My view of aging in 2026

For a working definition, aging can be thought of as the progressive loss of physiological function and resilience that makes an organism increasingly vulnerable to disease and death. Here, resilience means the ability to recover from stress or injury. Benjamin Gompertz described an approximately exponential increase in human mortality with age in 1825, using mortality tables, long before anyone knew what a gene was. Across much of adult life, the mortality rate roughly doubles every eight years, although the rate varies across populations and ages. To me, that regularity hints that something lawful is happening underneath all the biological mess.

That said, aging can feel as unavoidable as gravity. Scientifically, the analogy has limits. In long-term laboratory studies, Hydra maintained a very low, roughly constant mortality rate, and their reproductive output did not systematically decline with age. A comparison of 46 species found mortality trajectories ranging from steeply increasing to flat to declining. That does not make these organisms incapable of dying. It does challenge the idea that every organism must deteriorate in the same way. This is part of what makes aging interesting to me: how do living systems maintain themselves, and why does that maintenance so often become less effective with time?

What "aging" are we solving?

Two terms usually need clarifying in any discussion of aging research: lifespan and healthspan. Lifespan is how long an organism lives. Healthspan is how long it stays healthy and functional. I don't see them as competing goals. They can improve together, although improvement in one does not guarantee improvement in the other. Lifespan studies can also measure health and function rather than survival alone. My interest is in improving both. Distinguishing the two helps us judge what an intervention actually achieved; it does not force a choice between them.

A similar question is whether to extend average lifespan or maximum lifespan. Again, I don't think I need to choose: both interest me, and they can increase together. Mean lifespan is the average age at death; median lifespan is the age by which half the population has died. Maximum lifespan concerns the longest-lived members, but it matters how we measure it. The oldest observed age depends partly on the number of individuals studied and does not, by itself, establish a biological limit. Studies may instead assess late-life survival using a measure such as the 90th-percentile lifespan—the age by which 90 percent of the population has died. Exceptionally long-lived individuals deserve closer study, but I also want to know whether an intervention reliably extends survival across the whole group. For me, these are complementary questions.

There are also several related aims hiding inside "anti-aging": prolong the young state, delay the onset of the old one, or rejuvenate someone who is already old. These aims overlap, and preventing damage, clearing damage, or replacing what is lost may contribute to more than one of them. They can nevertheless require different trial designs and raise different ethical questions.

I like to start from theory. Theories of aging come in layers. The first is the evolutionary layer, which asks why we age.

Metabolism and nutrient sensing

Calorie restriction (CR). In my opinion, still one of the strongest starting points for understanding how a non-genetic intervention can extend animal lifespan. Of course the effect varies with genetic background and with the timing and degree of restriction: in 960 genetically diverse female mice, genetics explained more of the variance in lifespan than diet did, and in a separate study the same 30 percent cut extended male mouse lifespan by about 10 percent when food was spread across the day, or 35 percent when feeding included a daily fast and aligned with the animals' active period. Neither result means that greater restriction is always better, or that the same schedule would produce the same effect in people.

Rapamycin / mTOR. A drug with a particularly strong mouse-lifespan record: it extends lifespan even when started late. Work separating the two mTOR complexes found that disruption of mTORC2 contributes to rapamycin-induced insulin resistance, while reduced mTORC1 signaling can support longevity without that metabolic cost. Other adverse effects and mechanisms of lifespan extension remain less clearly separated. How to translate those benefits into aging treatment in people is the question Lamming explored in his 2026 review, “Bench to bedside: is rapamycin headed for the docTOR?”

Sirtuins and NAD⁺. One of the most hyped and contested corners of the field. Where a gene is expressed can matter as much as how much: one brain-specific SIRT1-overexpressing mouse line lived longer, with experiments implicating particular hypothalamic nuclei, whereas the whole-body overexpression models discussed in that study did not show lifespan extension. That is more specific than “more SIRT1 is better.” NAD⁺ precursors can raise NAD⁺ in human blood cells; functional outcomes need to be judged separately. For example, NMN improved muscle insulin sensitivity in a small trial of women with prediabetes, while other trials have reported no improvement in glucose control or insulin sensitivity. These are specific outcomes in specific populations, not evidence of human lifespan extension.

