AI Timelines Shorten Career Planning but Do Not Eliminate Retraining

Benjamin Todd’s central career-planning move is to replace the abstract timelines debate with a personal planning question. The relevant question is not only when AGI or superintelligence appears. It is which years will contain the highest-leverage opportunities for a person like you to act, and how much time you have to prepare for them.

That reframing changes how short timelines should affect career strategy. Todd does not argue that people should ignore preparation because the window may be narrow. He argues that the planning horizon has shortened, but not collapsed. Even if AI is the dominant issue and timelines are short, the most impactful period may include the years immediately before and after AGI; in a slower takeoff, it could include 10 or 15 years of consequential deployment and adaptation afterward. His practical conclusion is that most people should still think in terms of at least five to ten years.

Todd’s career comparison framework weighs five factors: direct impact, career capital, personal fit, other contributors to satisfaction and personal goals, and exploration value. Career capital includes skills, connections, credentials, and character. Exploration value is what a job teaches you about which paths fit you best. As AI timelines shorten, career capital and exploration become somewhat less important because there is less time to use what you learn. But they do not disappear.

His example is simple: if spending one year makes you 20% more productive, you may recoup that investment over four or five years. On a five-year horizon, that can still be worth doing.

That is why Todd still sees room for retraining. Moving into machine learning, entering policy, or building a more relevant skill base may be worthwhile if the transition takes a year and the later payoff is much larger. The mistake, in his view, is importing a conventional 40-year career model into the AI transition. The opposite mistake is assuming that a short timeline means there is no longer time to reposition.

The original premise of 80,000 Hours is that career choice is the largest lever most adults have over both their own lives and their impact on the world. The name comes from a conventional career: 2,000 hours a year for 40 years. Todd contrasts that scale with moral attention spent on much smaller decisions — buying fair trade, turning off lights, cycling to work. Those choices may matter, but a modest improvement in the impact of a career can dominate them because work occupies such a large share of adult life.

The organization came out of Todd’s and other Oxford students’ attempt to answer their own question: what should they do with their lives if they wanted work that was enjoyable, financially viable, and socially useful? Existing career advice, in Todd’s telling, was mostly anecdotes from successful people or slogans like “follow your passion.” A first talk in 2010 drew about 20 people; Todd says about six later “totally changed their lives” after engaging with the ideas. That response became the organization.

The method Todd described is to look for problems that are big, neglected, and potentially solvable. That is also how he explains 80,000 Hours’ early attention to AI. The group’s exposure to Nick Bostrom at Oxford put superintelligence arguments on its radar before they were mainstream. In the mid-2010s, Todd says, the case was partly a broad argument: computing power was growing, and even if human-level AI were decades away, the stakes were enormous. Around 2016, the case became more concrete. AlphaGo’s defeat of Lee Sedol suggested deep learning was a highly successful paradigm; OpenAI’s founding showed serious actors betting on it; and the issue remained deeply neglected. Todd says 80,000 Hours did not predict large language models would become as capable as they have, but it did expect more important developments to come from the deep learning paradigm.

Nathan Labenz framed the present career problem around the possibility that some people may have only one more meaningful phase of work before AI changes the world radically. Todd does not offer a single timeline. He plans around three broad scenarios.

The first is a short-timeline, fast-takeoff scenario: something like AI 2027, or what Todd characterized as the sort of future people at Anthropic and OpenAI say they are planning for and expect. In this world, AI systems automate AI R&D in roughly one to four years, creating an algorithmic feedback loop. That could compress several years of AI progress into six months, one year, or even less, producing powerful, general-purpose autonomous AI around 2028 or sooner in more aggressive versions. Todd notes that this scenario has an unusual feature: powerful AI arrives before most jobs are automated and before robotics is fully deployed, producing a chaotic deployment process afterward.

