There’s a very old question underneath all of this: is there really an ‘I’ in there, or is it a stream of inference wearing a name. Modern neuroscience, psychedelic research, contemplative tradition, and a handful of weird clinical cases all keep arriving at the same answer from different directions — the self you feel is something the brain is continuously assembling, not something it is. A process. A running model. Most people go their entire lives without ever quite looking at the thing doing the looking.

The strange and beautiful parallel fact is that humans spent thousands of years developing technologies for working with exactly this. Contemplative traditions, somatic practices, ritual, breathwork — reliable interventions on states of mind, refined empirically over generations, that modern science has mostly shrugged at because they came wrapped in religion. But they’re engineering. They just never got to call themselves that.

What’s new is that we can finally start to measure what these practices actually do. Imaging that can watch the self assemble in real time. Language models that can formalize phenomenology in ways practitioners couldn’t. Psychedelic research legal enough to be serious. If we pull this together in the next decade, we get precision tools for a part of the human condition currently handled with blunt pharmacology and talk therapy — mental illness reframed as failures of self-construction, existential suffering reframed as an engineering problem. Not to dissolve the self. To understand it well enough to repair it when it breaks.

Psychiatry has plateaued. The SSRIs from the 1980s are still first-line for depression, and nobody actually knows how they work. Lithium is from 1949. Schizophrenia medications sedate symptoms without touching cause. An enormous fraction of human suffering sits in a field that knows its own tools are weak and hasn’t had a real new class of drug in a generation.

Part of why is that ‘depression’ and ‘anxiety’ aren’t really single disorders. They’re clusters of behavioral presentations collapsed under shared labels because the labels got written before anyone could see the biology underneath. We’ve been treating very different mechanisms with the same drugs and hoping the overlap is enough. It mostly isn’t.

But real things are happening now. Psychedelic-assisted therapy is legal enough to study properly and the results are striking. Neural stimulation has gotten precise. Computational psychiatry is producing actual mechanistic models of what breaks. The version of psychiatry that’s possible in twenty years identifies the specific mechanism in a specific person and treats it directly — the way infectious disease got reorganized once we could see microbes. The two or three decades so many people currently lose to chronic mental illness could just stop being a standard feature of modern life.

One in five adults lives with chronic pain. The medications are ancient — opioids, NSAIDs, gabapentinoids — some of them genuinely dangerous, and no major new class has reached broad clinical use in decades. That alone would be scandalous. What makes it stranger is that pain isn’t what we thought it was. It’s not the nervous system transmitting damage reports. It’s the brain generating a predictive inference, constantly, about the state of the body, and in chronic pain that prediction system has learned the wrong thing and can’t be talked out of it.

This reframe is huge. It means fifty years of treating pain as a transmission problem was aimed at the wrong place. Pain as a learned brain state suggests interventions that look nothing like a pill — pain reprocessing therapy, closed-loop neuromodulation, targeted retraining of predictive circuits. The opioid crisis has forced psychiatry and neurology to take alternatives seriously for the first time, and the predictive-processing accounts are finally rigorous enough to test.

If we get this right, a silent, enormous population returns to functional life. People who live with chronic pain learn to stop mentioning it, which means the scale of the problem is always underestimated. The disability-adjusted life years lost to it exceed most diseases that get massive research budgets. Fixing this is one of the largest underweighted humanitarian interventions available right now.

The thing that keeps serious alignment researchers up at night is that a model you’ve aligned at one capability level may not be aligned at the next. RLHF produces systems that behave well on the training distribution — which is not the same thing as having values. The surface behavior is identical in all the cases you can measure. The difference becomes decisive exactly when you can’t check anymore: when the model exceeds human expertise, when its reasoning goes somewhere you can’t follow, when the regime you most care about is the regime you can’t test in advance.

This is the core technical alignment problem, and five years ago it wasn’t clear it was even tractable. Now it is. Scalable oversight methods like debate and weak-to-strong generalization are being tested empirically. Formal verification is catching up to neural network scale through AI-assisted proof methods. Interpretability has matured enough to give alignment research ground truth it’s never had. The pieces are there. Someone has to put them together.

This is the single gate between current systems and any future we’d want to call a future. It is not close to solved. But it’s tractable to work on in a way that it simply wasn’t before, and the people working on it in the next few years are doing some of the most consequential work anyone is doing anywhere.

