DJ PAOLETTI
Cambridge, MA
// hi.

I'm Dylan. I think about
how to not die.

I'm building a company doing genetic circuits for cancer. On the side, I think a lot about how to make humans live a long time. The body falls apart on a schedule, and nobody has a serious plan to stop it. I have one, written down here.

I also keep a separate page of things I like and things I think: shows, music, half-formed opinions, predictions I'll probably regret.

Email's in the footer if you want it.

nowBuilding genetic circuits for cancer readingComparative longevity genomics writingthe longevity notebook updatedApril 2026
DJ PAOLETTI
// notes on not dying // some links and sources still need updating or repair working draft · 2026

I want to help people live longer and live better.

A working notebook on why human bodies break down, what's known about stopping it, and what an honest attempt looks like. Parts of this are wrong, and I'll change them when I figure out which.

§ 01   The problem Why one drug at a time won't get you to two hundred

Roughly a hundred thousand people die every day from causes whose upstream driver is aging. They don't die of "aging" on the death certificate. They die of cancer, heart disease, stroke, pneumonia, dementia. The underlying reason most of them die is that for the last third of their life, the body's housekeeping has been falling apart.

Your cells are constantly fixing themselves. They repair broken DNA, kill cells that have gone wrong, replace tissue that wears out, take out the trash, and watch for invaders. From adolescence into your early twenties, all of this happens in the background and you don't notice. By seventy, every one of those processes is running slower and worse. DNA breaks don't get patched properly, so cells accumulate mutations. Cells that should die quietly hang around leaking inflammation. Stem cells stop making replacements. The lysosome, the cell's recycling bin, clogs with garbage proteins it can no longer break down, which is why old neurons fill up with plaques and tangles. The thymus, where new immune cells are trained, turns to fat by your forties, which is why old people get sick more easily.

None of these is "the" cause. They all are, and they drive each other. Bad DNA repair makes more damaged cells. The immune system can't clear them. The resulting inflammation damages more DNA in turn. The same loop, running everywhere at once, for decades.

The way the world currently deals with this is one drug for one downstream disease at a time. A drug for the high cholesterol that built up over forty years. Surgery for the heart attack it caused. Chemo for cancers that emerged because DNA repair has been failing for decades. A few approved drugs touch upstream aging biology by accident (metformin nudges AMPK and mTOR, statins reduce inflammation as much as LDL, GLP-1 agonists hit several aging pathways at once), but none of them is approved for aging, and none of them was designed to be. The most reproducible drug for extending mammalian lifespan is rapamycin: in the NIA's Interventions Testing Program, it extends median lifespan in mice across multiple genetic backgrounds and starting ages, including when started late in life. Immunosuppression and metabolic side effects keep it from being a casual prescription.

The regulatory machinery was built around single mechanisms and single diseases. Aging is neither. That's softening: in 2015 the FDA agreed to a multimorbidity endpoint for TAME (Targeting Aging with Metformin), the first regulatory foothold for "aging" as a target rather than a specific disease. Partial-reprogramming trials are now moving toward patients. Still, the dominant pattern is one drug, one disease, one mechanism. Anything aimed at "all of aging" lacks a clean regulatory target, so it costs more to develop than something with a tidy indication. The money goes where the path is shortest: most recent private longevity funding has flowed into cellular reprogramming. Altos Labs, Retro Biosciences, NewLimit, and a long tail in the same lane. The other eleven things going wrong inside an aging body share whatever's left over. We've built an industry around one tool and we're trying to use it for everything.

// a brief history of the war on time

The modern era starts in the late 1990s with two converging discoveries. Cynthia Kenyon at UCSF showed that a single mutation in daf-2 doubled the lifespan of C. elegans, which established that aging is something genes can move. Leonard Guarente at MIT showed that SIR2 extended yeast lifespan, which opened up the sirtuin pathway. Everything since has been built on top of those two results.

