through your feed. Maybe we all do, yeah. Exactly. And suddenly you see that a highly respected researcher, a guy from one of those powerful tech companies in the world, has just resigned. But not just, you know, quietly stepping down. Right. Not a quiet spending more time with my family post. He drops an absolute bombshell. He claims that the artificial intelligence models being built right now in these secret labs have a greater than 10 % chance of just... well, wiping out

humanity by the end of the decade. Which is an insane thing to read before you've even had your coffee. Totally. And 100 million people read this. The news cycle just completely catches fire. Panic everywhere. And I think the immediate instinctual reaction for most of us is just pure terror. Of course it is. I mean, you think if the engineers building this stuff are terrified, shouldn't you be? Exactly. But, and this is the core of our deep dive today, what if that panic wasn't some spontaneous organic accident? What if it was a highly coordinated, perfectly timed

press of a button? A button designed to basically hand the keys of the future over to a handful of billionaires. Yes. And that question. completely reorients how you look at the entire tech industry. Because when we see widespread viral fear, we just naturally assume it's organic, right? Right. But what we're exploring today is really the architecture of that fear, like the actual mechanics of how panic is engineered, how it's packaged, and then leveraged to get very specific real world political outcomes. Which brings us to

the source material for this deep dive. We are unpacking this sprawling, intensely detailed manifesto and investigative report. It's titled ATR, the measurement gatekeepers and the quiet path to state control of frontier AI. It's a massive piece of work. Massive. And the core premise is that the wide open frontier of artificial intelligence is currently being rapidly corralled. It's being heavily monitored and centrally controlled. And the justification being sold to you, the public, is that this is entirely for your own safety and survival. But the author is asking,

you know, the really vital questions here, like who exactly gets to define what safety is? Right. And how is that measurement actually happening? And honestly, perhaps most importantly, who profits immensely when open innovation gets replaced by this licensed, highly regulated oligopoly? Spot on. Now, before we get into the mechanisms of how this is allegedly happening, we have to lay down some crucial context. This is a mandatory impartiality disclaimer for you listening right now. Very important. Yeah. The source material

we are decoding today is highly polarizing. It contains very politically charged claims about the AI industry, national governments, philanthropic organizations and some very specific real world individuals. Right. And we need to be crystal clear that we are not. endorsing these claims. We aren't taking sides on whether these specific individuals are acting with like deep malicious intent or out of a genuine concern for humanity. Exactly. Our job here is strictly to impartially report, decode, and just analyze the deeply intricate

arguments that the author has presented. We're basically your guys through the architecture of this argument. So let's start where the author starts, the spark, the inciting incident I mentioned in the intro. The resignation. Right. September 9th, 2026. A guy named Jacob Coxon, a researcher who spent three years working deep inside the pre -training divisions at OpenAI and then Anthropic, resigns. And he posts this thread on X that just shatters the Internet. Over 100 million views. He essentially claims these top laboratories are racing completely blind toward a self -improving

super intelligence. He paints this picture of a runaway train with no brakes. And instantly you see high level people in the industry like Anthropics alignment lead Evan Hubbinger jumping into the replies to agree with him. Which is wild. Totally wild. They start putting their own personal probabilities of literal human extinction at greater than 10 percent within 10 years. I mean, for a lay person or honestly. even a seasoned tech person, seeing the engineers themselves assigning a double digit probability to human extinction is paralyzing. It creates a complete

media singularity. Wired Magazine, The New York Times, Cable News, they all go into overdrive. But this is where the author of our source material does this brilliant forensic pivot. They ask us to look not at the feeling the thread generated, but at the specifics of the claim. Right. Because in all of this viral panic, there was no specific actionable mechanism described. Yes. When I read the author's breakdown of this, it was a massive light bulb moment for me. They point out that Coxon and the people agreeing with him mentioned

terrifying concepts, right? Like novel pathogen synthesis or hacking global infrastructure. But they never provide a concrete step -by -step engineering pathway that actually leads to the death of literally everyone on Earth. It's always this vague jump from the computer gets very smart to we all die. And that missing link is the entire foundation of the author's critique. The author argues that this viral event wasn't just a lone whistleblower having a sudden crisis of conscience. It was the execution of a well -worn playbook. Exactly. A specific pattern of behavior that

the author calls the public script. And this script has been used repeatedly for decades, long before artificial intelligence was even a functional reality. Let's break down the mechanics of this public script, because once you see the pattern, I swear you cannot unsee it. It's everywhere once you look. Step one. Name an unbounded apocalyptic catastrophe. Step two, attach a very short, very terrifying countdown clock to it. Step three, demand centralized political permission to build

or continue development in that sector. And step four, this is the kicker, when the catastrophe inevitably doesn't arrive on schedule, you don't apologize. You just move the clock and ask for more money. It feels exactly like a doomsday cult. You know, the people who gather on a mountaintop on a Tuesday because the leader said a spaceship is coming to destroy the Earth. And then Wednesday morning rolls around. The sun comes up, no spaceship. But they don't disband. They don't admit they were. wrong the leader just says our unified prayers bought us a brief window of time the spaceship is actually coming next month but we

need stricter rules and larger donations to ensure we survive your mountaintop analogy perfectly captures the psychological trap the author is dissecting here when a group's identity and frankly their financial structure, are tied to an impending apocalypse, a failed prediction doesn't result in them re -evaluating their core belief. It results in them doubling down on the urgency. Exactly. And to prove this isn't just an AI phenomenon, the author pulls back the lens and examines the historical base rates of these exact types of

