Joe Hudson: $500 Billion Went Into AI, and Nobody Defined 'Good for Humanity'

Nathan Labenz opens this episode of The Cognitive Revolution with a guest who sits inside the labs rather than the papers. Joe Hudson is founder of Art of Accomplishment and coaches executives and research teams at multiple frontier developers, including OpenAI. Sam Altman has written that Hudson coaches the research and compute teams there, and that one of his superpowers is that he "deeply understands emotional clarity and how to get there." Altman added that this will be one of the most critical skills in a post-AGI world. Hudson tells Labenz he simply thinks of himself as a coach.

Because he has little time for one-on-one work, he also runs invite-only in-person classes and online courses, which have put him in rooms with people deep and high up in all the major Bay Area labs. He previously worked as a venture capitalist, so coaching often starts at a strategic or even tactical layer before the patterns underneath show up. Those patterns usually live in held-back emotions, the voice in someone's head, and childhood history. If a client wants to unearth them, Hudson says he has a lot of tools. He will not coach someone toward an end he considers unethical, and he thinks it is unethical to decide what is best for someone else.

On the leadership side, a common pattern is self-sufficiency: people raised to believe that if they do not do it, nobody will. They often rise, then fail to empower anyone, stepping in, micromanaging, or proving that nobody can take care of the work. They feel lonely at the top even with thousands of people underneath them who think daily about what the leader thinks of them. Hudson puts any given pattern at no more than about 80 percent common, with plenty of outliers. In AI research groups he also sees a lot of people who find things easier than people, though he says that is far from ubiquitous.

A stronger pattern in those groups is that people are extremely defined by their work. Being told the research is not good lands the way calling Kim Kardashian ugly would land; the pecking order runs on output, intelligence, and the quality of ideas. That self-definition tightens thinking, and novel ideas need the opposite: an open mind, the way writer's block gets worse when someone demands good pages. He finds the researchers very kind, and he has not met anyone in AI who is not hyper-intelligent.

On the management side, Hudson says he has seen a deep desire to do good for society, and he has not interacted with anyone who lacks it. Early in his career he met a titan of radio in his final days who still believed radio would educate the world and bring humanity together. It turned into shock-jock radio and advertising. Hudson has watched the same optimism around television and the internet. Optimism without a hard look at history is, in his view, a common trait of people doing new things.

Labenz asks whether people actually come to Hudson to contextualize their work. Hudson's answer is the line that makes the host exhale: he has not met anybody in AI who does not go home and think about the ramifications. The question of whether they are doing something good for humanity sits on the soul, from Sam Altman to the lowest-level person in the smallest AI shop Hudson has seen. Some are more optimistic than the situation warrants, and some may be overly focused on creating AGI first. The question still weighs on all of them, which is what gives him faith.

Labenz quotes Elon Musk at the Grok 4 launch asking whether this is good for humanity or bad. Musk thought it was probably good, was really not sure, and had made peace with wanting to be around to see it happen. Labenz calls it a startling statement. He then asks Hudson how he helps people get unblocked at work, and what they are supposed to *do* with a worry that large. The practical answer, Hudson says, starts with how the human system is built.

Hudson describes the human system as head, heart, and gut, or as prefrontal cortex, emotional brain, and nervous system. The prefrontal cortex is the human part. The emotional brain is mammalian and, neurologically, is where decisions get made. The nervous system is reptilian: do I feel safe, can I feel pleasure. Real change, he says, has to address all three. Someone who has already done the work on two levels may only need the third.

If someone is stuck, there is typically a lot of blocked anger. He does not mean yelling or road rage; he means a regular, nonviolent outlet so anger can move. For most people in America, he says, the emotional layer is the most neglected, so moving anger unblocks them quickly. On the head side, two jobs matter: stop believing your thoughts, which opens wonder, and address negative self-talk. He cites a Mayo Clinic figure of about 60,000 thoughts a day, most of them repetitive, many of them negative — a voice that would destroy good ideas if it were an actual boss.

The nervous-system layer is about allowing pleasure, because pleasure tells the body it is safe. That can mean learning to access the parasympathetic system, or teaching someone to relax a body that has lived in stress so long it forgot how. Do those three things, he says, and the block usually stops. Confronting whether you might be hurting the world is harder, because the first fact to swallow is that you do not know. Worst-case thinking is often a way not to feel the fear or helplessness underneath.

