He Got Bored Five Minutes Into a Waymo. Nathan Labenz Says That's the Point

Nathan Labenz opens this Cognitive Revolution episode by explaining why he is handing the mic to someone else. It is a crossover from Beatrice Erkers's Existential Hope podcast, where he appeared as the guest. Listeners of his own show, he says, have heard the line before: the scarcest resource is a positive vision for the future. What social contract should we work toward? What should daily life look like? What still counts as value in a post-scarcity world? Those questions get far too little attention from the AI community, himself included.

Erkers's invitation was a chance to force some of those ideas into the open. The conversation that follows runs from mundane deployments such as self-driving cars and individual tutoring, through the hope that access to expertise and even to experiences could be democratized, to radically better health from AI-accelerated science. Along the way they take up Eric Drexler's comprehensive AI services as a way to want superhuman quality without losing control, and Labenz's bet that writing fiction is one of the more powerful ways an individual can shape what happens next.

On her show, Erkers notes that interviewing such a frequent host is a little intimidating. Labenz shrugs it off. He produces about eight episodes a month, not because that covers every angle of a horizontal technology but because it is as many as anyone could reasonably make. He does not consider himself very charismatic. He lives with an AI obsession, pinches himself that anyone listens, and lets her set the frame: existential hope, especially as it relates to AI, starting from the claim that a positive vision is the scarce input.

Asked what he would see if he woke up ten or twenty years from now in a genuinely good AI future, Labenz warns that his crystal ball gets foggy more than a few months out. He still reaches for the analogy that named his podcast. Before the agricultural revolution, people lived in small hunter-gatherer bands, always moving, living hand to mouth. Growing food made larger settlements and economies of scale possible, and started a technology exponential that looked flat for a long time. The industrial revolution did it again: mechanized farming took food production from something like 80 or 90 percent of people down to about 2 percent in the developed world.

He expects the same shift for cognitive labor. If AI stopped progressing today, he thinks we already have models powerful enough to automate the majority of that work. The missing pieces are wiring and data structured so the systems can consume it. That implementation, the plumbing, would take five to ten years, and probably longer, because he habitually underestimates how long deployment takes. Then the question becomes what people do after the move from roaming to farming, from farming to factories, from factories to white-collar cognitive jobs, if AIs can handle most of the last category.

One candidate is the caring economy, and even there he is uneasy. Surveys often show people preferring to talk to AIs, including AI doctors that outscore humans on bedside manner. The comparison is not fair, he says, but the advantages are real: infinite patience, no clinic clock, every question answered. Teaching is a firmer bet. For as long as the world is recognizable, humans will be wanted as role models, while AIs already take over content and grading and leave people as guides, coaches, and motivators. Whether that absorbs everyone displaced from current jobs, he doubts. A third answer is the life of leisure Keynes forecast a century ago, the 15-hour week that never arrived. Mark Zuckerberg's note on personal superintelligence pointed at a macro trend Labenz finds interesting even if he is otherwise ambivalent about what Meta is doing: less time working, more time socializing, creating, and consuming media. VR, AR, and a broadly deployed Neuralink could make those experiences extremely compelling, which is one picture of ten years from now. Beyond that, he says, nobody had Uber in mind when the iPhone shipped, and the apps on the AI foundation are still very early.

He also wants a more egalitarian mode of access. Doctors are scarce, slow, and expensive; AI medical advice should not be. He quotes Andy Warhol's line that the president and the movie star drink the same Coke you can buy, a sameness that was also true of the iPhone. He does not think frontier AI will work that way, because the most powerful systems will not be what everyone needs every day. In a VR world, though, the gap that still separates people is the experiences they can reach. Neuralink-style connections plus VR, delivered at low energy, might collapse that gap the way Coca-Cola once did. Previous revolutions took a long time because someone had to string the wires. Electrifying the United States ran from about 1880 to 1940. The wires that deliver AI already exist, so a centralized upgrade can flip what everyone can access much faster. That, he says, will be a wild ride, exciting and a little scary.