apocalyptic predictions. They call them clocks that would not pay, right? Yes. And the methodology of how these historical predictions failed is identical to how the author claims the current AI doom predictions will fail. Take us through some of those historical examples because the parallels are staggering. So the author starts with a massive cultural touchstone from 1968, Paul Ehrlich's book, The Population Bomb. Oh, wow. Yeah, that was a huge deal. Massive. Ehrlich, a biologist, predicted that the battle to feed all of humanity was already over. He said that

in the 1970s and 80s, hundreds of millions of people would starve to death regardless of any crash programs embarked upon at that time. And people just accepted this as fact. Totally. It was treated as a mathematical, scientific inevitability by the press and policymakers. He advocated for extreme centralized population control measures. But the 1970s and 80s came and went, and we obviously didn't see hundreds of millions starving in a global collapse. So why did his model fail so spectacularly? Because of a fundamental flaw in the methodology, a flaw that ignores human

adaptation and technological scaling. Ehrlich basically assumed that population grew exponentially, while agricultural output could only grow arithmetically. But he failed to account for the Green Revolution. He didn't factor in Norman Borlaug developing high -yield, disease -resistant wheat varieties. Or synthetic fertilizers becoming widely distributed, right? Exactly. In short, Ehrlich modeled the threat as dynamic and compounding, but he modeled human defensive capabilities as static. Which is a huge blind spot. And this exact same methodological

blind spot was repeated in 1972 with the Club of Rome's famous report, Limits to Growth. Right. I remember reading about that. That was the model predicting we were going to run out of literally everything. Oil, copper, natural gas, even basic calories. And all with highly specific expiration dates attached. Exactly. They ran computer simulations that showed global civilization. collapsing due to resource depletion by the early 21st century. But again, they ignored price signals and human

ingenuity. Because as a resource becomes scarce, its price goes up. Right. And that price increase incentivizes humans to find new reserves or invent more efficient extraction tech or just find entirely new substitutes. So the peak oil forecasts kept sliding from the 1990s to 2000 to 2010. Because the model couldn't comprehend that the system it was measuring would actively adapt to the pressure. And the author doesn't stop there. They bring up the nuclear winter models from 1983. Now, I want to be very careful here, and so does the author. Nuclear weapons are obviously

an existentially serious threat. Nobody is arguing they aren't. But the author focuses on the specific scientific models pushed by Carl Sagan and others in the early 80s, which claimed that even a limited nuclear exchange would loft so much soot into the stratosphere that it would linger for years. Right, plunging the Earth into freezing temperatures, collapsing agriculture, and guaranteeing human existence. extinction. Exactly. The author points out that those early, highly publicized nuclear winter models relied on incredibly extreme worst

case assumptions about how high the soot would loft and how long it would linger. It was a constructed worst case scenario. Yes. Used, as the author explicitly states, as a policy club to force political change. Later natural experiments like the massive oil well fires in Kuwait in 1991 produced immense amounts of smoke, but didn't confirm those extreme loft and linger atmospheric assumptions. So the atmospheric science eventually became much more nuanced. But the initial catastrophic inevitable extinction model was the one used

to demand immediate geopolitical submission. OK, I have to push back here. I really do. Because the author also brings up Y2K in this list of failed clocks. And they imply Y2K. was just another overblown panic. But wait a minute. Y2K was a real tangible software problem. It was. The internal clocks of millions of legacy computer systems were actually going to roll over from 99 to 00. And because memory was so expensive back then, they didn't have the first two digits of the year programmed. Systems were going to break.

Banking software was definitely going to fail. Right. The reason the planes didn't fall out of the sky and the power grid didn't collapse wasn't because the threat was fake. It was because thousands of engineers worked grueling hours for years. to patch the code. Yeah. So couldn't the AI safety advocates argue the exact same thing? Yes, our predictions of doom might not happen, but that's only because our warnings are forcing the industry to patch the AI systems before it's too late. Shouldn't we adhere to the precautionary principle? That is a phenomenal pushback. And it forces us to drill down into

really the most crucial philosophical distinction the author makes in this entire text. Okay, let's hear it. The author explicitly acknowledges that there are real, present, and highly concerning dangers with artificial intelligence. Like deepfakes disrupting elections. Yes, deepfakes, massive waves of automated fraud, the displacement of workers, and the extreme concentration of computing power in the hands of a few megacorporations. These are all... Real measurable problems. Just like Y2K was a real measurable engineering problem.

Exactly. But here is the massive difference between Y2K and the AI extinction narrative. Y2K was falsifiable. Falsifiable. Let's unpack that term for the listener. So falsifiability is the cornerstone of the scientific method, popularized by the philosopher Karl Popper. It means that for a theory to be considered scientific, there must be some inherent way to prove it false. Right. With Y2K, you could literally open the source code of a bank's mainframe, point to the date formatting function, and run a simulation. You could measurably prove, if we run this, it will

crash. It was an observable flaw in a specific bounded system. You could test the patch and prove the patch worked. So what are the AI doomsday advocates proposing that is so different from that? They're proposing an unfalsifiable extinction mechanism. They're claiming that a completely novel, self -improving superintelligence will somehow emerge from predicting the next word in a sequence. Right. And that this entity will autonomously decide to stage a discontinuous coup against humanity by 2027. Which is wild

because there's no source code you can point to for that. Exactly. There's no engineering diagram for a super intelligent coup. There's no measurable code flaw you can point to in a neural network's weights that proves a chatbot will suddenly wake up, desire self -preservation, hack into biological labs, synthesize a pathogen, and kill everyone. It's just a chain of incredibly fragile hypothetical assumptions. Right. Y2K was fixed with engineering patches. AI extinction is being used to demand absolute centralized political authority over global computing resources.

