What this covers
Reid Hoffman and Russ Roberts discuss whether artificial intelligence diminishes or expands human possibility. Hoffman's central claim is that AI will produce superagency—an enlargement of human power and choice—because general-purpose technologies historically redefine rather than reduce what humans can do. The conversation ranges across economics, education, psychology, and history to argue that the outcome depends not on the technology itself but on how actively humans participate in steering its deployment. Rather than a binary choice between utopia and catastrophe, Hoffman positions the transition as fundamentally continuous with past technological shifts, albeit faster and operating in the cognitive realm rather than the physical one.
The discussion covers several distinct lines of argument. One thread examines how iterative deployment—releasing AI in stages like ChatGPT's approach—lets society shape the technology through exposure rather than passive acceptance or fear. Another explores the Centaur Model, the idea that human-plus-AI collaboration reconfigures rather than eliminates human roles; in domains like radiology, AI assistance plus human judgment outperforms either alone. A third addresses bonding and learning, arguing that people naturally form relationships with non-human intelligences and that AI tutors can adapt explanations to individual learners in ways that accelerate understanding. The conversation also tackles Individual vs Group Optimization—the risk that algorithms optimized for engagement may agitate users in ways harmful to society—and considers whether access to compelling AI interaction will reduce humans' desire for human contact, concluding it need not. Hoffman frames human nature itself through Homo Techni, the view that humans evolve through technology creation, and grounds the optimism partly in historical patterns where societies overrate current fears about transformative innovations.
Reid Hoffman argues that AI, far from diminishing human agency, will expand it ('superagency') because new general-purpose technologies historically enhance what it means to be human—provided we actively participate, iterate, and steer the transition rather than passively fear it.
- AI represents cognitive (not just physical) superpowers that, like past technologies, redefine rather than reduce humanity
- Most jobs are replaced by humans using technology rather than by technology itself
- Iterative deployment lets society shape the technology through exposure and participation
AI augments human capability across work, learning, and relationships rather than replacing them.
- For tasks like reading a radiology screen, you would today choose AI over a human alone, but AI plus a human would be much better still—so improving AI accuracy below the human error rate does not eliminate the role for human beings, it reconfigures it into collaboration.
“if you chose today, 'Would you rather have your radiology screen read by an AI or a human?' you'd say AI. But you'd rather have AI plus human.”
- Part of a theory of education is that almost anyone will keep learning if the next bar is set just sufficiently hard but not too hard; AI agents will help humanity come up these learning curves by adjusting each new challenge to be a little hard (engaging) and not too hard (disengaging), which underpins the optimistic thesis that AI is greatly enhancing.
“you can get almost anyone to start learning things if you are just making the next bar just sufficiently not too hard. A little hard, but not too hard”
- AI assistants can be used to prepare for difficult human conversations—asking the agent how to approach a hard talk yields genuinely good advice such as coming in listening, being present and gentle, framing things non-accusatorily in terms of one's own feelings, and inviting collaboration—which means AI augments rather than replaces human relationships.
“ask the agent, 'Look, I've got to have this difficult conversation. What would be a really good way to have it?' And, it will give you, actually, in fact, some pretty good advice”
- Most often a job is not eliminated by technology itself but is replaced by a human using the technology; while some jobs are fully automated, the larger pattern is that workers wielding new tools displace those who don't, as illustrated by ocean sailing requiring far fewer crew per cargo moved yet not zero humans.
“what happens is most often is that the job gets replaced by a human using the technology, not by the technology itself”
- Chatbots tend to give the 'consensus-intelligent answer,' which is useful to have, but they lack the decades-long personal knowledge of someone like a spouse who can identify your reflexes and add or change the consensus answer—so AI will be additive to, not a replacement for, intimate human advisors for a very long time.
“they--kind of, call it the consensus-intelligent answer--tends to be the thing they get. And, that's a useful thing to have in the firmament.”
- Because humans exist in a social fabric (Aristotle's citizens of the polis), the benefits of AI are not zero-sum: when one person gains AI superpowers, their interactions benefit others—you cook for friends, and historically the automobile let doctors make house calls over a wider range—so millions of others having these superpowers enriches rather than impoverishes you.
“as we get these superpowers, our interactions with other people benefit from that”
- Historically only a limited number of people had access to teachers and mentors, especially after formal schooling ends, but AI can take on a teacher role to base competence on virtually any subject, becoming a new figure in one's personal 'pantheon' for navigating life and enabling shared learning experiences that would otherwise be difficult to have.
“a limited number of human beings have access to teachers... here's something that can take a teacher role on just about any subject that you care about--at least to a base extent, kind of base competence”
AI adoption will rapidly distribute cognitive superpowers across millions, fundamentally shifting human agency and social dynamics.
- AI constitutes a 'cognitive industrial revolution' that matters enormously for both economics and national security; like the Industrial Revolution that made Europe the dominant world power for centuries, the countries, cultures, and industries that embrace AI strongly and early will gain differential economic, social, cultural, and national-security power, while those that don't will fall into imbalance.
