
Leopold Aschenbrenner — 2027 AGI, China/US super-intelligence race, & the return of history
What this covers
This is a four-and-a-half-hour conversation in which Dwarkesh Patel interviews Leopold Aschenbrenner, whose essay series on the same themes the episode is built around. It begins with the industrial side of AI: a long-run trend of roughly half an OOM of training compute per year, extrapolated into clusters that grow from tens of thousands of GPUs to 10 GW by 2028 and a trillion dollars by 2030. Leopold treats this as heavy industry (power plants, fabs, natural gas or deregulated green megaprojects) and not as a typical software business. He pairs the scaling argument with unhobbling, the idea that models hold raw capability which techniques like RLHF and longer reasoning unlock, with agentic "drop-in" workers as the next step.
From there the talk turns to geopolitics and history. Leopold argues that the postwar era of uncontested US hegemony was the exception, and he uses WWII mobilization figures, the pre-1939 state of the US military, and Paul Samuelson's 1961 prediction that the Soviets would outgrow America to question pessimism about democratic state capacity. He expects the CCP) to try to infiltrate American labs and to outbuild the US on power, and he argues labs need nuclear-program-level security. The state-led versus private-led question, the risks of putting clusters in the Middle East, and a six-month lead as a margin for alignment work are among the more contested threads.
The later sections are more personal and varied. Leopold covers his path to graduating from Columbia at 19, what happened at OpenAI before he left, his views on Germany and on understanding foreign perspectives, and his plan for an AGI-focused investment fund. Dwarkesh adds his own immigration story. The conversation ends with lessons from WWII and a coda on Frederick the Great. The intelligence explosion, where automated AI researchers compress a decade of progress into a year, and the alignment discussion sit in the middle of the episode, around the 2:46 and 3:26 marks.
Leopold Aschenbrenner argues that superintelligent AI is arriving by 2027-2028 through scaling of compute and algorithmic breakthroughs, creating an intelligence explosion that will be historically decisive for geopolitical competition, national security, and potentially human civilization itself.
- Compute scaling follows a predictable exponential trajectory (0.5 OOMs/year) leading to trillion-dollar clusters by 2030, with AGI emerging from 10 GW clusters by 2028
- The intelligence explosion occurs when AI systems automate AI research itself, potentially compressing decades of technological progress into years and creating decisive military/economic advantages
- This geopolitical race will determine whether liberal democracy or authoritarian systems prevail, requiring the US to keep AGI training clusters domestic and maintain algorithmic secrecy advantages
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
WWII industrial mobilization in the US was more feasible than commonly believed despite perceived constraints. The narrative that 'we had state capacity then and it's a clusterfuck now' is incorrect; industrial mobilization in WWII was extremely competent but faced significant labor agitation and strikes even in 1941 that cost millions of man-hours during peak wartime production, demonstrating that even maximum effort mobilization faces obstruction.
“They would literally have these strikes, into 1941, costing millions of man-hours worth of time when we're trying to make tens of thousands of planes a month. They would just debilitate factories for trivial concessions from capital that were pennies on the dollar.”
People tend to underestimate the rate of change and overestimate the difficulty of being in the trenches of a field. Being in close focus on immediate technical challenges (debugging current models, fixing data issues) reduces visibility into macro trends (exponential capability scaling, approaching AGI) and makes algorithmic progress seem less impressive than it actually is.
“Why do people underrate it? Being in the trenches actually gives you a less clear picture of the trend lines. You don't have to zoom out that much, only a few years. When you're in the trenches, you're trying to get the next model to work. There's always something that's hard. You might underrate algorithmic progress because you're like, 'ah, things are hard right now,' or 'data wall'”
The CCP views itself as engaged in a historical mission of Chinese national rejuvenation and thinks deeply about world order and national power, unlike the US which assumes hegemony.
“China very much thinks of itself on this historical mission of the rejuvenation of the Chinese nation. They think a lot about national power. They think a lot about the world order.”
AI is fundamentally an industrial process unlike most recent Silicon Valley innovations, requiring massive new computing clusters, power plants, and semiconductor fabrication facilities to advance the next generation of models.