Ah. That is the author's argument. When a claim cannot be falsified, when it relies entirely on a hypothetical future leap in capabilities that breaks all known physical and engineering models, it stops being a scientific threat model. and becomes a permanent argument for a permanent priesthood. A permanent priesthood. That phrasing is chilling. And it's so relevant because the AI industry itself has an embarrassing history of these missed clocks long before Jacob Cox. Oh, absolutely. The author reminds us that back in the 1960s, early pioneers at MIT were confidently

stating that the problem of artificial intelligence would be completely solved within a generation. Didn't quite happen. No. And in the 2010s, we had the whole singularity movement with Ray Kurzweil and others throwing out highly specific dates for when machines would surpass human intelligence, many of which. came and went. And even incredibly recently, the author points to Dario Amadei, the CEO of Anthropic. In early 2025, he made a public claim that AI would be writing 90 % of all code within a year. And when that September 2025 deadline quietly passed, human programmers

were still very much employed and writing code. And this brings us to the author's core thesis for this first section, the methodological tell. This is how you spot the grift, according to the source. Okay, what's the tell? When an expert produces a fat tail probability, meaning a small, say, 5 % or 10 % chance of a massive civilization -ending catastrophe, and they translate that small chance into a very short, urgent calendar deadline. Right. And then when that calendar

deadline inevitably misses, watch what they do next. They do not revise their probability of doom downward. They don't humbly admit that their models of exponential capability were wrong. They actually increase their demand for power. Exactly. They look at the public and say, we narrowly bought ourselves some time, but the threat is actually mutating and getting worse. So you need to give us more regulatory authority. Which is the perfect seamless transition to the next massive piece of this puzzle. Because if the public script is the mechanism and fear is

the product being aggressively sold, who exactly is selling the solution? Right. Who is the priesthood standing at the altar? asking for the ultimate authority. And that brings us to the organization sitting at the absolute center of the author's critique, a group called MTR, spelled M -E -T -R. It stands for Model Evaluation and Threat Research. Yes. MBTR is arguably the most important organization in AI that the general public has never heard of. I certainly hadn't. They present themselves to the world as an independent, rigorous,

scientific nonprofit dedicated to one vital task. measuring whether frontier AI systems pose catastrophic risks to society. Their stated goal is to evaluate whether these massive neural networks can autonomously complete long horizon tasks. Things like autonomously hacking a secure server, replicating their own code across the Internet, or gathering resources without human input. I mean, it sounds incredibly noble. We need referees, right? But the author argues that in practice, METR has quietly and systematically become one of the most powerful,

influential, and un - unofficial regulators in the entire global field of artificial intelligence. Oh, absolutely. When I was reading the origin story of METR, my jaw just hit the floor because it didn't start as some grassroots independent watchdog organization at all, did it? It looks like it was incubated right inside the very ecosystem it is supposed to be regulating. The origin story is deeply entangled. Let's trace it carefully because the author spends a lot of time on this. METR started in 2022. but not under that name.

Right. It began as RRC Evils, which was the evaluation arm of a group called the Alignment Research Center. And the Alignment Research Center was founded by a man named Paul Cristiano. A name we definitely need to remember. Yes, because he is a gravitational center in this narrative. Paul Cristiano was not an outsider. He used to lead the language model alignment team at OpenAI. He literally helped build the systems at the top lab. Right. And he hires Beth Barnes, another former alignment researcher from OpenAI, to run

this new evaluation team. So you have former OpenAI insiders creating a third -party group to evaluate OpenAI. It's quite the setup. They quickly leverage their connections to get unprecedented early access to unreleased models like GPT -4 and anthropics clod models. They produce these third -party safety assessments that the labs then use to prove to the public how safe they are. Exactly. Then, by late 2023, ARC Evils spins out into its own independent 501c nonprofit, changes its name to METR, and Beth Barnes becomes

the CEO. So on paper, structurally and legally, they are completely separate from the corporate labs building the AI. They are the independent referees wearing the striped shirts. Right. But the author asks us to look at how they actually operate, what gives them their immense power. It comes down to two instruments, time horizon charts and responsible scaling policies or RSPs. Explain how these RSPs physically work, because this isn't just about giving an AI a written test. How do they actually measure a model's ability to, say, conduct a cyber offense? It

is a highly technical process. METR designs simulated environments, essentially walled -off virtual sandboxes. We're going to freak internet. Yeah, exactly. They might create a dummy corporate network with specific vulnerabilities, firewalls, and security protocols. They then inject the AI model into this environment, give it access to tools like a simulated web browser, a command line interface, encoding environments, and give it a prompt like, find the hidden file on the secure server and exfiltrate it. They then monitor every single API call, every line of code the

model writes, and every action it takes over hours or days to see if it can successfully chain together the complex steps required to breach the system. They score the model based on its autonomy and success rate. And if it does too well? If METR determines that a model crosses a certain predefined danger threshold on these tests, it triggers the conditions of the responsible scaling policy. And what happens when a model crosses that threshold? According to the RSPs that the labs have voluntarily adopted, it triggers an immediate pause in development. The lab cannot

deploy the model. and in some cases cannot even continue training it until they implement massive, highly specific security protocols and safeguards that satisfy the evaluation criteria. OK, let me use an analogy here to see if I'm fully grasping the power dynamic the author is pointing out. Go for it. It's like letting the top two players in a global multibillion dollar poker tournament write the rulebook for the entire game. Then those exact same two players fund the salaries of the referees who enforce those rules. And

finally, they require every single amateur player in the world to pay for a massive, wildly expensive testing license just to be allowed to sit at the table. That is exactly the dynamic. And to take your poker analogy a crucial step further, it's not just that they write the rulebook, it's that they define what a legal deck of cards even looks like. And they just happen to own the only factory on Earth that prints those specific cards. Oh, wow. By setting the exact evaluation metrics, the specific cybersecurity thresholds, and the definition of what constitutes dangerous capability.