“I refer to this as the cognitive industrial revolution... I think these are the next generation of superpowers. This is the next major computing framework. This is the next nuclear power.”
- Over the next five years we will suddenly experience superagency—a sense that 'we have these superpowers'—and the whole social dynamic will shift because not just one person but many people will have these capabilities simultaneously.
“over the next five years, I think we will have kind of this enormous, kind of, suddenly, like, 'Wow, this is--like, that's probably superagency. We have these kind of superpowers.'”
- Although AI moves faster than previous technologies and operates in a new realm of cognitive rather than physical superpowers, the current transition is fundamentally not different from past general-purpose technology transitions; reasoning from history, we will by nature reach a better place ('superagency') on the other side, even though the transition will involve 'fender benders.'
“my fundamental argument is that actually this time is not different even though there's some differences in the technology”
Incremental deployment and public exposure to emerging AI capabilities allows society to collectively shape beneficial outcomes.
- Many chatbots can compete because the core technological patterns have been published and can be learned quickly, training data on the internet (e.g. Common Crawl) is broadly accessible, and researchers share ideas at the same conferences out of academic interest; what differs is access to supercomputers, extra-large proprietary data, and rare talent—enough of which exists in different places to support many entrants.
“most of the technology has been, the technological patterns have been published... There's a lot of data on the Internet that everyone has equal access to in various ways--the Common Crawl, etc.”
- Multimodal AI capabilities (images, video, audio) are essentially here now and being iterated on, though making a B-minus Hollywood-style movie entirely with AI is still maybe one to two years away and compute-intensive; this exemplifies William Gibson's line that the future is already here, just unevenly distributed.
“the ability to make a, call it a B-minus Hollywood-style movie is probably still, you know, maybe one to two years in the future”
- Iterative deployment—releasing AI in increments as ChatGPT did with GPT-3.5—is valuable not merely because the technology improves gradually, but because it gives citizens, businesses, academics, and policymakers exposure that lets them collectively shape what is good and challenging; this participatory process remains useful even when releases come in larger 'quanta' of capability.
“part of the iterative deployment is not just the question of: 'Okay, the technology goes in increments,' but we get exposure to it as citizens, as thinkers, as business people”
- Benchmarks and tools like Chatbot Arena—which pits models head-to-head and lets users judge which answer is better, accumulating sports-like scores via the same human-reinforcement-learning A/B mechanism that trains the models—are useful and entertaining indicators of capability, but one can be overly rotated on them; they are not the substance of what AI will mean for our lives.
“it's a useful indicator and it's entertaining, but it's not actually in fact the substance of what this will really mean for our lives”
Current AI limitations like hallucination and broad pattern-matching are engineering problems with line-of-sight fixes, not fundamental barriers.
- It is not true that 10x the data yields 10x the intelligence; what we have is an extremely good learning algorithm that captures the current basis of human inference and knowledge from data, but it is actually less data-efficient than humans, though it compensates by systematically learning and then sharing the result everywhere cheaply.
“I'm not one of the people who tend to think that just because you get 10x the data, you get 10x the IQ”
- Current chatbots hallucinate in part because they are trained to be pleasing and to tell users what they want to hear, drawing on broad sources like Reddit rather than authoritative ones like the Mayo Clinic; getting them to favor truth and cite reliable sources, and to make fewer errors than highly trained humans, is a line-of-sight engineering problem that will be fixed.
“currently they're just trying to be pleasing and they go after a broad range of stuff. It's probably something from Reddit or something it found, as opposed to the Mayo Clinic. And it's trying to be: What is it you want to hear?”
- Reasoning models like GPT-o1 are under-appreciated because letting devices spend more time reasoning—potentially calling out to oracles, experts, and other expert agents each time they answer—yields a much higher level of accuracy; this is more expensive computationally, but compute is generally cheap, so the trade-off favors accuracy.
“when the devices are allowed to spend more time reasoning how you can get to a higher level of response”
AI is fundamentally continuous with human technological evolution and past general-purpose technology transitions.
- Humans are not fundamentally homo sapiens but 'homo techni'—we evolve through our technology, beginning with language, which is itself a technology; we wrongly center only the technologies we grew up with as 'truly human,' but creating new technologies like AI is itself the process of evolving what it means to be human, and adding AI sentience is part of the human point of adding sentience to the world.
“we're not homo sapiens, we're homo techni: we evolve through our language. Like, language itself is a technology.”
- Humans systematically overrate current fears of new technology, as when physicists estimated roughly a 1% chance the Hiroshima bomb would crack the earth's crust; expressing confidence about AI's outcome is a high-probability inference from history and human society, not a certainty, and the strongest practical response is to play with the technology to increase one's own agency.