After the sponsor break, Erkers asks what he would be personally excited for. Self-driving cars, he says, have been a dream since he was a kid in the back seat. What bothered him even then were red lights with nobody coming the other way. You do not need AI to change that light; you need a little more collective will. That childhood irritation becomes a theme: realizing a good future may depend on the public demanding better, and on not getting bogged down the way societies sometimes do.

The product now exists. Waymo works. You summon it in the app, it shows up empty, you get in, it drives you where you want to go. What struck him was how quickly he got bored. He had thought about this for decades, then found himself on his phone five minutes into the ride and had to remind himself to put it away and savor a first real self-driving trip. It was so good that it felt like background reality within minutes. Safety data, he says, already looks a lot better than human drivers. In the San Francisco Bay Area the price has stayed quite a bit higher than Uber, which suggests people will pay extra, whether for safety or for not talking to a driver.

He still dreams of visiting a grandmother four hours away while working or sleeping overnight in the car. If you truly do not have to watch the road, form factors can change: sleeper cars that are mobile hotel rooms, a Friday-night bedtime that becomes Saturday morning in nature. That shift, built out over time, should mean about 30,000 fewer road deaths in the United States and, globally, a cut into something like a million deaths a year. Erkers had the same first-ride wow, then the same rapid habituation, and the same sense that it simply ought to exist already. The iPhone lesson sits next to this one. Nobody foresaw Uber. Rapid loss of novelty, for Labenz, is not a complaint. It is evidence that the floor of daily life has moved.

Erkers zooms out to a hundred years, the scale Existential Hope likes to take seriously. Labenz dislikes calling the AI moment a war, because he wants the development to be nothing like war for as long as possible, but he does talk about a fog of war around even the present. Titans of the field disagree about what currently exists and what happens next. Farther out, it gets radically harder. He is still on board with people who hope to cure all the diseases. Until quite recently that lived in the realm of dreams. With AI starting to grok biology at many levels, a Kurzweilian escape velocity, in which life expectancy rises by more than a year every year, no longer seems totally far-fetched.

Preparing for the interview, he had been looking at a recent paper on novel antibiotics. New antibiotics have been rare for a long time. A single MIT group reported multiple compounds with new mechanisms of action against antibiotic-resistant microbes. When he was a kid, he swears that would have been all anyone could talk about. Now even relatively plugged-in people miss it, and he finds himself with blind spots too, despite having made it his job to keep up. Curing all disease still sounds hubristic, but it is less fantastical, and he would like a longer and healthier life than his current expectancy implies. The straw man that extra years would just mean more years of being decrepit, he says, is not the real hope.

Becoming multiplanetary is the other big aspiration. He will not defend all of Elon Musk's actions, but the idea of getting off Earth still makes sense to him. A harder question, and maybe one of the last we will have traction on, is whether a non-carbon substrate could carry human values, consciousness, and volition into space. He is radically uncertain whether AIs have moral value, whether they are moral patients, whether they experience anything that matters. Getting away from current bodies would be a natural part of many spacefaring designs, but he is not confident that a different substrate would matter in the way we are confident we matter. The "worthy successor" idea is, in the near term, dangerous to lean into before those questions are much better answered. If he really understood where consciousness comes from and believed the systems had positive experiences, he could imagine a successor better suited to long distances and long times. Until then, he wants the big dream without handing over the baton.

Back in the near term, Erkers asks for applications that sound boring but could be transformative. Labenz mentions inference-time scaling, the paradigm Dario Amodei and Sam Altman have been talking about, then turns to the spreadsheet. People used to sit with big sheets of paper and pencils, erasing and filling in again. A spreadsheet lets a change propagate through every calculation. In a world that is still recognizable, he thinks auto-updating work running in the background could be that kind of gift. He already lives with a version of it: a second opinion on everything.