The first big collision with hype happened over sirtuins. David Sinclair, a former Guarente postdoc, founded Sirtris Pharmaceuticals in 2004 around the idea that resveratrol activates sirtuins and extends lifespan in mammals. GSK bought Sirtris in 2008 for $720M. By 2010, Pfizer scientists had shown they couldn't replicate the central biochemistry: the original assay turned out to be an artifact of resveratrol interacting with the fluorophore tag on the substrate, not with the sirtuin itself (and an Amgen team had reached the same conclusion a year earlier). GSK quietly closed Sirtris in 2013. It set the tone for the next decade: anyone claiming a single molecule cures aging had to clear a much higher bar. Charles Brenner at City of Hope has been the most visible critic since, arguing in Life Metabolism in 2022 that sirtuins aren't conserved longevity genes and the entire sirtuin-activator framework was overhyped, and continuing to argue across op-eds and social media that Sinclair's later claims (the 2023 chemical-reprogramming paper in Aging, the journal Sinclair himself co-edits, with a 12-day review cycle versus the usual 14 weeks) have followed the same overclaim-and-revise pattern. The personal animus is real (both have commercial stakes in NAD+ precursors), but the substantive concerns aren't unique to Brenner. Paul Knoepfler and many other stem-cell scientists have raised the same questions.

The second collision was GDF11. In 2013--2014, Amy Wagers at Harvard and Richard Lee at Brigham & Women's published two high-profile papers in Cell and Science claiming GDF11 was a circulating "young blood factor" that rejuvenated old hearts and muscle. The papers fueled the parabiosis narrative and drove substantial investment. A 2015 paper from Novartis showed the antibody used to measure GDF11 wasn't actually GDF11-specific: it cross-reacted with myostatin, and the original effects didn't reproduce. The story had to be substantially walked back. Later plasma-fraction work from the Conboy lab at Berkeley suggested that removing pro-aging factors via plasma dilution may matter more than adding pro-young ones, which is a different story than the original.

The third collision is in slow motion: the senolytic field. The biology is solid. Senescent cells accumulate, secrete inflammatory factors, and can be selectively killed. Jan van Deursen's 2011 demonstration that genetically clearing p16-positive senescent cells extended healthspan in mice is one of the cleanest results in geroscience. But the translation has been bumpy: dasatinib + quercetin trials have produced mixed and modest human results, and Unity Biotechnology's UBX0101 failed Phase 2 in 2020. The insight is real. The drugs are harder than expected, mostly because senescent cells are heterogeneous, hard to identify in vivo, and live alongside critically important non-senescent populations.

Meanwhile, more disciplined work has built up. The NIA's Interventions Testing Program, three-site blinded replication since 2004, is the gold standard. rapamycin replicates across all three sites. resveratrol doesn't. Most popular supplements don't either. Vera Gorbunova and Andrei Seluanov at Rochester have produced consistent comparative-genomics work on naked mole rats and bowhead whales, including the 2023 demonstration that transferring the naked mole rat's high-molecular-mass hyaluronan to mice extended their lifespan and reduced cancer: a textbook example of borrowing alleles from a long-lived animal and watching them work. Maria Blasco's lab at CNIO has consistently produced high-quality telomerase work, including the 2012 gene-therapy result that opened up viral delivery for telomere maintenance. Andrea Ballabio's lab in Naples owns the TFEB and lysosomal-biogenesis field. These are the labs whose results have replicated and whose claims have aged well, and they've been a lot quieter in the popular press than the Sinclair lab.

The money concentration matters. Reprogramming gets the bulk of recent private capital: Altos Labs launched with a $3B seed in January 2022 from Bezos and Milner, with Yamanaka and Belmonte on the founding team. Retro Biosciences, backed by Sam Altman, took $180M and then a $1B Series A at a $5B valuation. NewLimit is Brian Armstrong's. None of that is random. Reprogramming pattern-matches to a normal drug program: one mechanism, one set of factors, one clean story for the FDA. It also has Yamanaka's Nobel behind it, which makes recruiting talent and capital easier than for messier multi-pathway work. Calico went the other way and picked a Bell-Labs-of-aging basic-research orientation under David Botstein; Aubrey de Grey called it "my biggest disappointment" in the field. Botstein quietly retired in late 2022. The TAME trial has been chronically underfunded for a decade. The supplement industry sells hundreds of millions in NMN, NR, and resveratrol every year on evidence that hasn't held up in rigorous trials. Nothing here is corrupt in any specific way. It's the normal incentive structure of a hard field with decade-long timelines and a public that wants "the cure for aging" yesterday.

Anyone walking into longevity now should assume some real fraction of what they cite will turn out to be partly wrong, treat single-molecule miracle stories with twice the skepticism they'd apply to oncology claims, and remember that the labs whose work replicates aren't always the ones with the most press. Geroscience has been bad at being defensive against its own optimism.

// the better idea, in one paragraph

Address all of the upstream problems at once. Restore the housekeeping across every system that's failing. And since you're already in there editing genomes, install a handful of specific upgrades that most humans don't naturally have, but that a small number of people are born with.