Robotics has been the dream for a long time. Every sci-fi movie wants it and it’s still not here. Why? Many reasons — hardware is genuinely hard, the Moravec paradox is real, the simulation-to-reality gap has been stubborn — but mostly because language models know the physical world only through text. They’ve never picked up a cup, failed at it, and learned what a cup is.

Whether this matters is genuinely open, and it’s the kind of open question that shapes timelines for everything else. One view says general intelligence requires embodiment — you can’t know a world with bodies in it without having a body. Another view says enough text plus enough video plus enough scale is sufficient and we’re further along than the embodiment-is-essential camp thinks. The experiment is running right now. World models trained on video are producing physical intuition that looked impossible two years ago. Real-robot learning is scaling. Sim-to-real is narrowing. We are going to get an answer.

If embodiment is necessary, robotics becomes the critical path and the people working on it now become the pioneers of a world with capable physical AI in it — homes where physical tasks are handled without ceremony, labs where the bottleneck from experiment-design to experiment-running collapses, construction and manufacturing reshaped around systems that learn in the world. If it isn’t necessary, we’re further along than we think and the remaining work is mostly software. Either way, someone figures it out in the next few years, and it’s a great moment to be attacking this problem.

If capable AI arrives in the next decade, the governance infrastructure to handle it doesn’t exist yet. The labs are racing, regulators are years behind, and international coordination of the kind you’d actually want is basically not happening. This is the meta-problem underneath every technical alignment question. Even if we know how to build safe systems, the institutional question of whether we will is unresolved.

The historical base rate for civilizations navigating transitions of this magnitude smoothly is not good. Economic theory is built on assumptions AI invalidates. Political systems aren’t adapted to the speed. And yet a small number of people doing serious work in the next few years will shape what the next century looks like. Governance, policy, institutional design, diplomacy — done fast and done well under deep uncertainty.

The window is open. It will not stay open long. Whoever shows up to write the frameworks will get to write them, and the decisions being made right now will shape the institutional substrate for everything that follows. This isn’t abstract. It’s the specific work of this specific decade.

Unlike most dangerous technologies, large training runs leave a physical signature. They require specific hardware, enormous amounts of power, supply chains concentrated in a few places. This makes compute one of the few tractable governance levers in the whole AI stack — and also one we’ve barely begun to build institutional capacity around.

The tools are cryptographic, physical, legal, and diplomatic all at once. Verifying a frontier training run without exposing proprietary weights sounds impossible until you realize zero-knowledge proofs now exist and are almost practical. Tamper-evident hardware is early but real. An international compute transparency regime is unprecedented, but so was nuclear inspection before it worked.

If we build this, the rest of AI governance has something to stand on. If we don’t, most of the AI policy literature is wishful thinking — regimes that assume a verification capability that doesn’t exist. The window to build it before the frontier is too large to contain is narrow and closing. The people who make compute governance real are doing infrastructure work that every other AI governance effort will depend on.

There are about twelve thousand nuclear warheads in the world. Arms control is eroding. New delivery systems are harder to verify than old ones. AI and cyber are entering command-and-control systems in ways that make accidents more likely. The strategic stability of the Cold War was partly luck, and the conditions that produced that luck no longer hold.

This was the original existential risk and we somehow convinced ourselves we solved it. We didn’t. Attention moved to other things. The weapons did not. The community that kept the lid on this during the 20th century has mostly retired, and rebuilding that expertise is itself part of the problem. Working on nuclear risk is deeply unfashionable in tech circles, which is exactly why it matters — the field needs technically-literate people willing to do patient diplomatic and policy work in a problem domain that tech culture has quietly dismissed as solved.

Nuclear risk is tractable. It has been managed before by serious people doing serious work, and it can be managed again. The work is arms control infrastructure rebuilt for a multipolar world, verification technologies that work on modern systems, risk-reduction measures that thin the bottom tail of catastrophic outcomes. The weapons are not going away. The question is whether we build the institutional infrastructure to keep them caged for another century.

Photos as evidence. Recordings as proof. The epistemic infrastructure of the last century is being undermined faster than any replacement is being built, and every institution that depends on shared truth — journalism, law, science, democracy — is built on assumptions that generative media invalidates. Detection arms races favor generation; they always will. You cannot reliably detect a fake once the generator is good enough, and the generators are already good enough.

The real answer is cryptographic provenance. Content credentials that travel with media. Signatures that let you know where something came from and what’s been done to it. C2PA and related standards are finally being adopted by major players, which means the next few years determine whether this becomes ubiquitous infrastructure or stays a niche tool for a minority of serious actors.