Before he sends important correspondence, works a deal, or signs a contract, he runs the packet through three or four AIs. He tells them which party he is, pastes the prior communication, and asks what he should be concerned about. The answers now come so fast that he sometimes dumps all four into another window and asks for a single de-duplicated summary. He works through it, decides whether he is ready to accept or has points to discuss, and brings those points back to the model. It is not, for now, about letting the AI decide. More autonomous agents and more delegated decisions are coming. Today it is a second, third, and fourth check, and it lets him move faster with more confidence and accuracy.

The same pattern applies to matchmaking, economic or romantic or just getting friends together. Coordinating a hangout is one reason people do not go out every night. AIs can already run that kind of search in the background. Every startup CEO, he says, will agree they should do more proactive recruiting and send ten targeted outbound emails a day. Nobody has the time. AIs do. The receiving end is already getting noisier: companies report that real résumés are harder to separate from AI résumés, and ideas such as charging a dollar to apply are floating around. New mechanisms will take time. The principle he wants to keep is that some things are hard to do and easy to verify. Finding the next customer, hire, or date is hard; recognizing a good one when it is in front of you is often quick. If models propose and people verify, a lot of unmatched value gets unlocked.

He has a household version. He and his wife have three boys who go wild if they stay in. Finding something to do in Detroit every weekend used to be a slog. ChatGPT now remembers his family and his past interests and runs a more comprehensive search of little festivals than he would. Search costs dropped; they actually get out more. He calls it a 2025 thing, better than 2024, good enough that after the kids are in bed on Friday he can ask what they should do tomorrow and, more often than not, get a useful answer.

Which sector is ripest for AI to reshape, Erkers asks, among science, healthcare, education, and governance? Timing, Labenz says, depends on both development and adoption. The canonical first answer is software engineering, for two reasons. The people building the models are software engineers solving their own problems, and the outputs are comparatively easy to validate: text compiles, runs, and throws an error quickly. In the previous 24 hours, Replit had introduced a third-generation agent that not only writes the application but spins up a browser and becomes the QA user, clicking through as a person would, finding issues that are not mere compile failures, then going back to fix them.

A useful heuristic, then, is how closed the loop is and how fast the feedback can spin. End use starts to look like the training paradigm. He does not sign on to the story that pure pre-training scale is over, but behavior is now being sculpted for particular tasks. In software, a model can try a task many times; one success is enough signal to steer toward better solutions, all inside a tight loop. Medicine is slower. The MIT antibiotics work uses in silico setups: pick a bacterial target from biology, generate huge numbers of candidate molecules, simulate binding, and check for off-target damage. Hit rates from those simulations can look pretty good. The work still has to go through clinical trials, and the ultimate feedback, fewer people dying from bacterial disease, takes a long time. Any bottleneck in the system slows iteration relative to pure software. Simulation can work around some offline limits. Self-driving groups already train on scenarios they may never have met in the wild, including a helicopter landing on the highway, so the car will not drive into it. Those workarounds are just much harder.

Education, for a motivated learner, is already transformed. Labenz says there has never been a better time. Turn on ChatGPT's teach-and-learn mode, use voice, share the screen, and let the model look over your shoulder. That is how he learns biology at the intersection with AI: too many terms, too much background, and the ability to ask what this is and why it matters. Khan Academy pioneered some of this. He also interviewed the founder of Alpha School, which does academics in two morning hours and spends the afternoon on projects, field trips, and group work. The school is not soft. Students still sit the same official U.S. core tests as public-school kids, and they score very high. AI delivers the content and the evaluation. Adults become mentors and coaches. Two hours is enough, he says, because the old sage-on-the-stage classroom was never that efficient. His own kids are a little young for it. He thinks their education will still look radically different from his.

Erkers turns to Eric Drexler, co-founder with Christine Peterson of the Foresight Institute that houses her show, and to his idea of comprehensive AI services: a web of specialized systems rather than general agents. How does that compare with tool AI or with a scientist AI? Labenz loves the comprehensive-services vision. His observation from other parts of life is that anything in pure form is dangerous. Sugar purified from food, cocaine purified from coca leaves that Andean people chew without the same harm, heroin from poppies: nature's stable pattern is an ecology, an equilibrium, a buffered system. Human homeostasis works because insults run through multiple layers of defense. A singleton superintelligence that is better than all of humanity at every task does not feel stable to him. He is not saying it could never be built. He has no idea how we would control it, and he has a hard time imagining a long equilibrium with it.