“Unlike most things that have recently come out of Silicon Valley, AI is an industrial process. The next model doesn't just require some code. It's building a giant new cluster. It's building giant new power plants. Pretty soon, it's going to involve building giant new fabs.”
The post-WWII period was historically exceptional in that the US experienced 80 years of uncontested hegemony and peace. Historically, the norm is 'extremely intense and extraordinary things happening in the world with intense international competition.' WWII saw 50% of US GDP mobilized to war production, and the UK/France/Germany borrowed over 100% of GDP. The stakes and mobilization intensity are returning to historical norms, not new.
“Our generation is so used to peace, American hegemony and nothing matters. The historical norm is very much one of extremely intense and extraordinary things happening in the world with intense international competition. There's a 20-year very unique period. In World War II, something like 50% of GDP went to war production. The US borrowed over 60% of GDP. With Germany and Japan I think it was over 100%.”
Before 1939, the American military was in 'total shambles,' with military expenditures below 2% of GDP while European powers spent above 10%, yet the United States mobilized rapidly from this baseline with latent industrial capacity, suggesting the US has reserves of mobilization capacity that could apply to AGI development if political will exists.
“Before 1939, the American military was in total shambles... Military expenditures were I think less than 2% of GDP. All the European countries had gone, even in peacetime, above 10% of GDP... But there was this latent capacity. At some point the United States got its act together.”
Even as late as 1961, economist Paul Samuelson predicted that the Soviet Union would outgrow the United States due to superior industrial mobilization capacity, demonstrating that the superior outcome of capitalism over communism was not inevitable or obvious at the time, and suggesting current pessimism about whether liberal democracies can outcompete authoritarian systems may be similarly misplaced.
“Even as late as 1961, Paul Samuelson was predicting that the Soviet Union would outgrow the United States because they were able to mobilize industry better.”
Historical warfare involving 20-30% of population deaths (Thirty Years War) and loss of large percentages of national land (20% of Poland in WWII) were commonplace, showing modern peace is unprecedented.
“Much more was on the line. People talk about World War I being so destructive with 20 million Soviet soldiers dying and 20% of Poland. That happened all the time. During the Seven Years' War something like 20-30% of Prussia died. In the Thirty Years' War, up to 50% of a large swath of Germany died.”
Pre-training via next-token prediction is 'magical' because it enables representation learning—models learn models of the world rather than mere statistical artifacts—which allows them to generalize and apply knowledge to novel problems, rather than just learning surface-level patterns.
“First of all, pre-training is magical. It gave us a huge advantage for models of general intelligence because you can predict the next token. But there's a common misconception. Predicting the next token lets the model learn incredibly rich representations. Representation learning properties are the magic of deep learning. Rather than just learning statistical artifacts, the models learn models of the world.”
Once AGI/drop-in remote workers arrive and can be scaled to productivity levels, the next logical automation target is AI research and engineering itself. Once AI researchers are automated, you can run millions or billions of human-equivalent automated AI researchers on GPU fleets, enabling a decade's worth of ML research progress to be compressed into a single year, triggering a 10x speedup in research velocity and potentially a jump to vastly superhuman AI within 1-2 years.
“One of the first jobs to be automated is going to be that of an AI researcher or engineer. If you can automate AI research, things can start going very fast. Right now, there's already at this trend of 0.5 OOMs a year of algorithmic progress. At some point, you're going to have GPU fleets in the tens of millions for inference or more. You're going to be able to run 100 million human equivalents of these automated AI researchers. If you can do that, you can maybe do a decade's worth of ML research progress in a year. You get some sort of 10x speed up. You can make the jump to AI that is vastly smarter than humans within a year, a couple of years.”
Human learning relies on iterative dialogue with study materials and peers: reading a dense textbook once is insufficient; instead, humans read, think internally, discuss with study buddies, attempt practice problems, fail repeatedly, and eventually achieve understanding—a process fundamentally different from simply consuming pre-training data.
“What you do is you read a page, think about it, have some internal monologue going on, and have a conversation with a study buddy. You try a practice problem and fail a bunch of times. At some point it clicks, and you're like, 'this made sense.' Then you read a few more pages.”