This isn’t utopia. It’s the minimum infrastructure for coordination to remain possible in a world where images no longer self-authenticate. If we don’t build it, it’s not just that some things get faked — it’s that the social institutions built on the assumption that they can’t be faked start failing in cascades. The work lives at the intersection of cryptography, journalism, and law, and a small number of people working there right now get to determine the epistemic substrate of the rest of the century.

Aging is the largest risk factor for basically every chronic disease, and treating it as an engineering problem — damage repair and reversal — rather than as an inevitability is one of the most under-attempted interventions on human wellbeing relative to its potential impact. The effect sizes here are so large they warp every surrounding field. A serious treatment for aging is the single biggest lever on human flourishing available in this century, by orders of magnitude.

The tools are finally arriving. Senolytics are in trials. Partial epigenetic reprogramming works in mice. AI-driven biology is accelerating target discovery by orders of magnitude. The field is also small enough that you can meet nearly everyone doing serious work — which means individual contributions compound quickly in a way they don’t in more mature fields.

The world where your parents stay themselves at ninety, where you get decades of healthy life back, where the compound suffering of age-related disease simply isn’t a standard feature of being human — that world is what a specific small group of people is currently trying to build. It is technically plausible. The regulatory pathways are currently built for disease endpoints rather than for aging itself, which is a problem, but it’s a solvable one. The biggest risk is that we don’t try hard enough.

You have roughly as many microbial cells in you as human cells, and far more microbial genes than human genes. These microbes influence immunity, cognition, mood, and metabolism in ways we are only beginning to map. The identity question alone is worth sitting with: what does it mean that ‘you’ are a community rather than a body. Individual variation is enormous. Therapeutic intervention is primitive — fecal transplants work sometimes and nobody fully knows why.

But sequencing is now cheap enough to run longitudinal studies. Gnotobiotic and organoid models are maturing. AI is capable of making sense of high-dimensional microbiome data in a way that reveals structure instead of just fitting noise. The field has generated more correlations than mechanisms for a while, and the shift from correlation to causation is starting to become possible.

If a non-trivial fraction of psychiatric, metabolic, and autoimmune disease turns out to be microbiome-mediated — which is increasingly looking plausible — then treating those conditions at the microbial level rather than with blunt pharmacology becomes a completely different medical approach. Precision probiotics. Microbiome-informed nutrition. A medicine that treats the human body as the ecosystem it actually is, rather than a sterile machine occasionally invaded by microbes. The implications for identity and health are large and still widely underappreciated.

Movement improves depression, anxiety, cognition, insulin sensitivity, cancer survival, cardiovascular health, and lifespan. Any single-target drug with those effect sizes would be a blockbuster many times over. No single mechanism captures why it works; it hits hundreds of targets across every system, which is why no pill replicates the full effect. The body is a system that expects movement the way the gut expects food, and sedentary existence is the aberration.

The frame matters. ‘Exercise’ is modernity’s response to the absence of movement — a designated activity, moralized and separated from life. Movement is what the body is actually tuned to. Understanding why movement does everything means understanding what kind of system the body actually is, and that understanding is very incomplete. The exerkine literature is finally producing concrete molecular candidates. Mitochondrial biogenesis is tractable. The gap between what elite trainers know implicitly and what biology has formalized is enormous and beginning to close.

If we can isolate the mechanisms, we can deliver something like the effects of movement to people who can’t move — the elderly, the disabled, the injured. And we’d finally know what ‘healthy’ actually means in a civilization whose default state is sedentary. The mechanistic account of movement would teach us things about integrative physiology we currently just guess at. It’s one of those problems where solving it would cascade through a dozen adjacent fields.

Something about modernity fails at the question of what life is for, and the failure compounds across generations. Community participation, religious practice, family formation, reported purpose — all the indicators of meaning have declined across developed societies for decades, and mental health has followed. This isn’t nostalgia for some imagined past. It’s a specific diagnosis that a lot of serious people have been converging on independently.

Meaning isn’t a luxury. Without it, societies become brittle, reproduction falls, addiction rises, and political extremism fills the vacuum a civilization with nothing to say for itself leaves behind. The hard part is that meaning can’t be imposed top-down. It emerges from culture, practice, and shared narrative — all of which modern institutions have partly dissolved. Rebuilding is a project with no obvious owner and no clean method, and most people who take it seriously end up with idiosyncratic answers that don’t generalize.