Insulin/IGF-1 signaling. One of the most conserved longevity-related pathways, from worms to mice with reduced growth-hormone/IGF-1 signaling. In humans the picture is less straightforward: a meta-analysis found a U-shaped association between circulating IGF-1 and mortality. That is one reason I wonder whether optimal metabolic conditions differ among people (Part 5). The association alone cannot tell us whether changing a person's IGF-1 would improve survival, or what their individual optimum would be.

Senescent cells and the surrounding tissue

Senescent cells enter a generally stable state of cell-cycle arrest; many resist cell death and secrete inflammatory and tissue-remodeling factors. Genetically clearing p16-positive senescent cells improved several health measures and extended median lifespan in mice. Senolytic candidates such as dasatinib plus quercetin, fisetin, and navitoclax aim to eliminate senescent cells; senomorphics aim to alter their harmful behavior without necessarily killing them. These approaches do not all have the same evidence behind them. A small, uncontrolled human pilot reported improvements in some physical-function measures, while a phase 2 trial of UBX0101 for knee osteoarthritis did not improve its primary pain outcome relative to placebo. Neither result settles the entire approach. Targeting matters because senescence also contributes to wound healing and can suppress tumors. Which cells should we remove, in which tissue, and at what stage?

Genetics

How much of how long you live is inherited? The answer depends on what is being estimated. Twin studies have put the heritability of human lifespan around 20 to 30 percent, and when Ruby and colleagues analyzed a pedigree dataset containing more than 400 million people and accounted for the tendency to marry people with similar traits, their estimate fell below 10 percent. But a 2026 analysis of twin cohorts estimated heritability above 50 percent for lifespan associated with intrinsic mortality, after modeling the contribution of external causes of death. These estimates concern different definitions and assumptions; they should not be treated as interchangeable measurements. Heritability is a population statistic about variation, not the fraction of an individual's years that genes determine.

Genome-wide searches have repeatedly implicated APOE and FOXO3, but additional loci have been associated with lifespan and related traits. Exceptional longevity, parental lifespan, and an individual's survival are not identical outcomes. Even APOE is allele-specific: ε4 is depleted among exceptionally long-lived people, and ε2 is associated with longevity beyond its association with Alzheimer's disease. That makes me interested in changing a particular allele or its effects, rather than simply turning the gene off. Whether such a change would safely extend life is a separate experimental question.

Stem cells and their niches

Stem-cell aging is not always a matter of running out of cells. Some mouse blood-forming stem-cell populations become more numerous while their function changes; in the aging fly gut, progenitor cells accumulate while normal tissue organization breaks down. By contrast, stem-cell depletion has been observed in the mouse hippocampus and the fly germline. The species, tissue, and way the cells are identified matter. The niche ages too: immune-cell infiltration can alter neural stem-cell behavior, and in rodent experiments, age-related tissue stiffness suppressed oligodendrocyte progenitor activity, while softer substrates restored aspects of their function. Rejuvenation may therefore require attention to both the cell and its surroundings. My open question is what the end state looks like: replacing lost or dysfunctional cells with better ones, repairing the niche so the cells we already have can function, or combining the two? Senolytic CAR-T cells belong to a related but distinct approach: engineered immune cells clear selected senescent cells rather than replace them with new tissue.

Parabiosis

Joining the circulations of a young and an old mouse can improve regeneration in old tissues. There is no single established “youth factor” that explains all these effects. GDF11 became a prominent candidate, then its reported effects on muscle regeneration were challenged. Other candidates, including TIMP2, have shown benefits for particular outcomes in mice. In a separate blood-exchange experiment, the harmful effects of old blood were more pronounced than the benefits of young blood for several measures, and replacing part of old mouse plasma with saline and albumin improved several tissue measures without adding young plasma. That makes “remove old” an interesting possibility alongside “add young.” It does not yet establish which factors explain each benefit, or whether either approach will improve human aging outcomes.