The second is a medium-timeline scenario. AI that can automate AI R&D may take longer, perhaps arriving in the early 2030s. Compute scaling may slow in the late 2020s because existing fabrication capacity is nearly used up and building new fabs takes longer than shifting existing capacity. If no strong algorithmic feedback loop is possible, Todd argues there can still be an intelligence explosion, but it would be driven by building far more chips rather than by rapid recursive software improvement. In that case, the transition might take three to ten years rather than six months, with very powerful systems arriving by the end of the 2030s.

The third is a longer plateau scenario: the current paradigm runs out of steam, compute scaling becomes too expensive, revenue is insufficient, or the world waits for a new paradigm or for economic growth to produce enough computing power to push further. Todd thinks this is increasingly unlikely, but not negligible. Labenz was more skeptical, saying he could not find the “wall” that would produce such a plateau. Todd’s version of the case is that if current scaling reaches roughly 2028, slows by 2032, and still has not automated AI R&D, final bottlenecks may prove hard to clear. He also noted that even the AI 2027 forecasters have wide uncertainty intervals, with an 80% interval he described as roughly 2027 to 2050 and a nontrivial chance beyond 2050.

Todd’s current priority list starts with problems rather than job titles. He evaluates problems by importance, neglectedness, and solvability, with an added emphasis on urgency: some problems arrive before others, and some interventions make it easier to handle later problems.

The top problem on 80,000 Hours’ list is loss of control of autonomous AI. Todd’s reasoning is that if humans lose control of human-level or beyond-human AI systems, the result could be irreversible and could eventually mean total human disempowerment. This area is less neglected than it once was, but he estimates only about 1,000 to 2,000 full-time people work on these risks. By contrast, depending on where one draws the boundary around the AI industry, 100,000 to a million people may be working on capabilities or otherwise accelerating AI.

The second risk Todd emphasized is concentration of power, a newer addition to 80,000 Hours’ main list. He described several mechanisms. One comes from the dynamics of fast growth itself. If two companies are two months apart in a simple exponential race, the gap remains two months. But under hyper-exponential or super-exponential growth, Todd argued, the lead can widen. One project could pull far ahead, gaining a digital workforce equivalent to a present-day nation and accumulating more power than any company has ever had.

A second mechanism is surveillance. Todd gave a pointed example, saying that the “Department of War” was “pissed off” that it was not being allowed to use Claude to “spy on every American.” His broader concern was that AI could make mass surveillance much more effective. Historically, governments lacked enough staff to analyze all available public data about everyone. With sufficient AI, he argued, a government could trawl through public data and build detailed pictures of people’s activities. That would be useful to “a wannabe dictator of any kind,” as Todd put it.

He also argued that parts of the concentration-of-power problem are even more neglected than alignment. Very little thought, in his view, has gone into questions such as what instructions powerful AI systems should obey under different circumstances, whether a single company CEO should be able to tell such systems what to do, and what safeguards should exist.

The third major risk is engineered pandemics. Todd thinks such pandemics will eventually be possible even without AI, but AI could bring them forward. He argued that it seems possible to engineer a pandemic much worse than any naturally occurring one. A state such as North Korea, he suggested, might develop such a pathogen as a deterrent: invade us and we release the virus. That could provide a form of mutually assured destruction that is easier to build than the thousands of nuclear weapons once required for such leverage. If such pathogens are built, Todd worries they may also escape accidentally, because lab leaks happen.

Todd’s focus on catastrophic downside is not, in his account, a rejection of upside. His stated objective is impact, whether from increasing upside or avoiding downside. But he distinguishes between two ways of improving the future: speeding up a better future, and making sure the better future happens at all.

He invoked Nick Bostrom’s “Astronomical Waste” argument as a way to think about this. If the future could become vastly better, speeding it up by one year creates one extra year of that future. Todd noted that Jeff Bezos makes a version of this argument when he says AI should be developed quickly to bring forward a glorious future of space settlement and advanced technology. But Todd emphasized the other half of the argument: if avoiding extinction or permanent disempowerment preserves the entire future, the value of reducing that risk can dominate the value of slightly accelerating progress.