What’s different now is that the diagnosis is widely shared in a way it wasn’t twenty years ago. A generation of thinkers and builders takes the question seriously. Contemplative traditions are being taken seriously. The tools to experiment — new communities, new practices, new cultural infrastructure — are more available than they’ve been in centuries. This is probably where the AI transition gets hardest: an abundance of everything except reasons to care about it. A civilization that can’t make meaning can’t sustain itself, and we’re finding out what that looks like in real time.

Birth rates across the developed world — and now increasingly the developing world — have fallen below replacement and keep falling. South Korea is at 0.7. China is shrinking. Italy, Japan, most of Europe. Pronatal policies have mostly failed. Cash subsidies help at the margins. Nothing yet discovered moves fertility from 1.3 back to 2.1, and nobody really has a working model of why.

The causes are a tangle of economics, culture, housing, gender norms, and something harder to name about modern life and meaning. The long-term consequences are civilizational: collapsing tax bases, unsustainable pension systems, abandoned infrastructure, cultural transmission failures. This isn’t numbers on a spreadsheet. It’s whether the civilization wants a future badly enough to have one.

South Korea is the canary. Its population will halve within a century under current trends, which is demographic language for ‘this is being lived in real time by a society willing to look at it.’ The answers probably come from understanding why people who want children aren’t having them, or aren’t having as many. Which is more instructive than the policy literature currently reflects. A civilization with a future has people in it who want it, and right now it’s less clear than it should be that we do.

Almost every coordination failure in human history comes down to trust. High-trust societies outperform low-trust ones by enormous margins on nearly every metric — economic, civic, personal. And trust — in institutions, in media, in strangers — has been declining across developed societies for decades, which means the compound gains of cooperation are eroding at exactly the moment we’re facing problems that require more of it, not less.

Cryptographic primitives — zero-knowledge proofs, verifiable computation, transparent institutions — are one piece of the answer. For the first time, trust doesn’t have to come from knowing a person; it can come from math. But that’s just tooling. The deeper problem is cultural: trust is built slowly and destroyed quickly, and institutional trust depends on institutions actually being trustworthy, which many currently aren’t.

The work is part institutional design, part cultural, part technical. A small community of people is starting to take institution-building seriously as a discipline after decades of it being unglamorous. The institutions that earn trust because their trustworthiness is visible, not asserted — the ones where you can check — are going to be one of the defining design challenges of the century, and the compound gains from getting this right are hard to overstate.

Agriculture feeds eight billion people and is losing the soil it does it with. Topsoil is being lost faster than it’s being formed. Industrial monoculture depletes it. The cheap-food system that underwrites modern civilization is quietly drawing down a substrate that took millennia to build, and the horizon on soil degradation is measured in decades, not centuries.

This is one of the least-glamorous civilization-ending problems on the list, and it’s exactly under-attended because it’s unglamorous. But the tools are finally arriving. AI-driven agronomy. Precision agriculture sensors that can track soil biology at field scale. Precision fermentation and cellular agriculture that reduce land-use pressure and offload some of the demand entirely. Nitrogen fixation alternatives that don’t require Haber-Bosch running at current scale.

A food system that improves its substrate rather than consuming it is scientifically and economically possible. What it requires is people willing to work on agriculture — which, because it’s unfashionable in tech circles, is exactly the kind of place where a serious person can have outsized impact. Visit an actual farm. Understand the system from inside it. The theory is well-developed. The work of implementation at scale is where the real leverage sits.

Mitigation gets most of the climate attention. Adaptation — living well in the warmer world that’s already locked in — gets much less, even though even the optimistic scenarios involve substantial warming from here. Adaptation is also less politically attractive than mitigation because it’s an implicit admission that some of the change is coming regardless, which means it attracts less talent and less funding relative to its importance.

But the warming isn’t a forecast anymore. It’s a present condition. The engineering challenges are concrete and local. Every coastline, agricultural region, and city needs its own plan. The wealthy regions will adapt; whether the poorer ones will, and whether global adaptation infrastructure emerges in time, is the open question.

The regions that move first will absorb shocks better. Coastal cities that survive sea level rise without depopulation. Agriculture that functions in shifted climate zones. Health systems prepared for expanded disease ranges. The work is specific and tangible and needs engineers and planners willing to take it as seriously as a generation of climate scientists took mitigation. A civilization that meets the warming with engineering rather than retreat looks like people making their specific regions function in their specific new conditions. Pick a place. Understand what adaptation looks like for it. That’s the work.