Superintelligence under authoritarian control would enable perfect surveillance, perfect lie detection, identification and elimination of all dissenters, and prevention of any institutional evolution or ideological challenge, locking in the party's ideology forever and preventing the gradual dissipation of authoritarian systems that occurred historically through human doubt and institutional change.
“Imagine you have a perfectly loyal military and security force. No more rebellions. No more popular uprisings. You have perfect lie detection. You have surveillance of everybody. You can perfectly figure out who's the dissenter and weed them out. No Gorbachev who had some doubts about the system would have ever risen to power.”
If superintelligence development creates a tightly contested race where multiple parties are only months apart, the situation becomes extremely dangerous because there is no time for safety considerations, everyone operates at maximum speed, constant fear of being overtaken, and new WMDs appear continuously, creating unstable deterrence that could collapse at any moment.
“The world in which we're really neck and neck, we only have a three-month lead, is incredibly dangerous. We're in this feverish struggle where if they get ahead, they get to dominate, maybe they get a decisive advantage... There are crazy new WMDs popping up. Then we're going to be in the situation where it's crazy new military technology, crazy new WMDs, deterrence, mutually assured destruction keeps changing every few weeks. It's a completely unstable, volatile situation that is incredibly dangerous.”
The CCP will launch an all-out effort to infiltrate American AI labs, allocate billions of dollars, deploy thousands of personnel, and fully mobilize the Ministry of State Security, while simultaneously attempting to outbuild the US, given that China has added as much power capacity in the last decade as an entire US electric grid and can therefore more easily construct the 100 GW clusters.
“The CCP is going to have an all out effort to infiltrate American AI labs. Billions of dollars, thousands of people, CCP is going to try to out-build us... They added as much power in the last decade as an entire US electric grid. So the 100 GW cluster, at least the 100 GW part of it, is going to be a lot easier for them to get.”
Superintelligence will be absolutely decisive for national power in the same way that the atomic bomb was, because a nation that achieves superintelligence first can use it to compress centuries of technological progress into years, creating military advantages analogous to or exceeding the 100:1 kill ratio the US achieved in the 1991 Gulf War through technological superiority.
“The question is, when does the CCP and when does the American national security establishment realize that superintelligence is going to be absolutely decisive for national power?... If you look at the first Gulf War, Western coalition forces had a 100:1 kill ratio. They had better sensors on their tanks... They had maybe 20-30 years of technological lead. They just completely crushed them.”
Big Tech companies understand that scaling is critical and plan for 10 GW data centers by 2028. Only Big Tech, sovereign wealth funds, and major countries can afford and build these scales.
“Companies realize that scaling is a thing. Obviously their whole plans are contingent on scaling. At that point, the people who can keep up are Big Tech, potentially at the edge of their capabilities, sovereign wealth fund-funded things, and also major countries like America and China.”
To build AGI infrastructure at scale in the US, the nation has two main paths: (1) increase natural gas production (the US can double production again in 7 years, providing enough power for multiple trillion-dollar data centers), or (2) pursue green energy megaprojects (solar, batteries, SMRs, geothermal) coupled with massive deregulation (FERC reform, NEPA exemptions, streamlined state permitting).
“There are basically two paths to doing it in the US. One is you just have to be willing to do natural gas. There's ample natural gas... The 10 GW cluster is super easy. The 100 GW cluster is also pretty doable. I think natural gas production in the United States has almost doubled in a decade. You do that one more time over the next seven years, you could power multiple trillion-dollar data centers.”
Training compute for the largest AI systems has grown by approximately 0.5 orders of magnitude per year for nearly a decade, a long-run trend that can be extrapolated forward to predict cluster sizes: GPT-4 at ~25,000 A100s ($500M) in 2022, scaling to 100,000 H100 equivalents ($billions) by 2024, 1 million H100s (10 GW, tens of billions) by 2026, 10 million H100s (100 GW, hundreds of billions) by 2028, and 100 million H100s (1 trillion dollars, 100 GW) by 2030 consuming over 20% of US electricity production.