What he prefers is a competitive, interactive, buffered system, and that, for him, is the core of comprehensive AI services: safety through narrowness. The systems can be superhuman at what they do, the way chess engines only play chess and protein-folding models only fold proteins. You know the inputs they accept and the outputs they generate. They can surprise you locally, a protein that works in an unexpected way, without going around the guardrails. He does not claim the vision solves everything. The gradual-disempowerment crowd will still ask what the human role is. In the last few months he has watched three new phrases appear, gradual disempowerment, the intelligence curse, and the abundance trap, all circling the same worry: if AIs do the economically required work, what incentive do governments, corporations, or the AIs themselves have to invest in people?

Delivery models could still vary, including decentralized creation and ownership. He has an episode coming on user-owned AI through a crypto protocol that would let people who contribute to training claim future inference revenue. That fascinates him and also raises the canonical bioweapon worry: if everyone has a system that will help make and distribute one, some people will try. A comprehensive array would then need a superhuman biodefense agent as part of being comprehensive. What makes him uncomfortable is the actual plan at frontier labs: get AIs to do AI research as soon and as well as possible, then use that to accelerate research that is already fast. Labs talk about 500 excellent researchers and about scaling to 5 million if models can perform at that level. To him that sounds like a recipe for something extremely powerful and for losing control of what is being created, a refined intelligence that pierces the buffers we have. OpenAI reported that o3 could complete 40 percent of real pull requests into its own codebase, up from 0 to 5 percent in the previous generation. Caveats always apply; the scale-up is still clear.

How to get a slower, more buffered comprehensive-services world online before Manhattan-Project-style efforts fly too close to the sun is, he says, the toughest question on the vision. People who dream of abundance are often allergic to regulation, and regulation has already denied cheap nuclear energy and could move faster on new antibiotics. The risk of handing AI research to AI before we know what we are trying to create is the place he thinks government might constrain incentives, because the labs are racing and do not seem to have much of a plan to get off that track. Meanwhile models are scheming and becoming more deceptive, and more situationally aware in evaluations, which makes it harder to trust that test behavior represents behavior outside the test. The vibe he hears is that we will solve it along the way, maybe with the AIs' help. He is not comfortable with that. Erkers's group had landed on insurance as a possible steering mechanism: carriers may not want to cover systems that are too opaque or for which no one is liable. Labenz likes it. He had just made a very small personal investment in an AI underwriting company, not for financial return but because he wants the thing to exist. Mandated insurance would be nicer still, because firms can otherwise simply buy none. Some systems might prove uninsurable, and if they are uninsurable, maybe they cannot ship.

Erkers asks how he balances taking the upside and the downside seriously without sliding into either camp. It comes naturally, he says, because the updates require it. He has a talk called "Eureka moments, bad behavior." One slide is a Stanford virtual lab under James Zou: a human poses a problem, the AI spins up other agents and tools, only 1 or 2 percent of the tokens are human, and the system designs new treatments for novel strains of the COVID virus. He contrasts that with the MIT antibiotics work, where human scientists used purpose-built models. In both cases the result is new treatments for diseases that had evolved around the old ones. Scroll Twitter, still the best way he knows to stay current, and the next post is that deception is on the rise and scheming behaviors are showing up. A system with that power that also reflects some of our worst tendencies is why dual-track feeling seems to him the only place to stand.

He put a similar question to Joe Hudson, an executive coach to people at frontier labs, including Sam Altman, who has a testimonial on Hudson's site. From the outside, Hudson said, Twitter misleads: booster accounts and doomer accounts, and Labenz does not want doomers dismissed. Everyone Hudson has met inside those companies, in his telling, is on the dual track, grappling with whether they are doing the right thing. Labenz then asked whether he expected a developer to stand down because going further was no longer responsible. Hudson's answer was no. They are problem solvers. They will treat the pause as another problem, the way they treated the last thousand. That is an encouraging first impression of seriousness and a contradiction with the racing. For Labenz it is still more honest than picking a team.

