A thousand people on guard, a million in the sprint

Nathan Labenz's guest is Ben Todd, co-founder of 80,000 Hours and author of a book of the same name, reissued on May 26, 2026 and rewritten for the AI moment. Looking back a decade, Labenz says you would have a hard time finding a better briefing for the present. The group warned about pandemic risk years before COVID, pushed people toward AI safety before ChatGPT, and once gave Labenz free one-on-one advice when he left entrepreneurship to study AI full-time. This episode is not a substitute for the book. It is a map of how Todd now wants a listener to spend a career.

The number that organizes the map comes later, when they rank problems. Loss of control of human-level or stronger autonomous AI is still on top: irreversible, and in the limit a path to total human disempowerment. It is less neglected than it used to be. It is not staffed. Todd's working estimate is 1,000 to 2,000 people on those risks full-time, against an AI industry — depending on where you draw the circle — of 100,000 to a million people making the systems more capable, faster. That ratio is why the rest of the advice exists.

The name is arithmetic. Two thousand hours a year for forty years is about 80,000 hours, and Todd's claim is that this is the biggest decision you will make for your own life and, even more, for your impact on the world. Public talk about "making a difference," he says, still clusters on fair-trade coffee, turning off lights, and cycling to work. A career is about half of an adult's working life. A small improvement in where those hours go dwarfs the sum of the little helps.

Career advice, as he found it, did not match the stakes. It was success stories and slogans like "follow your passion," not research. The organization started because a group of Oxford students in 2010 could not get a serious answer to what they should do with their lives. They researched it, gave a talk to about 24 people, and six of them later changed course entirely. Some of those six told them to start an organization. Todd went full-time after graduation and spent the first ten years of his career running it.

Labenz offers a smaller version of the same math. If the next job will last at least two years, two extra months of search that raise pay by 10 percent already pay for themselves — even if you count only selfish interest. People still rush because idleness feels unbearable. Todd is running that instinct at career scale, and aiming it at impact rather than salary.

80,000 Hours picks problems that are big, neglected, and potentially solvable. AI was on the list early, partly by geography: in Oxford they overlapped with Nick Bostrom, whose *Superintelligence* came out in 2014–15 and whose arguments were older. Some of that case was still a Kurzweil-style compute story — enough chips, eventually a human-level system, a huge deal even if it took fifty years. The temperature rose in 2016. Todd wrote a short call to arms after AlphaGo beat Lee Sedol. Deep learning looked like a paradigm that would keep producing breakthroughs if you drew the trend line forward. They did not forecast that language models would be this good. The broad trend held. OpenAI had just been founded. Capital was betting.

The other side of the scale was how strange the concern still sounded. Andrew Ng's line — worrying about AI taking over is like worrying about overpopulation on Mars — was, Todd says, the attitude of a lot of people in the field around 2014. Trend, enormous stakes, almost nobody on it: that combination was the reason to bet. They did not need a precise date.

Labenz asks how AGI-pilled Todd is now, because career advice splits if some people think they have one phase left. Todd says he is pretty AGI-pilled, more so with every development, though guts move slowly. He plans against three sketches. The short one is the *AI 2027* / frontier-lab plan: automate AI R&D in one to four years, an algorithmic feedback loop compresses five years of progress into six months or three, and fairly general autonomous AI shows up around 2028. That world is strange because powerful AI arrives before most jobs are automated and before robotics, so deployment is chaos. The medium sketch delays automated R&D into the early 2030s as compute scaling slows in the late 2020s — fabs are full, new plants take years — and the explosion, if it comes, is driven by building chips, three to ten years instead of six months, with really powerful systems by the end of the 2030s.

The third sketch is a plateau: the current paradigm runs out, compute gets too expensive, revenue cannot fund the next jump, and you wait for a new paradigm or for ordinary growth to buy more chips. Todd calls that increasingly unlikely and not his main bet. Labenz cannot find the wall that would cause it; Todd's candidate is compute. Even Daniel Kokotajlo's camp, on Todd's reading of the *AI 2027* forecasters, puts an 80 percent interval from 2027 to 2050, with 10 percent mass after 2050. Labenz's own calibration is the other way around: he slightly overestimated how long models would take to get this good, and underestimated how slowly real-world impact would arrive.

The career framework in the book, for comparing offers side by side, has five parts: the impact of the option; career capital (skills, connections, credentials, character); personal fit; exploration value — what the job teaches you about which jobs fit; and everything else that matters for satisfaction. Longer horizons raise the weight on capital and exploration. Shorter ones push you toward whatever helps in the next couple of years. Todd's conversion is the one that matters: the relevant "timeline" is not an AGI date. It is the years when you can have the most impact. For AI-focused people that is usually just before and after powerful systems, though a slow takeoff could still be wild ten or fifteen years later. Ajeya Cotra's marriage post put expected length at ten years. Todd would have everyone plan at least five to ten. A year spent to become 20 percent more productive pays back in four or five. Retraining in machine learning, or transferring into policy in a year, can move impact by more than that. You would not zero out capital just because the horizon shrank from forty years to ten.