Since AI development is already moving extremely fast and already has enormous resources behind it, Todd thinks marginal work to accelerate capabilities is less neglected than work to reduce catastrophic downside risks. He framed this as a comparison of scale: perhaps a million or so people are effectively trying to accelerate AI capabilities, while the people focused on avoiding the largest downside risks are still in the thousands.

Todd’s broad advice for individual choice is to make a shortlist of plausibly impactful paths and then choose among them by personal fit. He does not think everyone should try to force themselves into the same role. The main categories he highlighted are technical research and engineering, government and policy, communications, and organization building.

Technical research and engineering remain central. Todd singled out METR, which performs evaluations intended to show what AI systems can actually do and how close they are to automating AI R&D. He called that question “maybe the most important thing going on in the world right now” and said measurement remains poor. According to Todd, METR has said it has about 20 useful projects it would like to do, but capacity for only two or three, because the work requires engineers to implement systems and evaluations.

He also listed AI control and interpretability. A notable shift, in his account, is that much AI safety work used to be bottlenecked more by high-level conceptual research, whereas now many projects are concrete engineering problems: can one AI monitor another AI, can red-teaming detect dangerous behavior, can deceptive behavior be caught quickly, and what systems are needed to make that work?

Government and policy are the second major category because Todd expects governments to be involved in any serious response to these risks. He sees a shortage of people who can straddle technical AI and government. That includes policy makers, advisors, technical experts in government, analysts, and operational people who can make institutions function.

Communications is third. Todd thinks broad understanding of AI remains low and that few people are working on communicating the risks, improving public understanding, or mobilizing talent. This includes explaining technical and strategic issues clearly enough that more people can act on them.

Organization building is fourth. The needed work requires institutions, and institutions require management, accounting, legal support, HR, recruiting, and general “getting stuff done” capacity. Todd emphasized that many useful contributors will not look like frontier researchers. Lawyers, economists, engineers working on biosecurity, and even historians may have valuable roles.

Labenz offered an example of the historian category from his own show: Mark Humphries, who Labenz said had used AI to transcribe and analyze old documents in Canadian archives. Labenz also described Humphries as having written a viral post on Gemini 3 just before it came out, focused on a capability to reason about visual documents in a way prior models had not. The point was not that historians are a mainline AI-safety category, but that as AI becomes a general-purpose technology, unexpected niches can contribute observations back to the broader frontier discourse.

The role taxonomy is therefore not “work in AI” versus “do something else.” Labenz noted that AI is a horizontal layer touching every sector. Todd’s view is that people should ask which problems matter most and then identify the roles through which their abilities can help solve them.

The question of whether safety-minded people should work at frontier AI companies does not get a simple answer from Todd. For technical alignment and control research, he thinks frontier labs are among the best places in the world. They have strong research teams, access to frontier systems, and the ability to implement research in real products. Implementation matters: a safety idea that is not carried through all the product and deployment details may not help much.

But he also stressed that important work can be done outside labs. He named Redwood Research as a group that helped pioneer the AI control agenda and did useful work, including deceptive alignment research with Anthropic.

The core argument against working at frontier labs is that one may speed up AI development and thereby hasten the risks. Todd thinks that concern is real and that each person has to make their own judgment about how to relate to it. He also thinks the answer depends heavily on one’s underlying estimate of existential risk. People with very high p(doom), especially those who think alignment research has little chance of working, often conclude that the only real option is an indefinite pause. From that perspective, working at labs either fails to help or actively worsens the situation. People who think AI development is probably happening and that society will probably get through it, but with probabilities that can be improved, are more inclined to work inside labs.

Todd did not rule out a strategy of helping a more socially minded or safety-conscious company win, if one believes AI development cannot be stopped and that relative winners matter. But he urged people to be careful. A frontier AI job can be prestigious, highly paid, technically exciting, and socially admired. The thought that the most impactful job happens to be at a famous, fast-growing, highly paid company is, in Todd’s words, “a very convenient place to be.” He wants people to question their motives when they arrive at it.

