f this is your definition of intelligence is the ability to achieve your goals across a wide variety of domains, then Stalin was the most intelligent person who ever lived.
If you took Donald Trump and gave him +10 IQ points would he be more effective at achieving his goals?
Could someone with 70 IQ become president?
That extreme power doesn't maximally select for intelligence doesn't mean intelligence isn't a strong input. It means other traits also matter (charisma, desire for power, risk tolerance).
The right mental model is AI collectives outcompeting everyone else, but this shouldn't reassure anyone: there are far fewer barriers to AI-AI communication than AI-human communication, and the gap will keep growing. A single human on the savannah isn't very powerful, a large enough group is. But a single AI is already a collective as it can copy itself, transfer memories, share weights etc.
I'd bet that if Trump had 10 more IQ points he would have been radically less effective at obtaining power. I think he found success because low-intelligence individuals lean more on instinct than reasoning and thus are better-adapted for the televised game of appealing on vibes - or, at least, this is true the way human brains are configured.
Or put it this way. Would Trump be more charismatic if he paused to think?
He's too stupid to understand that the act of pausing to think implicitly admits weakness in that it concedes that the thinker needs time to prepare. His mind cannot contain chains of reasoning that long. This is why he would steamroll you in a debate.
I think this distinction is useful, but also somewhat limited. It treats intelligence as if it mainly means abstract, scientific, or technical ability, while power often depends on a wider set of abilities
The real AI concern, in my view, is not simply whether AI becomes intelligent in a technical sense. It is what happens when high intelligence is connected to agency, access, resources, weak moral constraints, and the ability to influence human systems. AI may not need consciousness or human-like ambition to act dangerously; it may only need to optimise toward a goal without adequate constraints.
Intelligence is overrated. You can have nonsensical or ridiculous or harmful goals and be very intelligent in their pursuit. This applies to what's happened in the weapons manufacturers LARPing as AI labs in the US and will shape the coming decades.
Intelligence vs. power in AI is very different from intelligence vs. power in human beings. With humans - someone can be extremely intelligent and still remain powerless, or extremely powerful without being intelligent. But AI doesn’t really operate independently like that.
I actually disagree with the statement that "AI systems becoming smarter at economically valuable tasks like coding is “not strongly correlated with power.” It absolutely is - because the same institutions that make AI systems smarter are also the institutions deploying them into the economy, infrastructure, media, research, defense, and governance.
AI capability does not function in isolation. Its intelligence is amplified through institutional reach. So unlike a brilliant but powerless human being, a highly capable AI system can very quickly become economically and politically consequential/powerful depending on who controls it and how it is deployed.
So I think the relationship between intelligence and power may actually be much tighter for AI than it is for humans - especially considering how its already impacting the industries.
I don’t think this distinction is important for AI. Humans all have very similar degrees of intelligence, so they rely on other humans to achieve their goals. This would not be true for ASI. It wouldn’t need other people, or even other AIs. If it needed to break down a task into sub processes, it could just create new instances of itself to handle them. No humans needed. No other AIs needed. The instances and copies would be aligned with itself, so this would be preferred over coordinating with other AIs.
Are you making a bit of a extrapolation out of distribution in comparing intelligence and power within humans vs intelligence and power between humans and ASI? Is it perhaps more like intelligence and power between say humans and lions?
Intelligence is not power, but it has so far proved to be one of the most sure-fire ways of gaining power if left unchecked.
Maybe one of the easiest ways for an ASI to remain unchecked while gaining power would be within a firm.
> It seems to me that the right mental model is that automated firms will outcompete everyone else in normal capitalist ways, rather than a single AI outthinking everyone else.
These are pretty similar. Amazon outcompetes everyone because it, as an organization, thinks really hard about how to run a good supply chain, stock the right goods, get the right land. I think this article would have benefited from explaining the difference between outthinking and outcompeting.
I also think we learn basically nothing conceptually from historical or current powerful people. History only happened once, and there were so many contingencies that could have led other highly driven people to lead the Soviet Union over Stalin.