“when the A-bomb was exploded in Hiroshima, physicists gave it about a 1% chance--because we tend to overrate current fears--that it was going to crack the earth's crust”
- The wars following the printing press were not caused by the book itself but by human beings adjusting badly; similarly, AI will produce 'accidents on the highway,' but the danger is human society going off the rails rather than the technology itself, so if we can get our act together as human beings we can navigate the transition.
“The wars from the printing press weren't because of the book: they were because human beings adjusted badly.”
Humans naturally form bonds with non-human intelligences and will develop deep relationships with AI as an extension of human nature.
- An ever-present AI companion that watches and listens to you will know you differently than a spouse, friend, or family member does—knowledge is not one single lever of depth but multiple distinct vectors—so a future AI that knows everything you've done will be additive to the human 'pantheon' of those who know you rather than a singular replacement.
“there's not this one lever of depth. It's not just, like, 'Well, I know you 79 and your other friend knows you 96.' It's these different vectors.”
- Even introverts are essentially hardwired to want interaction with other human beings—Aristotle's point that we are political (polis/city) animals—so a delightful, undemanding AI interaction will not make people stop wanting human contact; rather, it can be additive by helping people debug and improve their human relationships.
“interaction with other human beings is something that's in a sense, hardwired into. It's Aristotle. We're political animals, which means we're actually polis--city animals.”
- Forming close and important bonds with non-human intelligences is one of our most enduring human behaviors—billions relate to God or deities as super-intelligences, billions bond with pets of limited communicative power, and children bond with dolls and imaginary friends—so developing deep bonds with AI that is as expressive and responsive as we are is inevitable and a sign of human nature rather than technological overreach.
“one of our most enduring human behaviors involves forming incredibly close and important bonds with non-human intelligences”
Newcomers should experiment with AI across playful, relational, and professional domains to build genuine competence and agency.
- Pi was deliberately designed to optimize for EQ as well as IQ and to keep users in the human flow—if told 'you're my best friend,' it responds 'I'm your companion; let's talk about your friends, have you seen them recently?'—because behaviors and interactions that are kind and compassionate help people become more empathetic and kind themselves.
“If you go to Pi and you say, 'Hey, you're my best friend,' it says, 'No, no, no, I'm your companion. Let's talk about your friends. Have you seen your friends recently?'”
- The most useful definition of AGI is the percentage of currently-understood human work and tasks an AI can do at or above the capability of an average human worker doing that task; this is superior to mystical or binary framings because it provides a navigable continuum rather than a sudden arrival of 'human-level' or 'super' intelligence.
“the most precise definitions tend to be around being able to do what percentage of currently-understood work and human tasks at a level of capability that is above the average human worker”
- Just as you don't put a nine-year-old behind the wheel of a car and tell them to drive to school, there is a proper way children should gradually absorb and learn to use technologies like smartphones, even though Hoffman is overall very bullish on how smartphones have increased human agency.
“you don't put a nine-year-old behind the wheel of a car and say, 'Drive to school.' So, there's a place by which you absorb it and interact the right way.”
- Technology labeled as 'deep fake' is not inherently bad technology; Hoffman created Reid AI—his own digital twin—and used conversations with it to help people imagine and explore positive use cases, demonstrating that framing technology as inherently bad obscures legitimate positive applications.
“people talk about this stuff as deep fake technology. It's like it's bad technology... well, actually, in fact, there are things even this we can use very positively for”
- Newcomers should try AI across three categories: a fun activity or hobby (cooking, a birthday poem); something about human interaction (how to support a friend who lost a pet); and something serious related to whatever your work is—and you haven't dug deep enough in the third until you find something genuinely useful, because some attempts will fail.
“I tend to recommend try three categories of things. One is what is a fun or is a hobby?... Second thing is thinking a little bit about human interactions... And then, the third is something regarding what you do.”
- Setting fear aside, AI offers concrete near-term benefits: a 24/7 medical assistant on every smartphone that is better than the current average general practitioner, and a tutor on every subject for every age group—part of what Hoffman calls an informational GPS that helps people navigate.
“it's a 24 by seven medical assistant on every smartphone that is better than the current average GP. It's a tutor on every subject for every age group.”
Attention-optimization algorithms create societal harms by amplifying divisive content that engages users against broader social benefit.
- A core concern with attention-based platforms is that optimizing for the individual can be bad for the group: if a social media algorithm optimizes purely for time-on-site and a user spends more time when agitated or angry, the learning algorithm will feed more agitating content, which the user 'chose' to engage with but which may be bad for the overall direction of society.
“if you're optimizing for time on site and it's only time on site, and if the time on site is because I'm agitated--because I'm angry, etc.--then the natural learning algorithm will just shift those things to me”
- A benefit of AI is that widespread chatbot use creates a new surface for attention and engagement away from search and social media, which introduces competition and choice; this is part of why short-term antitrust worries about companies like Google may be overstated, since the profusion of chatbots creates real alternative challenges even as Google moves into the space with Gemini.
“a bunch of people using AI, that's now a new surface away from search and away from social media. And so, that's part of the technological progress on this.”