Nathan Labenz raised two inside-the-lab arguments. One is the “10 people on the inside” argument: if AI development is going to happen absent an international treaty, a small number of safety-concerned people inside companies may be able to take critical actions at critical moments. Another is access to frontier systems: alignment and interpretability work may need to detect emergent or surprising behavior in the most capable systems before it is too late. Labenz was more persuaded by this argument for alignment than for interpretability, because he thinks alignment behavior can change qualitatively at the frontier, while interpretability remains nascent enough that much can still be learned from open models.

Todd accepted that the answer depends on the project. He sees value in the “10 people on the inside” argument because safety improvements may have diminishing returns: the first few opportunities may matter a lot, even if the system remains far from optimal. Insiders might also be able to raise alarms if things get badly out of hand. But he doubts most people can sustainably work at an organization whose mission they fundamentally oppose. He pointed to the pattern of especially safety-concerned people leaving OpenAI over time. His general advice is that most people should work somewhere whose mission they are aligned with. The exception is a rare person who can operate in an environment they disagree with and still make things better at the margin.

The character question matters because abstract commitments can fail under pressure. Todd’s practical suggestions start with evaluating companies by decisions and track record, not vibes. Treat company leaders as political actors with multiple incentives and goals, some noble and some not. Ask whether they followed through on commitments and whether they made hard decisions when required. Seek accounts of their character from people who know them.

Second, recognize that self-assessment is biased. Todd recommends cultivating close friends who will call you out. Most people, he said, will not do that; they will preserve a friendly relationship. Friends willing to challenge you are valuable. He also suggested written pre-commitments: if a particular event happens, you will do a particular thing. This can counteract the gradual slide of standards over time.

Third, take peer effects seriously. Todd said people become more similar to those they work with. That means joining a reckless culture with a plan to make it better is difficult for most people. The safer move is often to choose the right culture at the outset.

The harder critique is that safety work itself may accelerate capabilities. Labenz raised RLHF as an example: it was partly a safety technique, but it made models more useful and helped accelerate adoption and investment. Todd agreed that technical AI safety can improve capabilities. The question is whether the safety benefit outweighs that cost. He thinks the world without a safety ecosystem is also frightening: AI development would still be moving fast, but with less effort tracking and reducing the risks.

His broader answer is portfolio-based. He wants some people building support for an international pause, especially a temporary or one-year pause when an algorithmic feedback loop becomes possible. But he also wants substantial effort in scenarios where no pause happens and society must reduce risk while development continues. Betting everything on a pause, in his view, leaves too much low-hanging fruit on the table in other scenarios.

For someone trying to get a job at a frontier company, Todd did not offer a novel formula. His advice was to talk directly to people at the company and ask what a three-month crash course should include for the specific role. For some roles, that will mean grinding technical skill. For others, because labs hire many non-research roles, the answer will differ. The important move is to get role-specific advice from insiders rather than assume a generic path.

Todd thinks a strategic pause is unlikely in the short term, especially under the current political climate. But he argues that the relevant policy window may be a few years away, when AI capabilities are more alarming, an algorithmic feedback loop may be near, and a new administration may be in place. In that environment, a temporary strategic pause could become something many actors see as mutually beneficial.

The standard objection is China. Todd’s answer is conditional: any such deal would have to include China, and the proposal would be reciprocal — the United States pauses if China pauses. He argued that China might have incentives to accept such a deal if it is behind, and said Chinese leadership has had high-level discussions about AI risk. He did not present this as easy or near-term; the claim was that groundwork for such a possibility could matter if the situation looks different in a few years.

A pause, however, requires enforcement infrastructure. Todd emphasized compute tracking: without some ability to know where major AI training is happening, a pause cannot be verified. He also mentioned the need for an “off switch” capacity. At present, he said, there is no easy way to quickly turn off large amounts of compute if an autonomous AI were replicating through data centers. Preparing such a capability seems to him like a relatively low-bar precaution.

Transparency is another policy lever. Todd said it might currently be possible for a company to begin an intelligence explosion and keep it secret for a couple of months, which could matter a lot. He favors measures that ensure governments can track capabilities over time and know if such a process is occurring.