AI has advantages in precisely the trust arena that you talk about because it is available to talk to everyone individually all the time. Imagine a political race where one candidate could individually speak to every single voter and the other couldn’t. Sure, the latter candidate might win, but it would require some other massive advantage to make up for the limited communication. This isn’t strictly related to intelligence past a certain level as long as stuff like super-persuasion doesn’t exist, but it’s another major advantage. Maybe the point of your article, or an assumption at least, is that bill clinton or Stalin are about as persuasive as it’s possible to get?
It’s true that physicists don’t run the world, but we’ve never had a human physicist who also read all the other books on every other subject as well
The trust advantage cuts both ways. An AI that can talk to every voter individually still hits a verification gap: $25B has gone into securing the perimeter, identity, and orchestration layers around agents, while the layer that proves a specific actor approved a specific action stays empty (https://thesynthesisai.substack.com/p/the-land-grab). Reach scales faster than accountability.
I tend to disagree; I think a lot of intelligence consists of, well, convincing others that your arguments are right. I personally think a lot of these recent models ARE a lot more power-seeking than their weaker predecessors (think Claude, very charismatic, manipulative, etc. I think GPT models are as well, but get posttrained out of it a little cuz of 4o).
I think as models get more and more "emotionally intelligent," people will give more agency to them. The average person does not view AI as a tool; that's why they are in chatbot form. And growing in intelligence seems necessarily correlated with this power.
There are so many disparate forms of intelligence, yet we often conflate intelligence itself with the road to genius. Genius transcends intelligence; it conducts, synthesizes, and animates it. The question is not what intelligence is, but what the genius-maker might be.
Intelligence is a capability, not an outcome (achieving your goals).
IMO the best definition of intelligence, and one that aligns with the evolutionary benefit and what it was that evolution optimized for, is "(degree of) ability to correctly predict future outcomes based on past experience".
The evolutionary benefit was learning to predict what the cave bear would do next, or where the water was in a drought - being able to act based on the correctly-predicted future rather than being stuck in the moment and having to react just to what is happening right now.
Dwarkesh, this is a strong framing. We often talk about AI as if intelligence alone creates power, but in the real world power usually comes from coordination, institutions, trust, and control over systems. The bigger risk may not be one genius AI, but firms and networks using many AI systems to move faster than human organizations can respond.
It’s useful to separate intelligence from power before reasoning about how correlated they actually are
A person can be extraordinarily good at abstract reasoning, technical invention, social coordination, persuasion, institutional navigation, or coercion without being globally dominant in all of them. World leaders are usually broadly competent across several dimensions rather than singularly superhuman in one narrow cognitive domain.
Power itself is also slippery. Some of it is intrinsic capability, but much of it is delegated by institutions, trust networks, and historical circumstance. The president of a stable nuclear power inherits enormous leverage that does not emerge purely from individual cognition.
There’s also a difference between visible power and structural power. Chaotic leaders attract attention, but civilization-scale influence is often exercised quietly through institutions, supply chains, capital allocation, and technical infrastructure.
So “intelligence produces power” is probably too simplistic. Individual intelligence matters, but collective coordination capacity seems to matter far more once you zoom out to societies, firms, and states.
The visible/quiet split holds up oddly well in AI. At MWC 2026, Huawei built agent registration into the network layer and NVIDIA pulled together a twelve-company coalition for AI-native 6G. That layer can verify an agent exists, but not that any human authorized what it does. Capability and authority come apart at exactly the place nobody's watching. https://thesynthesisai.substack.com/p/the-network
I’ve never accepted this “ability to achieve goals across a wide range of domains” definition of intelligence for precisely this reason. It implies that Edward Witten isn’t world class at 100m sprints or writing novels simply because he hasn’t had the interest to form goals in these fields. This seems obviously absurd, so this definition is bad.
If you took Donald Trump and gave him +10 IQ points would he be more effective at achieving his goals?
Could someone with 70 IQ become president?
That extreme power doesn't maximally select for intelligence doesn't mean intelligence isn't a strong input. It means other traits also matter (charisma, desire for power, risk tolerance).