The comparative zoo

Evolution has already produced comparisons we could never create in a single experiment. Naked mole rats live more than 30 years at roughly mouse size and, in a large captive dataset, showed no detectable Gompertz-like increase in mortality over the ages studied. Their cancer resistance involves mechanisms including high-molecular-mass hyaluronan and unusually early contact inhibition. They are not cancer-proof: spontaneous cancers have been documented. Still, particular chemical carcinogen regimens produced no tumors in the tested naked mole rats while producing tumors in mice. The researchers linked this resistance to a dampened inflammatory response and the loss of functional necroptosis regulators, a possible mechanism rather than a complete explanation of their cancer resistance. They also maintain mineral balance despite extremely low measured vitamin D metabolites, reminding me how risky it is to assume that the same physiological measurements mean the same thing across species. Hydra and the jellyfish Turritopsis offer other challenges to familiar aging patterns: Hydra can show negligible senescence under favorable laboratory conditions, while the jellyfish can revert from the adult medusa to the polyp stage. These are unusual maintenance and life-cycle strategies, not proof that the animals cannot die.

Reprogramming

Cyclic expression of the Yamanaka factors improved several age-associated features and extended life in a premature-aging mouse model; related experiments improved tissue recovery in older wild-type mice. Partial reprogramming aims to recover useful function while retaining cell identity, but achieving that safely depends on the factors, tissue, dose, and duration. In June 2026, Life Biosciences announced that the first participant had been dosed in its phase 1 ER-100 trial for open-angle glaucoma and non-arteritic anterior ischemic optic neuropathy. The treatment uses three factors, OCT4, SOX2, and KLF4. An October 8 sponsor update described early observations in three participants. This is an early safety and tolerability study; those observations do not establish clinical efficacy, human rejuvenation, or a general explanation of aging.

The tools

Single-cell and single-nucleus sequencing have made it possible to characterize cell types and molecular states at a scale that was previously difficult to reach, and to ask how their composition and gene expression change as an organism ages. Many cell types were known long before sequencing; the advance is in resolving their diversity and comparing them systematically. We now have atlases and multi-omics readouts, as well as measurement tools such as DNA methylation clocks.

• Tabula Muris Senis, a single-cell atlas of aging across 23 mouse tissues
• The Human Cell Atlas
• Aging-specific databases
• Chromatin accessibility (ATAC-seq) and 3D genome structure (Hi-C)
• DNA methylation clocks, with later approaches extending from chronological-age prediction toward health-risk prediction and estimates of the pace of aging

Next step: multimodality, and perhaps the virtual cell. I am glad these maps exist. I also want to understand how we get from increasingly detailed descriptions to interventions we can test. Some mapping efforts have been running for a decade. What would make the next atlas more useful for deciding which experiment or treatment to try?

Missing negative results and excessive exaggeration

When negative results go unpublished, the literature over-represents interventions that worked. On top of that, a large percentage sounds like a breakthrough until you check what it is a percentage of. A worm living substantially longer after a genetic manipulation is exciting, but we still need to know the conditions, the strain and whether another lab can reproduce it. The same caution applies when a treated group is compared with a control that lived unusually briefly, when a change in median lifespan is presented as proof that aging itself has slowed, and when a pooled number hides a sex difference: acarbose and 17α-estradiol showed stronger lifespan benefits in male mice in the ITP. Median and late-life survival describe different parts of a survival curve; neither separates disease from aging by itself. A pooled figure may be technically correct and still hide something a reader needs to know.

Replication deserves the same attention as novelty. In a widely cited account, Amgen researchers reported confirming findings from only 6 of 53 selected landmark cancer studies. That was not a representative survey of biology, and it does not give us a failure rate for aging research. It is a reason to ask how independently a finding has been tested. When a result fails to confirm a hypothesis, the questions include dose, frequency, timing, target, sex, strain—and whether the hypothesis is wrong. These possibilities need experiments that distinguish them, rather than becoming permanent excuses for a failed prediction. When a result succeeds, the same scrutiny applies.

We pull one screw and watch what happens

The body can look like an exquisitely built machine, and one familiar move in biology is to pull out a screw and see what happens. Sometimes the system collapses; sometimes it compensates; sometimes very little happens at all. A well-controlled knockout can establish that a gene has a causal role under the conditions tested. What bothers me is the distance between that result and understanding how the whole system works. If knocking out A makes B decrease in one cell type but increase in another, I want to know why the result changes—and whether we can predict that change before running another experiment.