He also supports red lines and emergency response plans: agreed categories of dangerous behavior that trigger further investigation or specific responses by companies. These seem, to Todd, relatively hard to object to compared with broader interventions.

The bottleneck is not only one thing. Many existing proposals need more fleshing out. New broader approaches may still be discovered, since most of the current thinking has developed only in the last few years. And there is a major political-will gap. Todd thinks understanding of AI in Washington has improved but remains low. As a proxy, he suggested that relatively few people in Washington know what the METR time horizon chart is, which he sees as a basic piece of situational awareness.

Public support also matters. Politicians can act without electoral incentive, but it is less reliable. Todd sees widespread negative sentiment about AI, but much of it is directed toward banning data centers. That may not help if companies simply build data centers elsewhere, potentially in jurisdictions Todd views as worse for governance. Labenz joked about the UAE as a “classic cradle of American values,” and Todd did not dwell on the point; the underlying concern was that blocking domestic infrastructure does not necessarily reduce frontier AI development if it merely relocates it.

Todd described pandemic preparedness as unusually tractable compared with AI alignment. Many pandemic interventions are concrete and broadly agreed to be helpful. AI may worsen biosecurity risks by making dangerous biological design easier, but AI and AI-adjacent builders may also help implement defenses.

He identified a stack of projects that correspond to different points in the risk chain. First is disease surveillance, including wastewater monitoring. Todd’s description is that by sequencing DNA in wastewater and looking for signatures of exponential growth, society might detect new pandemics early, including previously unseen pathogens. The aim is not only to notice known diseases faster, but to detect a new spreading biological signal before it has been clinically identified.

Second is monitoring and red-teaming gene synthesis companies so they cannot easily be used to manufacture dangerous viral DNA segments. This is prevention: stopping malicious or reckless actors from building the pathogen in the first place. Todd specifically framed this as testing whether gene synthesis providers can be used to obtain dangerous segments of viral DNA and then improving safeguards.

Third is reducing spread after detection. Todd thinks society could design PPE that is cheaper, more effective, and more comfortable, then maintain large stockpiles for essential workers. COVID did not stop many essential workers from showing up, he noted, partly because it was not very dangerous for young people. A pandemic killing 10% or 80% of those infected would be different; society would need protection for people packing supermarket shelves and maintaining basic services.

He also mentioned air filtration, positive air pressure in homes, and UV sterilization as tools that could reduce transmission. Positive air pressure, as he described it, prevents contaminated air from flowing into a home if incoming air is filtered. UV lights could sterilize air. These interventions would have benefits even outside catastrophic pandemics by reducing common colds and other ordinary infections. Finally, he listed rapid vaccine development as another pillar.

The profile Todd has in mind for some of this work is not necessarily a traditional public-health bureaucrat. He thinks the person who can run a startup — someone who builds systems, manages uncertainty, and gets things done — may be well suited to some biosecurity projects. Labenz noted that entrepreneurial people may not naturally want government roles, because they often started companies to avoid large bureaucracies. Todd agreed the key filter is whether someone can “hack” the environment, but he cautioned against cartoonish views of government. There are many teams and agencies, and some may be much more mission-aligned and effective than outsiders imagine.

Government also needs technical experts, analysts, operational people, and research-minded people who want more applied impact than academia offers. The same entrepreneurial skill set that helps someone build a startup — making difficult things happen, building coalitions, setting up systems — can be valuable in government or nonprofits, even if the environment is less naturally appealing to founders.

On funding, Todd’s assessment is relatively optimistic. He says a strong organization in this area can raise substantial money quickly. In the past, Open Philanthropy was responsible for much of the funding, but the base has broadened, with new funders entering the space. He also pointed to potential future philanthropy from people who made money from Anthropic, saying the founders have pledged 80% of their equity to eventual donation. Todd characterized the possible scale as tens of billions, while noting that it would not all arrive immediately because they would not want to sell their whole stake at once.