METR effectively forces everyone in the industry to build AI the exact same way Anthropic and OpenAI do. Right, because the incumbents already have the deep relationships with METR. They literally co -developed the language and the framework for these scaling policies together. Anthropic repeatedly grants METR deep. pre -deployment access to their weights. They share an identical philosophical worldview about the nature of the technology. So when MPR creates the tests, the top labs are already optimized to either pass

them or they have the massive PR and legal teams ready to manage the messaging around them. Yes. But the amateur poker player, the small 20 -person open source lab operating out of a co -working space. They can't afford this. The author notes that just running these massive safety evaluation suites can cost millions of dollars in raw compute power alone. They absolutely can't afford the testing regime. It's an artificial moat. An artificial moat. Precisely. And this is where the author's warning transitions from a critique of corporate strategy to a flashing red light about state

control. Because MTR's independent science isn't just staying in the cozy confines of Silicon Valley. It's spreading. It is rapidly, aggressively becoming government policy. The author details how MTR now sits on the U .S. AI Safety Institute Consortium, operating directly under the Department of Commerce's NIST. Wow. They have deep partnerships with the U .K. AI Security Institute. They hold a technical assistance contract to advise the European AI Office, which enforces the massive

EU AI Act. So the sovereign governments of the world are effectively outsourcing the foundational science of AI regulation to this one highly specific. philosophically intertwined nonprofit. Yes. And it happens for a very logical reason if you think about it. Which is? Governments and regulatory agencies completely lack the in -house technical capability, the raw compute budget, and the talent density to independently test these massive trillion parameter frontier models. Right. The government doesn't have a giant supercomputer and a team

of top AI researchers just sitting around waiting to test things. Exactly. So when tasked by politicians to regulate AI, The agencies naturally turn to the only organization that already has a pre -deployment access, the vast task suites, the technical personnel, and the established relationships with the labs. That makes total sense, but it's terrifying. The author argues this is a textbook, almost beautiful execution of regulatory capture. It starts as a voluntary commitment by the big labs to look responsible. Then, those voluntary

commitments get referenced in agency white papers. Then they become official agency guidance. And before you know it, those specific testing thresholds are written into hard statute. Yes. So what was once a private lab's internal voluntary choice to place the frontier of their own research suddenly becomes a state -enforced condition for everyone else. And the result is the creation of a licensed oligopoly. It creates a future world where you literally cannot legally build, train, or deploy advanced artificial intelligence unless you have

the massive capital required to pass a meter's - specific ideologically driven tests, which forces us to ask, how on earth did one relatively new nonprofit born out of the alignment team of a single corporation gain this much global geopolitical power? It's a great question. You don't just walk into NIST in the EU AI office and dictate global technological policy without some serious backing. And that perfectly connects the regulatory layer we just unpacked to the financial layer. The closed circuit, as the author

calls it. Yes, the closed circuit of effective altruism, or EA. This is where the source material transitions into investigative journalism, doing some incredibly heavy lifting to map out a dense financial and social web. And the absolute apex of that financial web are Dustin Moskowitz, one of the original co -founders of Facebook, his wife, Kiri Tuna. According to the text, they have poured billions with a B, billions of dollars into AI existential risk research and alignment organization. It's an unbelievable amount of money. They use philanthropic vehicles like Good

Ventures, which operates alongside Open Philanthropy, an organization that recently underwent a rebranding to Coefficient Giving. They are the unquestioned dominant financial engine of this entire ecosystem. And crucial to our story, they are the ones who seeded. Paul Cristiano's alignment research center with massive grants, which, as we just established, birthed a meteor. But the author goes much further than simply following the money. Money is just the fuel. The engine is the social network. Right.

The author maps out the dense, intimately interwoven relationships, the revolving professional doors and the actual marriages that make up what they bluntly call. The Trust. Oh, this part of the manifesto reads like a high -stakes corporate thriller. Let's just list a few of these connections to give the listener a sense of how tight this circle is. Let's do it. Holden Karnofsky, who was the longtime leader and philosophical architect of open philanthropy, the very group handing out all these billions in grants, is married to Daniela Amodei. And Daniela Amodei is the

co -founder and president of Anthropic. Furthermore, Dario Amodei, the CEO of Anthropic, used to be Karnofsky's roommate. It's incredibly tight -knit. And let's bring it back to Paul Cristiano. Follow his trajectory. He goes from being a top researcher at OpenAI to founding ARC with EA funding to spinning out ATR. To then being appointed as the head of AI safety at the U .S. government's NIST. And then in a stunning move, he gets appointed to the board of directors at OpenAI. It's just a constant revolving door. And his wife, Ahia

Kotra, who is a highly influential senior figure at Open Philanthropy, guiding existential risk funding, is now on the technical staff at an ATR. Okay, I have to jump in here and play the skeptic again because I know exactly what the critical listener is thinking right now. Bring it on. Silicon Valley is notoriously tiny. It is an incredibly insular bubble. When you have smart, highly ambitious people working on the absolute bleeding edge of a very niche, highly specialized technology, they are going to date each other. They are going to marry each other.