Yuri Lazebnik explored this problem in 2002 in an essay asking whether a biologist could fix a radio. His discussion of the paradox that accumulating facts can leave us feeling less able to understand a process has stayed with me. Removing a part reveals both something about its function and the system's response to its removal. Many genes have multiple roles, and their effects can depend on age and context; antagonistic pleiotropy is one reason a helpful early-life effect may carry a later cost. The machine analogy has limits, too: living systems develop, adapt and compensate. I want perturbation experiments to help us build models that explain those responses, as well as identify the parts involved.

I once heard the development of a field described as a sequence of phases: a period when new findings pour in and each one changes the picture, followed by a period when findings keep pouring in and the picture stops changing. Findings stop adding to the structure and start piling up on it. I often get that feeling reading aging research, and honestly biology more broadly. It is a concern about how we connect findings, rather than a claim that the experiments themselves have stopped being useful.

Lack of direct, parallel comparison of interventions

Twelve hallmarks and growing, no agreed weights. Dozens of interventions, tested in different strains, on different diets, at different ages, in different sexes, by labs with different definitions of “control.” Dietary restriction alone covers many protocols, and genetic background can change an intervention's effect on lifespan. Some inconsistency is therefore expected. The consequence is that it is difficult to compare the levers on common terms. I think this is one of the field's larger structural problems, because a useful comparison would help us decide where to start.

But what would we actually rank? The feature that best predicts decline may not be the cause that is most useful to change. And the strongest causal contributor may not be the safest or most practical therapeutic target. I would want comparisons of how much an intervention changes a specified outcome, in a specified population, at a specified cost in side effects. The interactions matter too: controlling one process may change the benefit of targeting another. A universal ranking of the hallmarks may be impossible, while a ranking of interventions for a particular purpose could still be useful.

The NIA's Interventions Testing Program shows what coordinated testing can achieve. Its cohorts, studied since 2004, test compounds at three sites in genetically heterogeneous mice, using shared protocols and diets, with analyses that account for site. It has provided influential evidence for rapamycin, acarbose and 17α-estradiol, alongside valuable negative findings: metformin alone and resveratrol did not significantly extend lifespan under the conditions tested. That narrows the claims we can make about those regimens; it does not rule out every dose, population or possible health benefit.

The ITP is an important foundation, not the whole of mouse lifespan research. The Caenorhabditis Intervention Testing Program similarly tests across laboratories and genetic backgrounds, and the 960-mouse dietary-restriction study compared five feeding regimens in one large design. What I would like to see is a broader extension of this approach: interventions from different classes, their combinations and their timing tested against shared controls and common outcomes. A shared protocol would improve comparability, but it would also need deliberate variation and independent replication to show which conclusions survive outside that protocol.

Waiting on translational data

For many proposed geroprotectors, the evidence for lifespan extension is still preclinical. Some of the difficulty is biological: an effect in a short-lived animal does not tell us how much benefit to expect in an already long-lived human. An Ecuadorian cohort with growth-hormone-receptor deficiency, a cause of Laron syndrome, had strikingly low rates of cancer and diabetes, without an observed increase in overall lifespan. But the cohort was small and many deaths had other causes, so it cannot establish that reduced signaling offers no human survival benefit. Nor does the role of IGF-1 reduce to “lower is better”: its effects need to be considered in relation to age, tissue, health and the outcome being measured.

There is also a clinical-development problem. “Slowing aging” needs to become a specific, testable claim about benefit. Regulators distinguish clinical outcomes from surrogate endpoints; a movement in an aging clock does not automatically establish that people will function better or live longer. Trials with meaningful outcomes can be long and expensive, making their funding and commercial case difficult. That is a hurdle to solve, not evidence that there is no possible approval path through a defined disease or functional outcome.

For calorie restriction itself, human evidence includes randomized studies as well as observations of people who restrict their own diets. The two-year CALERIE phase 2 trial, following earlier randomized pilot studies, achieved about 12 percent restriction on average and improved several cardiometabolic risk factors. A subsequent analysis found a change in DunedinPACE, a methylation measure of the pace of aging, but not in the PhenoAge or GrimAge clocks. These findings did not demonstrate human lifespan extension. Human trials certainly can measure survival—SPRINT reported lower all-cause mortality with intensive blood-pressure treatment. The harder question here is whether an intervention extends human life by changing mechanisms that contribute broadly to aging, and how a trial would distinguish that claim from preventing particular fatal diseases.