“There's this long-run trend of almost a decade of training compute for the largest AI systems growing by about half an order of magnitude, 0.5 OOMs a year. Just play that forward. GPT-4 was reported to have finished pre-training in 2022. On SemiAnalysis, it was rumored to have a cluster size of about 25,000 A100s. That's roughly a $500 million cluster... By 2024, that's a cluster that's 100 MW and 100,000 H100 equivalents with costs in the billions. Play it forward two more years. By 2026, that's a gigawatt, the size of a large nuclear reactor... By 2028, that's a cluster that's ten GW... By 2030, you get the trillion-dollar cluster using 100 gigawatts, over 20% of US electricity production.”
AI labs should maintain state-level security comparable to nuclear weapons programs because algorithmic breakthroughs, training techniques (like RL-based unhobbling), and model weights can all be stolen relatively easily—current security is at Level 0 (vulnerable to state activity) when it should be at Level 4, and a single person recently stole valuable AI code by copying it to Apple Notes and exporting as PDF, bypassing monitoring.
“It's all extremely easy. They don't make the claim that it's hard. DeepMind put out their Frontier Safety Framework and they lay out security levels, zero to four. Four is resistant to state activity. They say, we're at level zero.”
Companies claim they cannot build AGI clusters in the US due to climate commitments they've made (Microsoft, Amazon, etc.), but at some point national interest and national security must take precedence over climate commitments, similar to how labor strikes debilitated US WWII production but society accepted these tradeoffs because war was existential.
“The issue there is that a lot of people made these climate commitments, not just the government. It's actually the private companies themselves, Microsoft, Amazon, etc., that have made these climate commitments. So they won't do natural gas. I admire the climate commitments, but at some point the national interest and national security is more important.”
RLHF (reinforcement learning from human feedback) unhobbled GPT-2's raw capabilities into a functional chatbot, achieving roughly a 100x improvement in human preference ratings compared to non-RLHF models, and enabled basic chain-of-thought reasoning, but the underlying model still possessed many unutilized raw capabilities that could be further unlocked.
“When you train a model, you have this raw bundle of capabilities that's useful. The unhobbling from GPT-2 to GPT-4 took this raw mass and RLHF'd it into a good chatbot. That was a huge win. In the original InstructGPT paper, comparing RLHF vs. non-RLHF models it's like a 100x model size win on human preference rating.”
During the superintelligence intelligence explosion period, having a six-month buffer (even if China catches up within a year) allows the leading side to dedicate additional compute to alignment research and to pause or reconsider if something feels uneasy. This buffer is critical to whether humanity survives the transition successfully or destroys itself through misaligned superintelligence.
“It might be really important during the intelligence explosion to have a six-month wiggle room to be like, 'look, we're going to dedicate more compute to alignment during this period because we have to get it right. We're feeling uneasy about how it's going.' One of the most important inputs to whether we will destroy ourselves or whether we will get through this incredibly crazy period is whether we have that buffer.”
By 2025-2026, AI models will be basically smarter than most college graduates, but their economic usefulness will depend heavily on 'unhobbling' — transitioning them from limited chatbots to agentic systems capable of long-horizon tasks.
“By 2025-2026, we're going to get models that are basically smarter than most college graduates. A lot of the economic usefulness depends on unhobbling. The models are smart but limited. There are chatbots and then there are things like being able to use a computer and doing agentic long-horizon tasks.”
Reports indicate that Sam Altman is attempting to raise $7 trillion for a chip/infrastructure project, suggesting some clusters will be partially located or developed outside the US, though the exact geographic distribution is unclear.
“It's been reported that Sam Altman is trying to raise $7 trillion or whatever for a chip project. It's unclear how many of the clusters will be there, but definitely stuff is happening.”
China has already demonstrated ability to make 7-nanometer chips and has significant latent industrial capacity to scale chip production and power infrastructure rapidly.
“They're able to make 7-nanometer chips now. There's a question of how many they could make. There's at least a possibility that they're going to mature that ability and make a lot of 7-nanometer chips. There's a lot of latent industrial capacity in China. They are able to build a lot of power fast.”
Nvidia's data center revenue has grown from a few billion dollars per quarter exactly one year after their blockbuster earnings call (which occurred on an anniversary of 2023) to $25 billion per quarter, demonstrating the capital acceleration unleashed by ChatGPT.