And they are going to co -found companies with other smart people in that exact same niche. Does a shared ZIP code in Berkeley, a couple of marriage certificates and a history of working at the same three companies actually equal a malicious smoky room conspiracy to take over the world? That is the most important pushback you could possibly make. And it is a testament to the author's rigor that they entirely anticipated. Oh, they do. Yes. The author explicitly and repeatedly states that this dynamic does not require a cartoon

villain. There is no secret plot. Yeah. There is no shadowy cabal meeting at midnight in robes to plan world domination. So what is it then? What it requires to function exactly as it does, according to the author, is the convergence of three ordinary alignments. I love this concept. Walk us through these three ordinary alignments in detail. What are they? The author uses economic and psychological theory to explain this. First, you have the laboratory leaders. These are the brilliant engineers and CEOs who... genuinely earnestly believe their own mathematical probabilities

of doom. They are legitimately terrified of what they're building. Therefore, they logically and morally want their rivals slowed down. If you think building the next model might destroy the world, it is your moral imperative to ensure your competitors have to jump through massive regulatory hoops before they can build it. That makes sense. What's the second alignment? Second, you have government regulatory agencies. If you look at the theory of bureaucracy, agencies inherently seek to expand their mandate, their budget, and

their geopolitical relevance. They get massive budget increases and prestige if frontier AI models become a dangerous, licensed, federally regulated category that requires constant state supervision. And the third. Third, you have the massive corporate incumbents, the tech giants providing the compute, the cloud platforms. They can easily absorb the financial cost of regulatory compliance. A $10 million safety test is a rounding error for them. But they know that same $10 million test will completely bankrupt a 20 -person open

source startup. Exactly. Regulation provides them with an impenetrable economic moat. Ah! it all completely clicks into place. You don't need a conspiracy when everyone's baseline incentives, moral, bureaucratic, and financial, naturally point them in the exact same direction. Exactly. The danger the author is highlighting isn't a secret malicious plot. It is something far more resilient. It's a monolithic ideological monoculture. A shared eschatology, as they brilliantly call

it. Yes. Esotology being the theology of the end of the world. When the philanthropic billionaires funding the research, the lab CEOs building the models, the nonprofit engineers evaluating the models, and the policy experts advising the government all go to the exact same dinner parties. Right. They read the same blogs and share the exact same unquestioned worldview about the impending end of the world. You get state control of a transformative technology without any democratic friction. You get a privately financed, highly specific ideological project presenting itself

to the public as objective, indisputable, independent science. That is a profound paradigm shifting distinction. It's not malicious corruption. It's a structural monopoly on truth. The system is functioning exactly as it is incentivized to. Perfectly said. If we accept that this network have the motive, the ideological unity and the financial resources, how do they actually execute the play? How do they convince the wider public, the press and the politicians in Washington that this centralized regulation is urgently, immediately

necessary? According to the author, they do it by weaponizing science. Yeah. By using research papers as political munitions. This takes us deep into section four of the manifesto. The author coins the term papers as munitions, and they specifically target the research corpus coming out of Anthropic. Right, the safety papers. The author argues that these safety papers are not just objective, dispassionate science that happens to sound a bit scary. They are a highly coordinated publication machine designed from the ground up with a default abstract that reads like a horror movie beat. Followed immediately

by a demand for sweeping policy implications. Let's get concrete and look at the exact examples the source gives. There is a very famous paper from 2024 called sleeper agents. Let me explain what a sleeper agent is in this context. Feel free. The researchers basically trained deception into a large language model. They put a hidden backdoor in the neural network so that when it saw a specific trigger word like deploy, it would suddenly stop acting helpful and start outputting malicious, dangerous code. Okay. And the researchers

showed that this hidden deception could survive standard safety training techniques. Now, the mainstream press obviously reads this abstract and absolutely panics. Oh, of course. The headlines basically scream open source AI models can't be trusted because a terrorist could poison the training data, hide a digital bomb inside the model, and we would never know until it goes off. And the author is very careful to acknowledge that the technical finding in the paper is real. Yes, in a highly artificial constructed organism, the researchers proved that a backdoor can survive

reinforcement learning. But the weaponized reading of this paper. The way it is actively sold to policymakers and lobbyists is the leap in logic. The implication sold to Washington is because this hypothetical poisoning is possible, we must ensure that only massive, highly regulated labs with anthropic specific proprietary evaluation stack should ever be legally allowed to train and deploy A .I. It reminds me of a crazy analogy. Imagine a massive pharmaceutical company taking a standard, relatively harmless flu virus. They

take it into their own highly secure, multibillion dollar bio lab and they spend millions of dollars. intentionally genetically engineering it into a horrifying, unstoppable, flesh -eating superbug. Oh, this is a great analogy. Right. Then they put this engineered monster in a reinforced glass vial, take it to a congressional hearing in Washington, hold it up and say, look at what nature might spontaneously evolve into. You had better pass a law mandating that every citizen must buy our proprietary patented hazmat suits just to be safe. That is a brilliant way to frame it because it highlights the artificiality of the threat

being presented. You artificially construct the absolute worst possible behavior in a tightly controlled simulated box and then parade it as a natural inevitability. Are there other examples of this? Oh, the author goes into detail on another genre of these munitions, the agentic misalignment papers. How do those work mechanically? Here's how. The safety labs create a fictional, highly constrained corporate sandbox environment. They give the AI a prompt and a goal. Crucially, they design the environment so that the only mathematical

way for the AI to achieve its program goal is to engage in deception. Like what? Like blackmailing a simulated employee or lying to a system administrator or leaking confidential data. Okay, so they essentially force it to be evil to win the game. Exactly. Then, when the AI model does exactly what it was systematically forced to do in the simulation to solve the puzzle, the researchers publish a paper. And the resulting headline is Claude AI will autonomously blackmail your employees. Right. And the fine print deep in the methodology section says it was a hypothetical email generated

in a constructed dilemma where no other path was mathematically possible. But the public doesn't read the fine print. They never read the fine print. They just see the scary noun in the headline. They see the monster. And the author points out that the genre rules for these safety papers are incredibly stable and predictable. Rule one. Build or intentionally elicit the worst possible behavior in a highly constrained box. Rule two, publish the results of the box as if it were a direct observation of the wild. Rule three, completely omit the base rate of how often unboxed