Measurement has outpaced intervention

Here is the concern I most want the field to outgrow. We have become very good at measuring and cataloguing aging, while the step from those measurements to reliable human interventions remains difficult. The clocks, the atlases and the hallmarks are genuine achievements. They support discovery and experiments, but a description of an old state does not itself tell us how to change it safely. Aging research has also produced causal experiments and interventions that extend animal lifespan; my frustration is with how much remains between those findings and a benefit people can depend on.

Many major gains in human survival have come through sanitation, vaccines, antibiotics and the prevention and treatment of cardiovascular disease. In a 2002 analysis, best-practice life expectancy had risen approximately linearly over the preceding 160 years. That historical pattern was not proof that aging had stopped or that it would continue indefinitely. It showed how much could change through improvements in the conditions of life and medicine. It also makes me reluctant to assume that future progress will respect disciplinary boundaries. The airplane required an understanding of flight that went beyond cataloguing bird anatomy. For aging, I am drawn to the work of connecting biological experiments with physics, mathematics and engineering: enough understanding to predict the consequences of a change and then make that change work.

Context changes the result

The same treatment can give different results depending on age at onset, sex, organ, cell type, genetic background and species. So I keep a checklist, and I'd suggest anyone reading a longevity paper keep one too.

We have nothing fast and complete

Aging intervention studies in mammals are time-consuming and resource-intensive: following a mouse cohort to the end of life can take about three years. A study can test multiple treatments in parallel, but adding doses, combinations, sexes and genetic backgrounds quickly increases the scale. The models fast enough for larger screens omit features that may matter for translation. Flies and worms have sophisticated biology, but neither has the vertebrate combination of adaptive immunity, closed circulation and a hypothalamic–pituitary system. The short-lived killifish is a vertebrate with those systems, but it is an ectotherm separated from us by hundreds of millions of years of evolution. Dietary restriction has extended male killifish lifespan, as some regimens do in mammals. What remains to be demonstrated is how reliably a screen in killifish can predict the benefits and failures of new interventions in mammals.

So the field is caught between models that are fast but omit important biology and models that capture more of it but are slow. Even a mouse is not a complete model of human aging. An effective screen is needed, and “effective” has to mean something specific: its readout should predict a defined outcome in the slower experiment it is meant to replace. Organoids and virtual cells are two candidates I keep returning to. Both already support useful research; neither should be treated as a general replacement for testing whole-organism aging.

Organoids offer human cells in a three-dimensional tissue-like setting. But brain organoids grown from pluripotent stem cells can resemble developing rather than adult tissue, and reprogramming to pluripotency resets epigenetic-age measures. That complicates modeling an old donor's state; it does not mean that every organoid is incapable of displaying aging-related changes. Adult-tissue-derived intestinal organoids can retain donor-associated epigenetic-age information, and fibroblasts converted directly into neurons, without passing through pluripotency, can retain donor-age signatures. Selected aspects of aged human tissue are already available in a dish.

There has also been progress in adding components that conventional organoids lack: brain organoid systems containing vascular-like networks and microglia are one example. But including these components is different from recreating their full interaction with circulating immune cells, hormones and the other organs of an aging body. I would want to know which age-related functions a particular model preserves, which ones culture changes, and whether a treatment's benefit in that model predicts benefit beyond it.

The in silico virtual cell aims to predict what a perturbation will do before anyone runs the experiment. Its performance remains a live question. In a 2025 benchmark, the deep-learning models tested did not outperform simple linear baselines on the evaluated gene-perturbation prediction tasks. An October 2026 study found that models could outperform uninformative baselines when evaluated with calibrated metrics, showing how much the answer can depend on what the benchmark measures. That does not settle which model will generalize to a new aging intervention. It makes careful evaluation, meaningful baselines and prospective experimental tests more important.

For aging, I would want training and validation data that represent different ages, cell states and responses to interventions, followed by tests in settings the model has not already seen. Predicting a transcriptional response is only one step toward predicting tissue function or survival. Both organoids and computational models therefore face a further task: establish which fast readouts reliably predict which later outcomes. A screen for cellular senescence can be useful as a screen for cellular senescence; calling it a screen for lifespan extension requires additional evidence. We can develop the screen and its validation together, but we cannot simply assume that a younger-looking measurement means a longer, healthier life.