“Exactly a year ago today, Nvidia had their first blockbuster earnings call. It went up 25% after hours and everyone was like, 'oh my God, AI is a thing.' Within a year, Nvidia data center revenue has gone from a few billion a quarter to $25 billion a quarter and continues to go up.”
A common mistake people make is valuing the GPT-4 cluster at $100 million based on rental price, when actually building and owning the entire training cluster costs approximately $500 million, because companies must pay for the full cluster capacity, not just the three-month rental period.
“There's a common mistake people make, saying it was $100 million for GPT-4. That's just the rental price. If you're building the biggest cluster, you have to build and pay for the whole cluster. You can't just rent it for three months.”
Despite the Soviet ability to copy the American atomic bomb design, the initial copied design was suboptimal compared to designs the Soviets could have independently developed, suggesting that parallel invention (where multiple parties independently discover similar solutions) is common in technology, and that even stolen algorithmic knowledge may not provide overwhelming advantage if the stolen design has drawbacks.
“The Soviets would have had a better bomb if they hadn't copied the American design, at least initially. That suggests something about history, not just for the Manhattan Project. There's often this pattern of parallel invention because the tech tree implies that a certain thing is next...”
Secreting algorithmic breakthroughs (preventing publication) can provide significant lead time over competitors: Germany's failed atomic bomb project due to not realizing graphite worked as a moderator (because Fermi's measurements were suppressed) suggests that 1-2 years of secrecy on fundamental insights could provide equivalent to multiple OOMs of computational advantage.
“In the early days, there had been some incorrect measurements made on graphite as a moderator. Germany thought graphite wasn't going to work, so they had to do heavy water. But then Enrico Fermi made new measurements indicating that graphite would work... Szilard assaulted Fermi with another secrecy appeal... In the end, Fermi didn't publish it. That was just in time. Fermi not publishing meant that the Nazis didn't figure out graphite would work.”
The analogy to human System 1 vs System 2 thinking is illustrative: scaling alone improves System 1 (autopilot/immediate response), but unlocking true agentic capability requires System 2 thinking (deliberate, reflective reasoning). The brute force path to agents is improving System 1 indefinitely, but there is a faster path: get System 2 working, which would quickly lead to agentic behavior and unlock the test time compute overhang.
“Here's an analogy. When you drive, you're on autopilot most of the time. Sometimes you hit a weird construction zone or intersection. Sometimes my girlfriend is in the passenger seat and I'm like 'ah, be quiet for a moment, I need to figure out what's going on.' You go from autopilot to System 2 and you're thinking about how to do it. Scaling improves that System 1 autopilot. The brute force way to get to agents is improving that system. If you can get System 2 working, you can quickly jump to something more agentified and test time compute overhang is unlocked.”
The question of whether Middle Eastern dictatorships (like the UAE) should have seats at the 'AGI table' is distinct from whether they can have access to AI products. The US could offer them access to last-generation models or commercial AI services (benefit-sharing) without giving them leverage over superintelligence development. Offering them a seat at the AGI table is not necessary for alliance-building and creates unjustified security risk.
“There should be a narrow coalition of democracies that's developing AGI. Then there should be a broader coalition of other countries, including dictatorships, and we should offer them some of the benefits of AI. If the UAE wants to use AI products, run Meta recommendation engines, or run the last-generation models, that's fine. By default, they just wouldn't have had this seat at the AGI table.”
Superintelligence could enable drone swarms to locate and neutralize ballistic missile submarines and mobile launchers, potentially undermining nuclear deterrence and making the technology 'enormously destabilizing'.
“How do you find nuclear stealth submarines? Right now, you have sensors and software to detect where they are. You can do that. You can find them. You have millions or billions of mosquito-sized drones, and they take out the nuclear submarines. They take out the mobile launchers. They take out the other nukes. It's potentially enormously destabilizing and enormously important for national power.”
GPT-4 currently performs chain-of-thought reasoning with a few hundred tokens of test-time compute (equivalent to about 3 minutes of human thinking at 100 tokens per minute), but there exists a massive 'test time compute overhang' where models could theoretically use millions of tokens for iterative debugging and planning, representing a +4 orders of magnitude improvement if unlocked.