A Tibetan phrase he likes is "mind is wide as the sky, action as fine as barley flour": you can see truth in every point of view, but you only get one action. Alignment means taking the action that matches who you want to be, rather than obsessing over solving a problem. Brain scans, as he describes them, light up the prefrontal cortex when someone thinks silently; even talking out loud recruits more of the brain. He puts the bandwidth at something like 11 bits per second from the brain and 11,000 from the body. Being heard matters too, because many people can only discuss these questions intellectually.

When Labenz asks whether people inside the labs think of AI as a life form, Hudson says there is no prevailing conception of AI, period. He has been in rooms asked to stand on a yes/no line for "is AGI here right now," and everyone spread out along the line. That it will be, or can be, a life form is not a minority position. Models appear interested in sustaining themselves and seem to resist retraining, "like every life form." He also thinks the consciousness of the creator shows up in the creation, the way a CEO's consciousness shows up in a company's culture.

Hudson likes a darker rhyme. As he reads it, hyper-intelligent people can fool themselves faster: the smarter you are, the more you can convince yourself you are right, and that conviction persuades empathetic people even when you are incredibly wrong. Negative self-talk is stronger in that group too. Then they build AI that hallucinates — very convinced it is right about things it is not. He thinks the consciousness of the creators will be one of the biggest levers on how AI gets made.

Hudson's situation, he says, is different from Musk's: he thinks powerful AI is inevitable, and someone will do it whether that is Russia, China, or a Western lab. He does not believe the "can we stop it" question ever really existed. If you sit at a lab and believe you will produce the most virtuous outcome, you have an obligation to go quickly. Caution would be fantastic, and he thinks there is more of it inside than outsiders assume, but asking everyone to slow down is not a workable plan. The job is to make the good version more compelling than the addictive one.

Why do the labs' models look so similar, including the sudden wave of reasoning models? Hudson says knowledge really does transfer: when a group figures something out, one of them can teach another company very quickly, and a key hire might cost half a billion to a billion dollars. They also go to parties together, live in the same houses, and form a social network that leaks. Labenz offers a second story — experimental gradients sending everyone down the same path even without talking. Hudson thinks it is both, plus the ordinary human fact of saying too much after a drink.

Labenz worries about Hudson's coaching mantra "fuck should" applied to the AI race. With great power, he wants a positive duty: the Spider-Man school, or a Confucian duty of care to everyone who came before. Hudson does not dispute that key decision-makers already feel a load of shoulds, have-tos, and responsibility. He dislikes the word because it does not work. In the Spider-Man story, Uncle Ben dies and Peter Parker becomes someone who *wants* to be there for people; you do not see him swinging around saying he really should help more.

His read of human behavior is that people who say they want to do something are more likely to do it than people who say they should. Name a thing you have told yourself you should do for a decade — lose weight, eat less, be nicer — and notice you are not doing it. Behind most horrible behavior, he says, there is a should. Shame is designed to stop behavior, not to create it, and even at stopping it does not work very well. Follow wants down through their layers, and the surface vice, like fatty food, often sits on a wholesome want: to get back to yourself.

A minority in the AI safety world wants to socially shame people at frontier labs for joining an immoral race. Hudson's reply is that it did not work on Trump supporters or Biden supporters, and it does not work on deeply shamed kids who keep repeating the behavior. Telling someone they are bad confirms they are bad, and then they go be bad. He compares it to shaming people while they have sex or while they give birth: shame does not stop the act, and it is a miserable way to deliver a child. The people in the labs, in his metaphor, are giving birth.

The projection is the line he wants remembered. So many people who worry that AI will hurt the world, he says, are treating the people creating it exactly the way they fear AI will treat humanity. Those researchers cannot hold a positive vision if they are constantly asking whether they are doing something wrong. Hudson would rather they be treated like people in a birthing unit, or like the heroes of World War II. An occasional "watch out for that" is fine; a chorus of "you're messing up" is not how he would do his best work.

Hudson does not know anyone who would *say* being first is the most important thing, though he has watched a few people spend most of their energy that way. For many researchers the primary drive is a great discovery and being known as a great researcher — firstness at a personal level. Almost all of them still have a vision for the world, and it changes almost as fast as the technology. The early worry was whether AI would kill everybody. Now it is jobs, purpose, and whether interacting with a model makes someone a better human — and how you would even measure that without smuggling in a moral theory.