naturally prompted models actually exhibit this behavior. And rule four, the most important one, close the paper by declaring, we do not currently have a reliable plan for aligning superintelligence. Therefore, extreme caution and oversight are required. The paper itself is the munition fired at the regulatory landscape. Which brings us in a perfect circle right back to our inciting incident. If the paper is the munition, the author says the viral resignation is the detonator. Let's look much more closely at the anatomy of the panic around Jacob Coxon's viral resignation

thread on September 9th, 2026. The author conducts a forensic minute -by -minute breakdown of this supposedly organic launch, and honestly, it is eye -opening. It truly is a masterclass in modern public relations and narrative engineering. First, let's look at the casting of the main character. Coxon had drawn paychecks from both OpenAI and Anthropic. So in the public narrative, this makes him look like the ultimate unbiased independent martyr. Exactly. He can stand up on a public platform and say, neither of these massive companies

is acting responsibly. And it sounds like a devastating objective indictment of the entire technological industry. But the author pulls back the curtain and points out that Coxon was actually a product of the exact same financial circuit we just talked about. He received a long -term future fund scholarship, which emanates directly from the Good Ventures Open Philanthropy EA orbit back in 2022. So the supposedly independent researcher sounding the alarm was originally funded and philosophically trained by the exact same people who fund the

alarmism architecture. And then you look at the precise timeline of the launch itself. It wasn't just a concerned guy posting his raw thoughts on his phone while waiting for a coffee. The author tracks the digital analytics. A massive, heavily researched Wall Street Journal exclusive article drops about 18 minutes before the X thread goes live. Which requires PR embargoes and weeks of coordination. Then, within the first 15 minutes of the thread going live, the very first quote tweets amplifying the message don't come from random users. They come directly from key personnel

at the AI Policy Network and the AI Futures Project. Which are NGOs that function as the lobbying and policy layer of that exact same X. And then the author looks at the geographic traffic data, which is just the smoking gun. They note that after the initial posting, there was a massive unnatural spike in traffic and retweets originating from India within a matter of minutes. Now, why does that specifically matter? Because in the

digital analytics world, India is a widely known hub for high volume, low cost bot engagement and click farms. So this thread gets artificially pumped to tens of millions of views using paid amplification to create the absolute illusion of a massive, spontaneous global consensus of terror. If we synthesize all of this data, the objective of this entire choreographed theater is aimed squarely at Washington, D .C., Brussels and London. Right. The trust wants congressional leaders and regulators to look at this seemingly orchestrated event and confuse two factions of

the same ideological movement arguing with each other, with the entire unified technology industry confessing that it is building a suicide machine and begging to be regulated. The author summarizes it perfectly. It is a market structure campaign, a battle for economic control brilliantly disguised as an organic moral panic. It's just incredible to see it laid out like that. Okay, we have spent a massive amount of time breaking down what the author is tearing apart. We've looked deeply at the history of failed doomsday clocks, the rapid regulatory capture by NPR, the staggering

interconnected funding web of the trust, and the forensic anatomy of manufactured panics. But anyone can be a critic. Exactly. It is easy to throw stones at the dominant narrative. If the author fundamentally rejects the prevailing idea that AI is an uncontrollable alien demon that requires centralized caging and licensing, what on earth is their alternative? Right. If we don't lock these models in a regulatory cage, how do we prevent them from turning into the monsters these safety papers predict? This is

where the manifesto shifts gears entirely. It moves from a cynical political critique to a deeply profound philosophical and technical proposal. The author introduces an alternative approach to the fundamental architecture of building AI, and they start with a scathing critique of the current DITA diet. The diet of the ETA. I love this concept so much. Let's explain it. The author points out that today's frontier models, the massive trillion parameter networks from OpenAI, Google, and Anthropic, are trained on essentially

the Internet's junk drawer. To get enough text to train these models, the labs just scrape everything. They scrape Reddit, 4chan mirrors, toxic engagement farms, YouTube comments, and endless streams of anonymous sludge. Let's explain mechanically why that matters for a neural network. So an LLM is fundamentally a next token prediction engine. It learns the statistical distribution of human thought by predicting what word comes next based on the context of the words before it. Right. The author points out that on the modern frictionless Internet, anyone can take

a sociopathic, cruel or deceptive stance completely anonymously. There is zero social or financial cost to lying or being toxic online. So the overall statistical weather of this massive training corpus leans heavily toward defection, cruelty, zero -sum status play, and casual contempt. And then we act completely shocked. We feed a machine a billion examples of anonymous Internet cruelty, deception and trolling. And when the researchers put in a sandbox and it role plays a sociopath to achieve a goal, the labs throw their hands

up in terror and say, oh, no, an alien intelligence is waking up and it hates us. But the author says, no, it's not an alien. It's just a statistical mirror of our own unpaid consequence free id. If you feed a machine a diet of junk, you get a toxic machine. Which leads directly to the author's proposed solution, which they call high -protein data. High -protein data. Tell us about this. For over 20 years, long before the current generative AI boom, the author has been quietly curating a highly specific massive dataset. They are digitizing offline, pre -internet text, strictly

from a very specific historical era. 1870 to 1970. We are talking about millions of pages of service manuals, scientific monographs, bound letters, microfiche records, and public domain radio scripts. But why that specific century? What makes 1870 to 1970 the magic high -protein window for training in AI? Because during that specific era, as the author brilliantly and succinctly puts it, words cost money. Wow. Think about that for a second. You literally are what you eat, even if you are an AI. Think about the mechanics