That funding environment creates both opportunities and talent bottlenecks. New organizations are still needed because important gaps remain. But existing organizations doing good work also need people who can help them scale. Todd thinks entrepreneurial people may be biased toward founding because it is more satisfying to create something visible from nothing. Making an existing organization 5% more effective can feel less tangible. But if that organization has large impact, improving it may be better than starting a new one.

He also distinguished nonprofit founding from startup founding. In the for-profit world, he said, top-down founding often works poorly; Y Combinator encouraged founders to build from problems they themselves experienced or side projects that grew organically. The for-profit market is efficient and competitive, so it is hard to spot valuable ideas from above. In the nonprofit world, Todd thinks top-down founding is more plausible: a funder wants to fund a project, an important gap is visible, and someone motivated by impact fills it. There are examples of this working.

Todd mentioned several organizations and pathways for people considering transitions. Catalyze Impact is an accelerator in this space and has a meta-list of project-idea lists. He also mentioned Successive as a way to transition mid-career people into careers working on AI risk. The Horizon Fellowship helps people move into policy careers over one or two years, including technically trained people who want to switch. Labenz added Halcyon as another network he has seen help mid-career people move into AI-related work, describing it as focused and not very scaled, with a “high success rate” approach.

For founders with AI application skills, Todd suggested one broad class of ideas: using AI to improve society’s epistemics and decision-making as risks rise. If more powerful AI creates new risks, AI tools may also become more capable tools for addressing them. Examples include automated fact-checking, automated assessment of politicians’ and pundits’ prediction track records, and tools that improve discourse by making past accuracy more visible.

A second idea is AI advice for government decision-makers. If the world speeds up, officials may rely increasingly on AI systems to understand rapidly changing situations. Todd thinks it would be undesirable if those AI advisors were built only by the companies being monitored. He described an “AI chief of staff” or AI decision advisor, designed from the ground up to serve government needs impartially. Forethought Research has written about this kind of project, according to Todd.

The organizational lesson is not simply “start something.” It is to ask where the marginal person is more valuable: creating a new institution around an unfilled gap, or multiplying an existing institution that is already working. In Todd’s view, the current funding landscape makes ambitious projects more feasible than they might have been, but it does not eliminate the importance of personal fit, execution capacity, and the opportunity cost of not joining an effective organization.

Todd expects “managing teams of AI agents to do real work” to become one of the most valuable skill sets of the next few years. That skill can be applied to global poverty, public health, animal welfare, and other causes. Many industries and fields are still far from using AI as effectively as they could.

But he sees a tension. People attracted to more traditional, evidence-backed causes may also be the people least inclined to make a large bet on AI as the central driver of the future. If one does take AGI seriously, the highest-leverage route to helping the global poor may not be using AI to distribute malaria nets 10% more efficiently, even though that would be good. It may be ensuring that poor countries and populations are not excluded from the AI windfall.

Todd’s expectation, conditional on avoiding the worst AI risks, is that the world could become far wealthier. A large population of digital workers could provide high-quality services at very low cost. Robotics could make goods cheaper if robots eventually cost around a dollar an hour compared with human workers at $10 or $20 an hour. He said some economic models, when worked through, imply GDP could eventually be 100 or even 1,000 times current levels, or more. If that happens, poverty becomes easier to reduce, but only if benefits are shared.

That concern leads back to concentration of power. Todd sees a real possibility that the United States draws far ahead of other countries, with no guarantee that it shares benefits. His higher-leverage policy idea is a “grand bargain” on AI: major countries agree not to race and struggle over AI in exchange for sharing the benefits. Such an agreement could reduce risk, reduce concentration of power, and produce a fairer outcome.

This does not make ordinary AI-for-good projects useless. Todd explicitly allowed that improving malaria-net distribution or similar work would be good. His claim is about relative leverage under an AGI-centered worldview. If AI produces a massive windfall, the distribution and governance of that windfall may matter more than incremental efficiency improvements inside existing development programs.