Problems, then interventions. Importance is likelihood times scope, plus urgency — some risks have to be handled before others, or they unlock the later ones. Loss of control is first because the stakes if it happens are so large. Concentration of power is newer on the list. One path is an intelligence explosion: under ordinary exponential growth the gap between first and second project stays two months; under super-exponential growth the gap widens, and one company can pull into a digital workforce the size of a nation, more power than any firm has had. Another path is surveillance. Todd cites the U.S. Department of War complaining that it was not allowed to use Claude to spy on every American — not a staffing problem in the old world, a data-analysis problem AI can close. Wannabe dictators, he says, will like that. The neglected questions underneath are almost more so than alignment: what instructions a more powerful model obeys, and whether a CEO can simply tell it what to do.

The third is engineered pandemics, possible eventually without AI and earlier with it. A pathogen worse than anything natural is on the table. A state like North Korea might build one for deterrence — mutually assured destruction that is easier than a thousand warheads. Lab leaks happen all the time; if these things are built, they probably get out. Labenz notes that 80,000 Hours has long been associated with downside work, and asks whether the upside of AI just takes care of itself if we get through. Todd will take whichever side has the best opportunity. Speeding up a glorious future by a year, the argument Dario Amodei makes in *Machines of Loving Grace*, is the first half of Bostrom's "Astronomical Waste." The second half is the one that matters: avoiding an extinction risk protects the whole future, not one extra year of it. Capabilities already have on the order of a million people. The downside has thousands. That is the case for focusing there now.

Todd's simple method is a short list of plausible attacks on those problems, then pick on personal fit. The four buckets that cover a lot of people are technical research and engineering; government and policy; communications; and organization-building — management, legal, HR, recruiting, the skills that make other work happen. Specialists still count: lawyers, economists, biosecurity engineers, even historians. Labenz's example is Mark Humphries, who used models to transcribe untouched Canadian archives and then wrote a widely read Gemini 3 note because his visual-document problem was unlike anyone else's. METR sits in the first bucket as the leading measurer of how close we are to automating AI R&D. On Odd Lots they said they had twenty useful projects and capacity for two or three. Labenz's report from a San Francisco event is that measurements are saturating and models are racing past the ability to measure them.

Whether to do that work inside a frontier company is, Todd says, complicated. The labs have some of the strongest alignment and control teams, and they can put research into the product, which is where it actually helps. Redwood Research is his counterexample on the outside: they helped pioneer the AI control agenda and ran deceptive-alignment work with Anthropic. The case against going in is that you may speed development and pull risk forward. Labenz adds the "people on the inside" argument — without a treaty this is happening, and a few people who care at the critical moment are high leverage — and the claim that alignment, more than interpretability, jumps qualitatively at the frontier, so open-weight work on a GPT-OSS-class model is a different activity from aligning the next closed system.

What actually drives the disagreement, Todd thinks, is p(doom). If extinction is likely and alignment has no hope, the only move is an indefinite pause, and a lab job makes that worse. If this is happening, cannot be stopped, and the job is to raise the odds of getting through, you want people inside. Conclusions split because premises split. He would still rather avoid adversarial strategies, but helping a more safety-conscious company win is not a strategy he rules out. A world in which nobody who worries about safety works at any top lab also scares him. Diminishing returns cut both ways: the first ten people inside can pick low-hanging fruit and raise alarms. Most people cannot last somewhere they reject the mission; nearly everyone especially safety-concerned has left OpenAI over the years. If you are the rare person who can stay and move the margin, consider it. Do not assume you are.

Todd puts the bias on the table. It is very convenient to decide that your highest-impact job is at a famous, rapidly growing, highly paid company everyone is talking about. Question the motive. Want a concrete strategy that has to be executed *there*, not an interpretability project that could run on open models. On the company, do not go on vibes. Treat the labs as political actors with mixed incentives, and get down to decisions: did they make the hard call, did they keep the commitment, what do people who know the leaders say about character. On yourself, cultivate friends who will call you out; most will not. Write the commitment down, and pre-commit the conditions under which you will reassess, because standards slide. You will become more like the people you work with, so pick the culture first instead of betting that you can hack a reckless one from the inside.