I sort aging research into three overlapping kinds of work.

1. Describing and characterizing: biomarkers, clocks, atlases, phenotyping; increasingly combining experimental measurement with computation.
2. Intervening: dietary, pharmacological, genetic, cellular.
3. Digging into mechanism: the molecules along the pathways, how CR works, what senescence is.

My concern is that our growing ability to describe aging and investigate its mechanisms has not yet translated into a comparable ability to intervene reliably in humans. Mechanistic research can be valuable long before a treatment emerges. What I want is a closer connection between that understanding and experiments that test whether it helps us change an outcome. Below is what I would build, in roughly the order I would build it.

Everything on this list would benefit from people working across disciplines. The history of biology is also a history of instruments that made new questions possible. Lenses made cells visible. X-ray methods and cryo-electron microscopy opened up molecular structures. Sequencing combines chemistry with engineering; GFP, super-resolution microscopy and protein structure prediction brought together biological discoveries and new ways to measure or model them. Work on GFP, super-resolution microscopy and protein structure prediction received Nobel Prizes in Chemistry in 2008, 2014 and 2024, respectively. Those examples show how closely the disciplines can meet. Resolution, in space, in time, in molecules and across populations, is a recurring bottleneck. Aging confronts us with all of those limits at once. But measuring more precisely will not, by itself, tell us which change causes another, how to deliver an intervention, or whether it is safe. I expect the next decade to depend heavily on collaborations among aging biologists, physicists, chemists, engineers and computational researchers. AI belongs here too. It cannot make the years required to observe an animal's full lifespan disappear. What it might do is help us choose which studies are worth that wait. That depends on reliable predictions: some benchmarks have found little advantage over simple baselines, while a 2026 study found advantages under better-calibrated metrics. The useful question is what a model can predict in a new experiment, under conditions it has not already seen.

1. Making experiments easier to compare

I would like shared records that connect a hypothesis to the model, protocol, intervention, results and limitations, including well-documented negative results. Researchers should be able to find out what has been tried without repeatedly reconstructing the same history from scattered papers. Those records should retain details such as sex, genetic background, diet, dose and treatment timing. Otherwise, an apparent contradiction may simply reflect two experiments asking different questions. The first test of such a resource would be practical: can it help another laboratory reproduce a result or explain why a result changes?

2. Building platforms

If you want to solve something big, look for a sub-problem whose solution would make the rest easier. In aging I think there are two: automation of the wet lab, and massively parallel intervention screening. Both could reduce cost and shorten the tedious part. The worm field already has machines that repeatedly image large populations to record survival and open-source robots that do this at the bench; chemistry has self-driving laboratories that select, run and analyze experiments toward a specified goal. I would like killifish and mouse lifespan facilities to borrow some of the discipline of a chip fab: shared reference protocols, well-characterized inputs, automated records and a consistent pipeline. Animals introduce difficulties that chips do not, including husbandry, welfare, infection, variable drug exposure and years of biological variation. A common protocol would improve comparability, but it could also reproduce the same bias everywhere. The Interventions Testing Program provides a starting point for coordinated testing across sites. My ambition would be to expand that kind of infrastructure dramatically, perhaps eventually by two orders of magnitude, and connect it to short-lived vertebrate screens. Scaling would have to earn its way through pilots showing reliable endpoints, lower cost per informative result and replication across conditions.

Combination is the clearest reason to want this scale. Interventions do not necessarily add. In one mouse study, CR did not further increase overall survival in growth hormone receptor knockout mice, although it increased the study’s measure of maximal lifespan in females. An ITP study found that metformin alone did not extend lifespan at the tested dose, whereas metformin plus rapamycin did; its suggestion of an added benefit over rapamycin came from historical comparisons, not a concurrent rapamycin-only group. A more direct example is the 2025 rapamycin–trametinib study, in which the combination extended mouse lifespan more than either treatment alone. Finding useful combinations of genetic, dietary and pharmacological interventions is an experimental-design problem as much as a data problem. Pairwise experiments are useful, but they cannot exhaust the possibilities. Screening needs to prioritize combinations, then confirm the promising ones with the component treatments and appropriate controls tested alongside them.