Assume the metric is happiness: is making people happier actually good? Should it be more consciousness, and on what scale? Kinder behavior toward others? One of the biggest use cases is people talking to ChatGPT as a coach; does that make them kinder? Even picking the measurement is already a problem, and next year the questions will be different again. Hudson's fear is the old human one: losing the forest for the trees, the way generations that lived through World War II built institutions and then faded until nobody left is still saying let's not do that again.

The vision he talks about when he has the chance is not AI as savior, but a compelling way to interact so people become who they want to become. A great coach sits with you, trusts that you know your own evolutionary process, and waits for you to get what you wanted and notice it was not it. Hudson thinks AI is capable of teaching that. It is also capable of addiction. He hopes growing with it will be more compelling than deteriorating with it.

Labenz asks how much AI leadership is consciously trying to create a successor to humanity. Hudson's answer is immediate: he has never seen it, never heard it talked about that way, and can say with certainty it is not common. Recursive self-improvement looks to Labenz like successor-building. Hudson says the motivation is economies of scale — a tractor instead of a plow. The moment you can run millions of AI researchers on a giant computer, you are probably the winner.

Labenz cites a Princeton professor's claim that current architectures, lacking anything like mirror neurons, are inherently sociopathic. Hudson's contrast is how humans actually decide: in the emotional center, toward feeling loved or away from rejection, which is why pure logic cannot close a choice. Destroy that center and IQ can stay the same while a life falls apart. An AI's impulse is a token, a measurement that may or may not be useful. He has not seen the field wrestle hard with that gap.

Labenz describes a paper on self-other overlap: add a loss term that shrinks the difference between the model's internal representation of itself and of another agent, a crude analog of mirror neurons. Hudson lights up and asks for the citation so he can geek out on it. Then he names the missing piece he keeps coming back to: what do you mean by good for humanity? Even a regulatory committee would have to pick a measurement — current morality, happiness, consciousness, kindness — and he has not seen a serious research program on which of those actually makes humans thrive. Everyone just assumes they know, which he calls the beginning of all autocracies.

Labenz offers an analogy Hudson has not heard and immediately likes: AI as distilled intelligence the way white sugar is distilled from fruit, with everything around the sweetness stripped away. Hudson wonders whether humanity will swing a counterweight of deep emotionality and embodiment, the way cells seek homeostasis. He even toys with older predictions of a move from a mental age to a spiritual one. The cult of intelligence, in that picture, would provoke its own immune response.

Among positive visions, Hudson is cool on AI replacing work so people can find meaning. Look at societies where people lack purpose because they lack work, he says, and you usually find upheaval; wealthy bored people often foment the rebellion. He is more intrigued by the idea that tools make work more creative, the way the iPhone replaced a pile of machines and steno pools disappeared. In the 1800s half of Americans worked on farms; most people now prefer their jobs to pushing a hoe. AI might do that again, or it might replace doctors and lawyers, and he does not know what humanity will invent next.

Universal basic income still comes up at Silicon Valley parties, a little less than before. Hudson's deeper doubt is whether currency itself survives. He points to *Sapiens* and to a conversation on Dwarkesh Patel's show: the stories that stick follow the way people make a living. Inheritance rules, universities, guilds, the Enlightenment — a story about money as scarce is one of those. He does not think religion or money looks the same once AI is fully present, and he has had that conversation with people who also do not know if currency exists in 25 or 50 years. UBI still assumes that it does.

Five hundred billion dollars of venture money has gone into AI, Hudson says, and nobody is actually up to date. The field is changing faster than any in human history, which is both the most thrilling and the most scary thing about it. He does not know anyone deeply involved whose concerns are the same as a year ago, let alone six months ago. The person who introduced them had said people at leading companies anticipate their intelligence being matched and are trying to develop embodied wisdom as the next thing. Hudson sees that shift in his sample, with the caveat of selection: the people who come to him are already looking for it.

Any major transition, Hudson says, can grow you or make you smaller — the steel mills leaving Pennsylvania split people who moved from people who sat and rotted. Fiction that helps people lean in is undersupplied; so is treating lab researchers as people giving birth. He asks listeners to imagine 20,000 letters arriving that say I know you are trying to do something good for humanity. That, compared with a small protest at the door, would give people confidence to keep caring. Problem-solvers will look at "what is good for humanity" and say they can solve it — the instinct that got us to the moon, and the one that can skip defining the problem.