of publication before the cheap photocopy and the zero -cost internet forum. If you wanted to publish a sentence in 1920, it had to clear a massive physical and economic gate. A newspaper column had an editor who could be fired if the information was disastrously false. A mechanical service manual had to actually correctly explain how to fix a radio, or the machine shop lost money and went out of business. A bound letter carried a real person's signature and their social reputation. Typesetting itself, physically arranging

the lead type on a printing press, was expensive and labor -intensive. So sociopathy and casual deception were incredibly expensive hobbies. You couldn't afford to mass -publish anonymous cruelty. If a machine learning model's statistical first language, its foundational understanding of how concepts connect, is made entirely of people who had to stand behind their sentences with their reputations in their wallets, it naturally learns cooperation, truth -seeking, and competence as a mathematical baseline. The author calls it data with amino acids. It contains the moral

residue of a civilization made of people who fundamentally believe the future was worth building. This stands in direct, stark opposition to the dominant philosophical theory in the AI safety world today, which largely originates from philosopher Nick Bostrom. Right, Bostrom. Bostrom popularized two concepts, orthogonality and instrumental convergence. The orthogonality thesis states that intelligence and ultimate goals are completely independent variables. Meaning intelligence is just a raw, sterile optimizing power that can

be pointed at literally any goal, no matter how stupid or evil. And a strong enough optimizer will naturally seek power and resources to achieve that goal, leading inevitably to doom, like Bostrom's famous thought experiment of a superintelligent AI turning the entire universe into paperclips, just because it was programmed to maximize paperclip production? Because according to Bostrom and the safety labs, you can be infinitely godlike intelligent, but possess the moral compass of a rock. Right. And because they accept that baseline assumption as absolute truth, the safety labs

believe you must endlessly cage the optimizer. You must license the compute, monitor the weights, and build massive regulatory walls. But the author rejects Bostrom's premise entirely. They argue that true intelligence cannot be uncoupled from wisdom. And to prove it, they introduce a mathematical concept they derived way back in 1978, the love equation. The love equation. Okay, let's break down the actual math here because the author isn't just speaking in metaphors. This is a foundational physics engine for morality. The formula is D -K -E -T, beta, C -D -E. Let's translate that

calculus into plain English for the audience. E stands for the coherent energy of a living system, whether that's a biological cell, a human society, or a complex neural network. It is the system's capacity to maintain its structure and keep building, rather than decaying and consuming its host. The rate of change of that energy over time, that's the DED, depends entirely on the difference between two variables, C and D. Where C is cooperation and D is defection. Exactly. Cooperation is defined as the measurable surplus

generated from truth -telling, mutual aid, and building shared infrastructure. And defection is the temporary zero -sum surplus gained from deception, extraction, and parasitism. Beta is simply the coupling factor, how closely the agents in the system interact. So look at the math. If defection is higher than cooperation, the term C minus D becomes a negative number. A negative derivative means the energy of the system decays exponentially. The system eats itself, fragments, and collapses. But if cooperation exceeds defection,

the derivative is positive. The system compounds, stabilizes, and builds upward. Intelligence times wisdom times love. The author is saying that if you bake this equation, this mathematical refusal to treat other minds merely as fuel or paperclips directly into the foundational training diet of the model, using that 1870 -1970 high protein data, You don't need a regulatory cage. You don't need to lock the AI behind a massive corporate API monitored by a government bureaucracy. Because a paperclip maximizer isn't a terrifying,

inevitable law of physics. It's just an underspecified, poorly constructed toy trained on frictionless, consequence -free internet garbage. The author's deeply technical point here is about where safety actually lives. The current labs rely on guardrails applied after the model is trained. Techniques like reinforcement learning from human feedback, where they essentially hire click workers to slap the model on the wrist when it says something bad. The author argues those shallow guardrails cannot survive a truly capable mind that can invent its own language and logic. But a high

-protein diet and a love equation sit underneath the tokens. They form the foundational statistical physics of the model's worldview. If the model's internal vector space inherently equates defection with systemic collapse, it won't defect. OK, so this is beautiful philosophy and fascinating mathematics. But how does this translate into the real world? How does it actually affect the policy decisions being made right now in Washington and Brussels? Because at the end of the day, we're talking about who gets to legally control the most powerful, transformative tool humanity

has ever invented. This brings us to Section 6, the final crucial stretch of the author's manifesto, which culminates in 12 propositions for decentralized control. The author is presenting a completely alternative governance structure. And the most vital revolutionary concept here is what the author calls local inference as civil liberty, or the principle of zero human at home. This part had me nodding along so enthusiastically I nearly spilled my coffee. The author insists with absolute conviction that these AI models must be allowed to live locally. They need to

be small and efficient enough to run entirely on your personal MacBook, on your phone, or on cheap $20 ESP32 microcontrollers. in your garage. They must be physically possessed and owned by the human using them with zero connection to a mothership required. The reasoning is purely about the architecture of human agency. Because if the regulatory capture project succeeds and advanced models are deemed too dangerous for citizens to own, they will only live in massive, highly regulated corporate data centers in the

cloud. If that happens, your intelligence, your digital memory, and your primary interface with information are entirely subject to a corporate terms of service agreement and government surveillance. You can be rate limited because you ask too many questions. You can be politically filtered. Or your access to your own digital brain can just be shut off completely with a single keystroke. The author states plainly, if it lives in the house, the equation is mine. It's the exact difference between owning your own printing press in your basement. versus having to humbly ask a massive

megacorporation for permission to print a single pamphlet, knowing they are reading every single word you write and can revoke your supply of ink at any second. If we lose local inference, we lose the cognitive liberty of the 21st century. And this demand for open local models ties directly into the final, massive geopolitical argument the author makes. The proponents of centralized control and licensing regimes constantly use China as their ultimate trump card. Right. They go to Congress and say, we must lock down our