The other platform is data. Organizations already generate longitudinal, deeply phenotyped human-aging data: the Baltimore Longitudinal Study of Aging began in 1958, and its repeated measurements span body composition, energetics, homeostatic mechanisms and neurological function. Biobanks and other long-running cohorts are substantial foundations. What I want is a better-connected combination of those strengths: frequent measurements, molecular depth, diverse participants, records that can be compared across studies, and carefully documented interventions. Wearables could add daily observations between clinic visits, although they measure only part of what matters. Repeated sampling is expensive, devices change, and people drop out; those problems belong in the platform design. No matter how good the computation is, poorly measured or systematically unrepresentative data can still produce a misleading model.

3. Continuous recording

Much of what we know about aging cells comes from snapshots: kill the animal, fix the tissue, sequence. I want to watch instead. Genetically encoded sensors can report redox changes, ATP levels and aspects of mitochondrial turnover in living cells; chronic cranial windows let researchers repeatedly image the same cells in the same mouse over months. Mitochondria would be my first target. We already have ways to watch them in living cells, but following their function through aging in intact tissue remains much harder than measuring a change before and after an intervention. Repeated monitoring could turn snapshots into trajectories, which would help us test dynamical models. It could also reveal when an in vitro experiment begins to drift from its intended conditions. I suspect that some failures to replicate would become easier to understand if we could see that history, although monitoring would not explain every failure.

The measurement itself has to be part of the experiment. Illumination can damage cells, reporters can perturb what they measure, and access to deep tissue is limited. A useful recording system would need controls showing that the sensor, surgery and imaging schedule have not created the changes being interpreted as aging. Continuous recording is the aspiration; in many tissues, repeated observations at carefully chosen intervals may be the more realistic starting point.

4. Higher resolution, especially in microscopy

Tissue-level measurements and bulk RNA samples can hide the heterogeneity that makes aging difficult to understand. Several kinds of tool let us ask more specific questions. Expansion microscopy physically enlarges preserved samples to reveal fine structure, while lattice light-sheet microscopy supports rapid three-dimensional imaging in living specimens. They solve different problems: an expanded, fixed sample cannot show a living process unfolding. Cryo-electron tomography can show protein complexes in their cellular surroundings in vitrified samples, again providing a structural snapshot rather than a live recording. And label-free metabolic imaging uses endogenous fluorescence, including NAD(P)H and FAD intensities and lifetimes, to probe metabolic state without adding a fluorescent label. These are informative proxies, influenced by molecular binding and the local environment; they do not directly measure ATP production or every aspect of mitochondrial function. Spatial transcriptomics can also place RNA signals at subcellular resolution with some methods, though resolution, sensitivity and gene coverage involve trade-offs. What I want from higher resolution is the ability to say which cell, which organelle, and when, then connect that observation to an experimentally tested mechanism.

5. Learning from nature's completed experiments

Evolution has produced a much wider range of lifespans and maintenance strategies than we can recreate in a laboratory. Naked mole rats, bowhead whales and bats are the famous cases, but the ocean holds other extremes: a Greenland shark estimated to be roughly four centuries old, with substantial dating uncertainty, black coral colonies more than four thousand years old, and an ocean quahog aged at 507 years from its shell growth record. A coral colony's persistence is not directly equivalent to the lifespan of a single shark or clam; the comparisons need biological context. We have growing genomic resources for long-lived animals, but for many marine species, keeping, breeding and manipulating them remains difficult. That is a gap I would like to work on. The AnAge database already organizes longevity information for thousands of species. Finding another long-lived animal is valuable, but I am especially interested in carrying a comparative observation forward to a causal test. Species comparisons must account for shared ancestry, ecology and uneven sampling; multiple individuals can be measured within a species, but that does not make the species independent evolutionary replicates.

It can be done. Comparing DNA repair across eighteen rodent species pointed to SIRT6, and versions from longer-lived species supported more efficient repair in cellular assays. Expressing the naked mole rat's hyaluronan synthase gene, Has2, in mice increased lifespan and improved several measures of health. Those two results illustrate the program I have in mind: identify a candidate mechanism in a long-lived animal, then test what it does in a tractable system. They do not establish that a single transferred gene explains the donor species' longevity. The deep sea adds another difficulty, because temperature, metabolism and ecology can be entangled with the mechanisms we want to identify. Separating those influences requires experiments as well as genomes. Cells derived from a biopsy could offer one route, if we can establish cultures under suitable conditions; cell survival in a dish would still need to be distinguished from whole-animal longevity.