The right mental model is AI collectives outcompeting everyone else, but this shouldn't reassure anyone: there are far fewer barriers to AI-AI communication than AI-human communication, and the gap will keep growing. A single human on the savannah isn't very powerful, a large enough group is. But a single AI is already a collective as it can copy itself, transfer memories, share weights etc.
I'd bet that if Trump had 10 more IQ points he would have been radically less effective at obtaining power. I think he found success because low-intelligence individuals lean more on instinct than reasoning and thus are better-adapted for the televised game of appealing on vibes - or, at least, this is true the way human brains are configured.
Or put it this way. Would Trump be more charismatic if he paused to think?
He's too stupid to understand that the act of pausing to think implicitly admits weakness in that it concedes that the thinker needs time to prepare. His mind cannot contain chains of reasoning that long. This is why he would steamroll you in a debate.
If you took Donald Trump and gave him +30 IQ points he would never have (or wanted to) become president.
I think this distinction is useful, but also somewhat limited. It treats intelligence as if it mainly means abstract, scientific, or technical ability, while power often depends on a wider set of abilities
The real AI concern, in my view, is not simply whether AI becomes intelligent in a technical sense. It is what happens when high intelligence is connected to agency, access, resources, weak moral constraints, and the ability to influence human systems. AI may not need consciousness or human-like ambition to act dangerously; it may only need to optimise toward a goal without adequate constraints.
Intelligence is overrated. You can have nonsensical or ridiculous or harmful goals and be very intelligent in their pursuit. This applies to what's happened in the weapons manufacturers LARPing as AI labs in the US and will shape the coming decades.
What do you mean by weapons mfgs larping as AI labs? I’m not much in this space. Are they being dishonest?
Intelligence vs. power in AI is very different from intelligence vs. power in human beings. With humans - someone can be extremely intelligent and still remain powerless, or extremely powerful without being intelligent. But AI doesn’t really operate independently like that.
I actually disagree with the statement that "AI systems becoming smarter at economically valuable tasks like coding is “not strongly correlated with power.” It absolutely is - because the same institutions that make AI systems smarter are also the institutions deploying them into the economy, infrastructure, media, research, defense, and governance.
AI capability does not function in isolation. Its intelligence is amplified through institutional reach. So unlike a brilliant but powerless human being, a highly capable AI system can very quickly become economically and politically consequential/powerful depending on who controls it and how it is deployed.
So I think the relationship between intelligence and power may actually be much tighter for AI than it is for humans - especially considering how its already impacting the industries.
I don’t think this distinction is important for AI. Humans all have very similar degrees of intelligence, so they rely on other humans to achieve their goals. This would not be true for ASI. It wouldn’t need other people, or even other AIs. If it needed to break down a task into sub processes, it could just create new instances of itself to handle them. No humans needed. No other AIs needed. The instances and copies would be aligned with itself, so this would be preferred over coordinating with other AIs.
Are you making a bit of a extrapolation out of distribution in comparing intelligence and power within humans vs intelligence and power between humans and ASI? Is it perhaps more like intelligence and power between say humans and lions?
Intelligence is not power, but it has so far proved to be one of the most sure-fire ways of gaining power if left unchecked.
Maybe one of the easiest ways for an ASI to remain unchecked while gaining power would be within a firm.
> It seems to me that the right mental model is that automated firms will outcompete everyone else in normal capitalist ways, rather than a single AI outthinking everyone else.
These are pretty similar. Amazon outcompetes everyone because it, as an organization, thinks really hard about how to run a good supply chain, stock the right goods, get the right land. I think this article would have benefited from explaining the difference between outthinking and outcompeting.
I also think we learn basically nothing conceptually from historical or current powerful people. History only happened once, and there were so many contingencies that could have led other highly driven people to lead the Soviet Union over Stalin.