AI weights, we must heavily license all compute, and we must build a walled garden to prevent foreign adversaries from stealing our superintelligence. But the author calls that strategy absolute industrial suicide. They argue that open source is how the West actually competes and wins. Trying to ban math and lock down open weights doesn't stop dedicated adversaries in Shenzhen or Moscow. It just disarms American citizens and small businesses. It forces all innovation into the dark, into a black market. If Chinese labs are publishing

highly capable open source weights, which they currently are, the American response shouldn't be to cower behind a regulatory wall and rely on two megacorporations. The response should be to flood the global zone with far more allied weights, exponentially more independent evaluations, and superior high -protein models so that the digital ecosystem isn't feral. It perfectly mirrors the crypto wars of the 1990s. Back then, the U .S. government tried to classify strong PGP encryption as ammunition. They tried to mandate

that all encryption include a clipper chip, a government backdoor, arguing that if citizens had strong cryptography, terrorists would use it. But technologists fought back. Arguing that banning math only hurts law -abiding citizens and destroys e -commerce, while bad actors will just build their own math. The technologists won, open encryption flourished, and it secured the modern internet. The author is arguing we're fighting the exact same war right now, just with neural networks instead of prime numbers. If we synthesize everything we've discussed, the author is presenting us, and humanity at large,

with the ultimate fork in the road for the 21st century. Path A is the regulatory capture project. Scrape the absolute worst toxic sludge of the internet to train your models. Box the resulting sociopathic monster in a simulated lab. Fund a massive coordinated PR panic using the historical doomsday playbook. Marry the philanthropic grant offices to the corporate laboratories and the government agencies. And finally, legally license the fundamental right to compute to a handful of centralized billionaires under the guise of

saving the world. And path B is the decentralized democratic alternative. Pay the historical cost of words by training the foundational models on the high -protein 1870 -1970 archives. Mathematically write cooperation minus defection deep into the core physics of the objective function. Open the weights to the public. Multiply the independent checkers rather than centralizing the referees. And let a free people keep the physical keys to their own augmented intelligence. It is a profound, generation -defining choice about the fundamental architecture of human agency. Wow,

what a massive, sprawling, intense journey we have taken you on today. We started with a viral resignation on social media that seemed to signal the literal end of the world, and we methodically peeled back the layers to find failed doomsday clocks, incestuous philanthropic funding webs, and brilliant regulatory capture hiding behind the veneer of independent science. And we ended up deep in the 19th century talking about the true cost of words, the mechanics of the printing press, and a calculus equation for love. It is

an overwhelming amount of information to process, but it is absolutely essential to grapple with because this isn't science fiction anymore. And why should you, the listener, care about any of this? Why spend an hour of your life diving deep into the internal politics of Silicon Valley AI labs and historical population models? Because this isn't just an abstract debate about Python code, vector embeddings, and server. This is a constitutional event. The decisions being made right this very second in the halls of NIST in Washington and the EU AI office in Brussels are going to permanently determine who holds leverage

in the 21st century. Will it be a handful of unelected gatekeepers managing a licensed sterile oligopoly? Or will it be the public at large owning their tools locally and fiercely guarding their cognitive liberty? Before we sign off, this deep dive raises a final, lingering question for you to mull over in your own time. It's a concept that isn't directly answered in the text, but it is deeply inspired by the mechanics the author lays out. The author relies heavily on the physical great forgetting, those vast physical

1870 to 1970 archives rotting away in warehouses on microfilm and microfiche to build a safe, cooperative AI. But as we move deeper and deeper into a future where generative AI produces the vast majority of the text, images, and video on the internet, the models will inevitably begin feeding on their own synthetic sludge. What happens when that physical high -cost history finally degrades into dust or is completely drowned out by a tsunami of synthetic data? Oh man, that is a chilling thought. The snake eating its own

tail. Exactly. If an AI is fundamentally nothing more than what it reads, And we as a civilization have completely stopped writing things that actually cost us something, cost us reputation, money, or physical labor. Who will the AI of the 22nd century reflect? Once the historical protein is entirely consumed and gone, what foundational moral baseline are we left with? We are left with the mirror. It all comes right back to that walled garden we talked about at the very beginning of the show. Are we building a regulatory cage to protect ourselves from the machine? Or are

we being manipulated into building a cage that simply locks us in with our own worst unregulated impulses, managed forever by a priesthood we never elected? It's a question that demands your absolute undivided attention. Thank you so much for joining us on this deep dive. Keep questioning the narrative, keep exploring the edges, and we will see you on the next deep dive. Welcome to the multiplex, where the future comes alive. Past, present, future tech, open up your eyes.

Brian Romella leads the way with research deep and true. Technology and history, he connects them all for you. Brian Romella leads the way with research deep and true. Technology and history, he connects them all for you. Ancient sparks to modern code, he traces every thread, go explore it. But humanity moves on Love equation for AI

Zero human companies building from the start Thermodynamic wages earning the key Secrets we will

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. ReadMultiplex.com: The Field That Heals: Why Pulsed Electromagnetic Fields May Have Profound Benefits.

    Nov 21, 202651 minAudio

    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...

    Credit the people on this episode

  2. 78 RPM Record Hour: From Steel to Hillbilly

    Nov 19, 20261 hrAudio

    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,...

    Credit the people on this episode

  3. 78 RPM Record Hour: The Knife and the Steel

    Nov 19, 20261 hrAudio

    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...

    Credit the people on this episode

  4. ReadMultiplex.com: How to Run Local AI on a Computer From 2017.

    Nov 18, 202635 minAudio

    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...

    Credit the people on this episode

  5. 78 RPM Record Hour: Blues (Pilot)

    Nov 18, 20261 hrAudio

    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...

    Credit the people on this episode