6. Personalization

Finally, a hypothesis I would like to explore: the individual metabolic set point. I suspect that people may differ in the combination of body composition and energy use that best supports their long-term health. The stronger version of my hypothesis is that much of what diet and exercise do for aging comes from helping a person stay near that state. I do not know whether that is true; it is the claim I would want to test. By “set point” I mean a possible optimal range, not necessarily a single fixed number or the level the body naturally defends. It might change with age. There are observations that make the question interesting: IGF-1 has shown nonlinear and age-dependent associations with health outcomes, while higher adiponectin can accompany worse outcomes in older adults despite its association with favorable metabolic traits in other settings. Neither finding establishes a person-specific optimum. Illness can itself change hormone levels, muscle and fat, so reverse causation is a serious alternative.

To make this testable, the variables and outcomes would need to be defined in advance: for example, repeated lean and fat mass, energy expenditure and glucose regulation, followed by physical function, disease events or survival. The model would need to predict future outcomes in people it was not fitted to, and a later intervention would need to show that moving toward its predicted range actually helps. The study of 960 genetically diverse female mice is an important challenge to a simple version of the idea: reduced adiposity and lower fasting glucose under dietary restriction did not predict longer life. A good model has to account for that, rather than defining “optimal” after seeing who lived longest.

If the idea holds up, measurement could help guide diet, exercise or other interventions, somewhat as continuous glucose monitoring provides feedback about one variable. The monitor would not itself keep a person at the desired state, and we do not yet have a continuous monitor of all the relevant features of body composition and metabolism. I hold this loosely. Longitudinal data could test specific predictions and expose weaknesses; demonstrating a benefit for human aging would take more than a year of observation. That is another reason I want the dataset.

The strategies we have

I think many current anti-aging strategies can be grouped into three overlapping categories. These describe intentions, not equal levels of evidence.

1. Clear the junk. Senolytics, including experimental senolytic CAR-T cells; approaches that promote autophagy, such as rapamycin and spermidine; and attempts to remove particular aggregates or accumulated material, including amyloid and lipofuscin. Plasma dilution is also being investigated as a way to alter the circulating environment, although “junk removal” does not establish its mechanism. These approaches make aspects of the damage hypothesis experimentally actionable.
2. Turn bad cells into good ones. Partial reprogramming, genetic or chemical; niche repair, such as changing stiffness or inflammation; and restoring factors whose activity has declined. For example, raising SIRT3 in aged mouse hematopoietic stem cells improved regenerative capacity. Its importance depended on age and stress, rather than implying that it has no function in young cells. “Bad” and “good” are shorthand here; the important question is which function is restored and whether that restoration lasts.
3. Replace. Stem cell transplantation and engineered cells that might restore lost function or compete with dysfunctional clones. CAR-T cells belong mainly in the clearance category above: killing a target cell does not automatically replace it. Regeneration in other animals shows what biology can achieve: a deer regrows its antlers from the pedicles, and some sea cucumbers regenerate internal organs after ejecting them. Transferring that capacity to an aging human tissue would require control over growth, integration and function.

What AI is for

I see two concrete uses for AI in this field, and a harder problem that remains. First, using existing data to make predictions worth testing. Protein structure prediction is an example of what computation can achieve for a clearly defined molecular task. Virtual-cell programs have a different and more demanding ambition: predicting how a cell responds when a gene, molecule or environment changes. Knowing a protein's structure does not by itself tell us the causal effect of removing it from a pathway. I want models that help identify which perturbation to try, with predictions checked prospectively against simple alternatives. Second, synthesizing the literature so that a researcher can see contradictory views on a question in a page rather than discover them over a year of reading. That summary needs traceable sources and enough experimental detail to show whether the studies really disagree.

The harder problem is interpreting an experiment: deciding whether a failure arose from dose, timing, strain, measurement or the hypothesis itself. AI may help propose and compare explanations, but a model cannot settle a question that the experiment did not distinguish. That still calls for judgment and, often, another carefully chosen experiment. Better data will make that reasoning more useful; it will not make the need for it disappear.