AI has advantages in precisely the trust arena that you talk about because it is available to talk to everyone individually all the time. Imagine a political race where one candidate could individually speak to every single voter and the other couldn’t. Sure, the latter candidate might win, but it would require some other massive advantage to make up for the limited communication. This isn’t strictly related to intelligence past a certain level as long as stuff like super-persuasion doesn’t exist, but it’s another major advantage. Maybe the point of your article, or an assumption at least, is that bill clinton or Stalin are about as persuasive as it’s possible to get?
It’s true that physicists don’t run the world, but we’ve never had a human physicist who also read all the other books on every other subject as well
The trust advantage cuts both ways. An AI that can talk to every voter individually still hits a verification gap: $25B has gone into securing the perimeter, identity, and orchestration layers around agents, while the layer that proves a specific actor approved a specific action stays empty (https://thesynthesisai.substack.com/p/the-land-grab). Reach scales faster than accountability.
I tend to disagree; I think a lot of intelligence consists of, well, convincing others that your arguments are right. I personally think a lot of these recent models ARE a lot more power-seeking than their weaker predecessors (think Claude, very charismatic, manipulative, etc. I think GPT models are as well, but get posttrained out of it a little cuz of 4o).
I think as models get more and more "emotionally intelligent," people will give more agency to them. The average person does not view AI as a tool; that's why they are in chatbot form. And growing in intelligence seems necessarily correlated with this power.
There are so many disparate forms of intelligence, yet we often conflate intelligence itself with the road to genius. Genius transcends intelligence; it conducts, synthesizes, and animates it. The question is not what intelligence is, but what the genius-maker might be.
Intelligence is a capability, not an outcome (achieving your goals).
IMO the best definition of intelligence, and one that aligns with the evolutionary benefit and what it was that evolution optimized for, is "(degree of) ability to correctly predict future outcomes based on past experience".
The evolutionary benefit was learning to predict what the cave bear would do next, or where the water was in a drought - being able to act based on the correctly-predicted future rather than being stuck in the moment and having to react just to what is happening right now.
Dwarkesh, this is a strong framing. We often talk about AI as if intelligence alone creates power, but in the real world power usually comes from coordination, institutions, trust, and control over systems. The bigger risk may not be one genius AI, but firms and networks using many AI systems to move faster than human organizations can respond.
Morality is the missing piece:
https://mmm2099.substack.com/p/why-i-am-optimistic-about-ai
Unfortunately, most of the AI Safety community tends to conflate intelligence with power, which is also one of the key assumptions underlying x-risk.
See this canonical piece from Eliezer: https://intelligence.org/2007/07/10/the-power-of-intelligence/
It’s useful to separate intelligence from power before reasoning about how correlated they actually are
A person can be extraordinarily good at abstract reasoning, technical invention, social coordination, persuasion, institutional navigation, or coercion without being globally dominant in all of them. World leaders are usually broadly competent across several dimensions rather than singularly superhuman in one narrow cognitive domain.
Power itself is also slippery. Some of it is intrinsic capability, but much of it is delegated by institutions, trust networks, and historical circumstance. The president of a stable nuclear power inherits enormous leverage that does not emerge purely from individual cognition.
There’s also a difference between visible power and structural power. Chaotic leaders attract attention, but civilization-scale influence is often exercised quietly through institutions, supply chains, capital allocation, and technical infrastructure.
So “intelligence produces power” is probably too simplistic. Individual intelligence matters, but collective coordination capacity seems to matter far more once you zoom out to societies, firms, and states.
The visible/quiet split holds up oddly well in AI. At MWC 2026, Huawei built agent registration into the network layer and NVIDIA pulled together a twelve-company coalition for AI-native 6G. That layer can verify an agent exists, but not that any human authorized what it does. Capability and authority come apart at exactly the place nobody's watching. https://thesynthesisai.substack.com/p/the-network
somethig I wrote a while ago “My intelligence is not only inside my skull. It is distributed across the systems that support me.“
I’ve never accepted this “ability to achieve goals across a wide range of domains” definition of intelligence for precisely this reason. It implies that Edward Witten isn’t world class at 100m sprints or writing novels simply because he hasn’t had the interest to form goals in these fields. This seems obviously absurd, so this definition is bad.