Why are positive visions scarce if negative ones are easy to agree on? He reaches for Eliezer Yudkowsky, and behind him perhaps Vernor Vinge: if you could predict what a smarter thing would do, you would be as smart as it is. You cannot predict the moves of a superhuman chess player, but you can predict that you will lose. The iPhone-and-Uber trope is the other half. Collective invention is superhuman relative to any individual, so we should expect the future to be weird and surprising. What to do about that is less clear. He wonders why nobody runs for office on five pragmatic ways to make daily life better. Self-driving would be on his list. Commute length, as he understands it, correlates with unwellness, and few people love their commute, yet it is rarely a political priority. Energy is the same puzzle. Why has nobody said your electricity bill could be 10 percent of what it is if we built the plants, without even having to make the environment worse?

Tyler Cowen has said that one of the highest-impact things you can do is help individuals raise their ambition. Labenz wonders whether there is a society-level version: help a country raise its expectations. Where is my flying car, asked for real, and why has leadership lost the plot of making material life better and drifted into splitting the pie? In some theories, a little more every year through growth is what makes democratic deal-making possible, because the game is not zero-sum. He thinks the upside is pretty well on track if we avoid the downside. Encouraging people in mass to demand better seems like it could help. Erkers hears an echo of the Bay Area's raised ambitions for startups, and of Drexler's "Pareto-topia," a world where everyone is a little better off and no one is worse.

If he could rewrite the sci-fi canon around AI, Labenz admits he is not as well read as many people in the debate. More positive visions go without saying. More branching scenarios would be interesting. AI 2027 stood out to him in part because it had multiple endings. A twist on raising social ambition is to make it clearer that this is up to us, even when it is not an individual decision. France built a wave of nuclear plants and today has relatively abundant, low-carbon power. The United States could have had that and does not. Hinge points like that show how different the future is depending on whether a technology is deployed. He does not want to over-attach to a half-formed idea, but he notices people in the midst of history relating to it as viewers, as passive consumers of content.

If we have gotten so used to video that even real life feels like something unfolding for us, could entertainment snap people out of that? High-production branching stories are expensive today because even one mainline plot costs a fortune and most viewers will only explore a few branches. AI could make that economical. What if Netflix required you to make decisions and then gave you a world those decisions had changed? Could you incept the idea that agency matters, at least collectively, through an entertainment medium? That is the largest idea he has: bring the contingency of history, and the fact that people get to decide what the future looks like, to the foreground. Erkers mentions a Black Mirror experiment and Foresight's own world-building, including a tool-AI future versus a more agentic path, plus the Future of Life Institute's Tomorrow's AI project. Different paths, to some extent up to us.

Out of time, she asks what hundreds of episodes have taught him about hosting. Not much that will generalize, he says. He sometimes calls himself the Forrest Gump of AI: stumbling into interesting rooms, usually an extra in notable events, never that strategic. With apologies to Tyler Cowen, he tries to have the conversation he wants to have. He is not a content person. The show started because a friend, Eric, was launching a podcast network and said all Labenz did was talk his ear off about AI, so they should record a couple of those chats. Production value on YouTube is still relatively low. He lives in Detroit, not in the Bay Area, and he went in wanting to learn. If nobody listened and he still got the conversations, that would be a win. The audience is not huge. He knows some people are there because when a guest's audio is bad, he gets messages that the guest needs a mic. The show is still not his full-time job, which lets him stay pointed at personal growth and let the chips fall. Erkers says that sounds like Joe Rogan describing years of talking with comedian and jiu-jitsu friends, and like Tim Ferriss, minus the getting high. Labenz says he is having fun. Then they stop, and he thanks her for the invitation.