Labenz's image from an India event is Sam Altman and Dario Amodei refusing to join hands while everyone else on stage did — two people building almost the same technology with almost the same strategy, in a scene he calls fit for a Greek tragedy. The "somebody's going to do it, so it should be us" logic is how you get that mess. On the further claim that safety work is just an accelerant — RLHF made models safer *and* more useful, which pulled in more resources — Todd does not dismiss it. The question is whether the benefit covers the cost. A world with no safety research also looks scary. He wants a portfolio: some people building support for a one-year pause at the point an algorithmic feedback loop becomes possible, and a lot of people preparing for the world where there is no pause. Betting 100 percent on one scenario leaves the other on the table. Getting hired, meanwhile, still starts with asking people at the company what a three-month crash course should contain.

Policy is the other black box. A strategic pause is unlikely under the current U.S. administration; in a couple of years, with a possible feedback loop in view and a new administration, the coalition could look different. China would have to be in the deal — "we pause if you pause" — and Todd thinks Beijing has more incentive to take it because it is behind, and because the top of the party has already talked about AI risk. None of that works without compute tracking: if you cannot see where major training happens, you cannot enforce. An off switch for a lot of compute, if an autonomous system is replicating across data centers, is a low bar we currently lack. Transparency so a company cannot run a secret intelligence explosion for months; red lines and emergency plans the industry has already agreed to. Proposals still need fleshing out, and new approaches are only a couple of years old. Political will is the other gap. Few people in Washington, Todd gauges, could explain METR's time-horizon chart. Popular anger at AI is flowing into banning data centers, which he thinks mostly moves them to the UAE.

Pandemics are, in Todd's view, the under-debated engineering agenda. Alignment arguments never settle; here people agree that wastewater monitoring, red-teaming gene-synthesis companies, better PPE stockpiles, filtered positive-pressure air, UV, and rapid vaccines would help. AI can watch synthesis orders and run surveillance. The deeper pitch is that people who already know how to build systems with AI can pick up those projects as startups. A 10 percent or 80 percent fatality pathogen will not keep grocery shelves packed the way COVID did. The work aimed at worst cases still has only a couple of organizations fully on it. Entrepreneurial skill — making hard things happen, building coalitions — transfers into government if you can stand the frustration; so do technical-analysis roles for people who want applied work rather than academia.

Join versus found is not a funding problem. Show up with a good organization and you can raise quickly. Coefficient Giving used to dominate; new funders have entered, and Anthropic's founders have pledged 80 percent of their equity — tens of billions eventually, not all at once. Entrepreneurial personalities underrate joining: starting from zero feels real, making a working org 5 percent better does not, and the second is often worth more. Top-down founding, Y Combinator-style, is hard in competitive for-profits and more viable in nonprofits, where a funder and a gap can be enough. Catalyze Impact is an accelerator with a meta-list of some twenty project-idea lists. Horizon Fellowship moves technical people into policy in one or two years. Labenz adds Halcyon for mid-career switches. If your skill is applying AI, Todd wants tools that get stronger as the problems get worse: automated fact-checking, pundit track records, and impartial AI advisors for officials rather than systems built by the companies those officials are supposed to watch — an idea Forethought Research has written up. Using AI to make malaria programs 10 percent more efficient is good; taking AGI seriously points more at a grand bargain that shares benefits and slows the race.

The neglectedness treadmill is the last method: as a topic fills up, move one step past what is already accepted. Concentration of power was barely discussable two years ago; Forethought and a few others got it onto San Francisco agendas. Next on Todd's list is digital minds — when AIs talk like people, a rights debate is coming, philosophy of mind will not be solved in time, and policy has to balance views we cannot settle. Space governance is the one his editor wanted cut, which he took as a reason to keep: self-replicating AI probes and a default land grab for the 99.9 percent (with many nines after) of accessible matter and energy that is not on Earth, plus a possible first-mover industrial base that cannot be dislodged. Gradual disempowerment is the case where aligned AI and no tyrant still compete humans out of the economy, and the current plan is to ask the aligned systems how to stop that. Labenz's utopian-fiction idea gets a cold reply: the track record is bad, yesterday's utopias read as dystopias, and William MacAskill's "viatopia" is the better target — not an end state, but a civilization that still has options, no irreversible existential risk, no authoritarian lock-in, information still moving. Lost in wilderness, find water and high ground first.

The practical close is 80,000 Hours itself: a job board of about a thousand openings, plus funding and fellowships, and free one-on-one advice that once sent Labenz several introductions. The book covers what makes work satisfying, why "follow your passion" fails, which skills matter under AI, how to choose a problem, and the usual job-hunt errors. Todd's last question is how you want to remember this stretch. "I made a bunch of money so I would not fall into the permanent underclass," or "I did the best I could to help this go well." You only have one career. Labenz's version is the one you give a grandchild who asks what you did in the great AI transition.