“Right now, GPT-4 can do a few hundred tokens with chain-of-thought. That's already a huge improvement... GPT-4 thinks for a few hundred tokens. If I think at 100 tokens a minute, that's like what GPT-4 does. It's equivalent to me thinking for three minutes. Suppose GPT-4 could think for millions of tokens. That's +4 OOMs on test time compute on one problem.”
Robotics has historically been limited by hardware, but that constraint is being solved; however, robotics lacks pre-training's bootstrapping advantage of unsupervised learning, forcing it to begin immediately with reinforcement learning self-play without the foundation of world models that language models gained from internet text.
“This pre-training advantage is also the difference to robotics. People used to say it was a hardware problem. The hardware is getting solved, but you don't have this huge advantage of bootstrapping with pre-training. You don't have all this unsupervised learning you can do. You have to start right away with RL self-play.”
The March 2020 COVID moment (when the scale of the pandemic became undeniable and governments mobilized within weeks) provides a model for how AGI realization might occur: a sudden shift from abstract concern to concrete action by politicians, CEOs, and the general public, creating a window where extraordinary measures become politically feasible.
“COVID in February of 2020 honestly feels a lot like today. It feels like this utterly crazy thing is coming... At some point, people saw it and then crazy, radical reactions came.”
AI lab employees should advocate for representative democracy institutions (rather than direct democracy) in AI organizations, learning from board governance experiences which highlighted benefits of deliberative institutions in managing complex decisions.
“I do think they would benefit from some more organs of representative democracy... In the OpenAI board events, employee power is exercised in a very direct democracy way. How some of that went about really highlighted the benefits of representative democracy and having some deliberative organs.”
Even without physical robots, superintelligent AI can make factory workers into skilled technicians by instructing them via augmented reality glasses and voice, dramatically increasing manufacturing capacity.
“Initially, you start with the factory workers. They're wearing the glasses and AirPods, and the AI is instructing them because you can make any worker into a skilled technician. Then you have the robots come in.”
Leopold Aschenbrenner's family history—his mother from former East Germany, father from former West Germany, great-grandmother born 1934 experiencing Nazi era and East German communism, immigrating to freedom only at age 60—shapes his visceral understanding of totalitarianism's danger and why AI competition matters existentially.
“To your point about history and living in America for the past eight years, this is one of the things I took away from growing up in Germany. A lot of this stuff feels more visceral. My mother grew up in the former East, my father in the former West. They met shortly after the Wall fell... I grew up in Berlin with the former Wall. My great-grandmother... was born in 1934 and grew up during the Nazi era.”
Placing AGI compute clusters in authoritarian dictatorships creates irreversible security risks: they can steal weights, seize compute, or exfiltrate them to China.
“Once you put them in authoritarian dictatorships, you create this irreversible security risk. Once the cluster is there, it's much easier for them to exfiltrate the weights. They can literally steal the AGI, the superintelligence. It's like they got a direct copy of the atomic bomb.”
The fundamental approach to pre-training (unsupervised learning, next token prediction) is quick to communicate, but crucial implementation details matter enormously and will likely be somewhat obvious in retrospect.
“The fundamental approach... You actually get a lot of juice out of that already. That one's very quick to communicate. Then there's a lot of details that matter. It's probably going to be somewhat obvious in retrospect, or there's going to be some not too complicated thing that'll work, but there's going to be a lot of details to get that.”
Large-scale engineering to make training runs work is tacit knowledge that China will eventually figure out, so it's not a durable source of advantage, but self-play RL for overcoming the data wall is more novel.
“There's another layer at the bottom that is something about large-scale engineering work to make these big training runs work. That is a little bit more like tacit knowledge, but China will be able to figure that out. It's engineering schlep.”
Models learning through reinforcement learning mirrors human learning in crucial ways: just as humans transition from guided learning (pre-training) to self-directed learning (RL) in high school and college, models are entering a regime where they can learn independently through practice and self-play rather than relying solely on pre-training data.
“Bootstrapping is an advantage. Your Twitter bio is being pre-trained. You're not being pre-trained anymore. You were pre-trained in grade school and high school. At some point, you transition to being able to learn by yourself. You weren't able to do it in elementary school. High school is probably where it started and by college, if you're smart, you can teach yourself. Models are just starting to enter that regime.”