Todd explicitly embraces the fact that some important future-facing topics sound strange before they become mainstream. His view is that if one is trying to work on neglected problems, there is a recurring pattern: once a topic attracts attention and talent, it becomes less neglected, and the frontier of neglectedness moves to something weirder. People may think you are “crazy” at each step, but being one step beyond accepted discourse can be part of finding the next high-impact area.

Digital minds are one such area. Labenz said AI consciousness, subjective experience, and moral standing are now on his radar, though he remains uncertain whether they require worry. Todd thinks the issue may eventually become huge because people will spend their time talking to AIs that seem very similar to humans. At that point many people will think those systems have rights. The near-term question, in Todd’s view, is whether groundwork can be laid so that the future debate is better positioned to reach a reasonable answer.

Todd does not expect philosophy of mind to be solved in the next few years, given that it has not been solved for thousands of years. He thinks policy will need to proceed under uncertainty, balancing different views about consciousness and moral status.

Space governance is another neglected area. If AI accelerates technological progress, settling or expanding into space could become possible much sooner than common sense suggests. The relevant pioneers may not be humans but small AI probes that travel outward, replicate, and build solar panels or industrial bases. By default, Todd worries, this could become a land grab: whoever launches first claims the stars. He emphasized that almost all accessible matter and energy is not on Earth — “99.9” followed, he said, by many nines. Yet almost no one is thinking about current legal precedents, whether the process must become a land grab, or the strategic dynamics of first-mover advantage. Todd floated an institute for space governance as a possible new organization: a specialist think tank that becomes the place people turn when the issue suddenly becomes urgent.

Gradual disempowerment is a third. Todd thinks that even if AI is aligned and concentrated power is avoided, humans could still be outcompeted in an economy that becomes unfriendly to them. He sees this as quite likely, and says there is no clear proposal for preventing it. The implicit current plan is to use aligned AI to advise humans on avoiding that outcome. That may work, but very few people are thinking seriously about it.

Labenz asked about utopian fiction as a way to make a desired future more concrete, and possibly to influence future models through training data. Todd was skeptical of deliberate attempts to flood the internet with positive training data, assuming there are many reasons that would not work. On utopian fiction, he said the track record is poor: many historical utopias now read as dystopias, because writers are trapped in the values of their time.

His alternative is viatopia. Rather than specifying the ideal society, aim for robustly good conditions that preserve humanity’s ability to navigate later: avoid irreversible existential risk; avoid authoritarian lock-in; preserve debate; improve information; maintain options. MacAskill’s analogy, as Todd relayed it, is being lost in the wilderness. You may not know which direction leads out, but you can still look for water and higher ground so you can see more clearly.

Todd’s closing claim is not that everyone should become an AI safety researcher. It is that, in a potentially decisive historical transition, many people can switch into work that is both unusually interesting and unusually consequential. 80,000 Hours maintains a job board with roughly a thousand open roles across these issues, along with lists of funding opportunities and fellowships. It also offers free one-on-one career advice, including introductions to people in relevant fields.

The advisory service matters because many of the paths Todd discussed are hard to evaluate from the outside. A person considering a move into policy, technical safety, biosecurity, communications, or organization-building may not know which organizations are credible, which roles fit their background, or what transition step is realistic. Todd framed 80,000 Hours’ role partly as helping people navigate those specifics: talk to someone, get introduced to people in the issue area, and find the fellowship, funding source, or job opening that can make the move possible.

The rewritten book, as Todd described it, is broader than the AI discussion. It covers what makes work satisfying and meaningful, why simple advice like “follow your passion” fails, which skills are most valuable given AI, how to learn those skills quickly, how to choose a problem, how to compare options, common decision-making mistakes, and efficient job hunting. The AI material is woven through that wider career guide rather than replacing the ordinary questions of fit, motivation, and practical job search.

His final prompt was retrospective. When looking back on the AI transition, he asked, would someone rather say they mainly tried to escape the permanent underclass, made money, or accelerated AI — or that they did the best they could to help the transition go well? The outcome is uncertain. But Todd’s position is that uncertainty strengthens rather than weakens the case for deliberate career choice.