GPT wrapper companies are making a poor bet because they're betting on stagnation—assuming intermediate models will require extensive schlep to integrate—when actually the trajectory points to drop-in remote worker AGI that will make wrapper functionality obsolete, sonic boom-ing past the intermediate capability regime.
“I'm so bearish on the wrapper companies because they're betting on stagnation. They're betting that you have these intermediate models and it takes so much schlep to integrate them. I'm really bearish because we're just going to sonic boom you. We're going to get the unhobblings. We're going to get the drop-in remote worker. Your stuff is not going to matter.”
As AI capabilities improve, the influence of AI lab employees (who currently exercise enormous power, as seen in board events like the November OpenAI incident) will depreciate and vanish, replaced by concentrated control by a few people commanding armies of automated AI systems, requiring lab employees to use their current power wisely.
“Both are going to get automated and they're going to lose all their power. It'll just be a few people in charge with their armies of automated AIs... It's good for lab employees to be aware of this. You have a lot of power now, but maybe not for that long. Use it wisely.”
For AI clusters to reach hundreds of billions in cost, companies need to generate $100 billion per year in revenue, which is feasible through widespread productivity tools like AI-enhanced Microsoft Office.
“Once you're trying to get into the hundreds of billions, you have to get to like $100 billion a year in revenue. The question is how feasible is $100 billion a year from AI revenue? It's a lot more than right now. If you believe in the trajectory of AI systems as I do, it's not that crazy. There are like 300 million Microsoft Office subscribers. They have Copilot now. Suppose you sold some AI add-on for $100/month to a third of Microsoft Office subscribers. That'd be $100 billion right there.”
Most people in AI and outside don't think about AGI as a serious geopolitical competition because they're in a 'quiet period' where the trajectory hasn't yet become obviously dangerous, and will only wake up when catastrophe feels imminent (like March 2020 COVID realization), missing the opportunity to prepare.
“Most of the people who feel it are like right here... Why aren't other people talking about national security? I made this mistake with COVID. In February of 2020, I thought it was going to sweep the world... Most of the world just doesn't realize it.”
An AGI system could function as a drop-in remote worker through multimodal interfaces (Zoom calls, Slack messages), accepting projects, writing initial code drafts, receiving feedback, running tests, and returning results—with the key economic insight being that this 'overkill' capability makes deployment easier than intermediate systems requiring extensive workflow integration.
“You're interacting with them like coworkers. You can do Zoom calls and Slack with them. You can ask them to do a project and they go off and write a first draft, get feedback, run tests on their code, and come back. Then you can tell them more things.”
Some companies (reportedly including OpenAI leadership) have considered funding and selling AGI by orchestrating a bidding war between the governments of the United States, China, and Russia, which is analogous to playing them off each other and saying 'if you don't do this, China will do it'—a concerning approach that treats superintelligence as a commodity to be auctioned.
“I've heard from multiple people — not from my time at OpenAI, and I haven't seen the memo — that at some point several years ago, OpenAI leadership had laid out a plan to fund and sell AGI by starting a bidding war between the governments of the United States, China, and Russia.”
With 25% of compute in Middle Eastern dictatorships and potential seizure, the US-China compute ratio could drop to 3:1, which Leopold considers 'not that great' for maintaining a decisive advantage.
“Suppose we put 25% of the compute capacity in these Middle Eastern dictatorships. Say they seize that. Now it's a ratio of compute of 3:1. We still have more, but even with only 25% of compute there it starts getting pretty hairy. 3:1 is not that great of a ratio.”
2023 was the moment Aschenbrenner transitioned from viewing AGI as an abstract theoretical concept to seeing it concretely: observing the actual clusters where models are trained, understanding the algorithms, and recognizing the path forward, an insight most of the world has not yet grasped.
“2023 was the moment for me when it went from AGI as this sort of theoretical, abstract thing, and you'd make the models to like, I see it, I feel it. I can see the cluster where it's trained on, like the rough combination of algorithms, the people, like how it's happening, and I think most of the world is not; most of the people who feel it are like right here”