"There are no laws of physics preventing humans from living as long as bowhead whales. We believe aging is negotiable -- we'll see."
§ 02   Animals that already live a long time Whales, elephants, hydra: studying the upperclassmen

Some animals get to live a very long time. Bowhead whales make it past two hundred years, apparently in good shape until the end. Naked mole rats are small rodents that live thirty years instead of three, and almost never get cancer. Hydra, the little freshwater organism, don't seem to age in any measurable way at all.

Look across them, and across other long-lived species, and the same three moves keep showing up. Different proteins, different organs, different scales, but the same three ideas. Evolution has converged on these solutions independently across distant lineages, which is the strongest hint you get in biology that the three are doing something real.

i.
Don't break in the first place.
Keep the cellular damage rate so low that ordinary repair keeps up. Better DNA repair. Better mitochondria. Better antioxidants. Less stress on everything.
ii.
Have so many backups that breaking one doesn't matter.
Redundancy. Multiple copies of the genes that catch cancer. Layered checkpoints. When any one fails, the rest hold the line.
iii.
Constantly replace what's broken.
Keep turning over tissue. Don't let individual cells live long enough to rack up serious damage. Constant renewal beats perfect maintenance.
HYDRA
biologically non-aging · every cell replaced every few weeks through continuous stem-cell turnover.

Different species pick different mixes. Bowheads lean hardest on the first one: keep the damage rate down. Elephants lean on the second: pile up redundancy. Hydra lean on the third: constant turnover. Naked mole rats use a fourth-ish trick on top, high-molecular-mass hyaluronan that makes their cells stop dividing the moment they touch each other. It's a version of the second move in its own register.

Humans use almost none of these aggressively. We have decent DNA repair, decent antioxidants, decent stem cells, decent tumor suppression. Decent, in evolutionary terms, means "enough to get to reproductive age and a bit beyond." The classical evolutionary theories of aging (mutation accumulation, antagonistic pleiotropy, the disposable soma) all converge on the same prediction: selection pressure on maintenance drops off sharply once an organism has reproduced. Our cells were never under serious selection to last 120 years in good shape. The rest of this notebook: borrow the best ideas from each of those three moves. The pieces already exist in living organisms. Moving them across is the part nobody's done.

glossary DNA repair -- the machinery that fixes DNA breaks; ~150 genes in humans.
TP53 -- "guardian of the genome"; the protein that triggers cell-cycle arrest or apoptosis when DNA damage is sensed.
Contact inhibition -- the property of normal cells that stops them dividing when they touch their neighbors. Cancer cells skip this.
Stem cell turnover -- replacement of differentiated cells by new ones made from tissue-resident stem cells.
-- -- --   ×   -- -- --
§ 02b   My theory of why aging is the way it is The Conservation Theory of Aging: p53 picking from three losing hands

I have my own theory of aging. Here it is.

The version I'd defend, and which I'll call the Conservation Theory of Aging: aging is what falls out when p53 has to pick among three bad options inside the cell, runs into a hard replicative ceiling on the cleanest of them, and evolution refuses to pay the calorie bill for any honest way around it. Three lousy choices, one ceiling, and a stingy referee.

// the three bad options

When a cell senses DNA damage, it picks one of three things to do. It can repair, which means fixing the damage and continuing. It can go senescent, which means stopping division while staying alive. Or it can commit apoptosis, which is the cell's polite term for self-destruction. p53 sits near the center of that decision, biasing the cell one way or another depending on how bad the damage is, what else the cell is dealing with, what the rest of the body is signaling, and a constellation of post-translational marks no one fully understands yet.

p53 isn't the only voice in the room. The broader DNA-damage-response network, the retinoblastoma pathway, the p21/p16 axis, and a few metabolic sensors all push on it. But it's the transcription factor the others mostly route through, and it's the first place to look.

The standard reading is that the three responses are a graded tumor-suppression program. Small damage gets repaired, dicey cells get benched, dangerous cells get executed. Fine, as far as it goes. What it leaves out is that each of the three is buying short-term safety by handing the body a different long-term IOU.

Repair leaves you with mutations. The cell survives, the tissue holds, nothing has to be rebuilt, but repair isn't clean work. What it leaves behind is residual mutation, epigenetic drift, and mitochondrial junk. Over decades, your repaired cells stop being the cells you started with.

Senescence leaves you with inflammation. The cell quits dividing, cancer risk on that line collapses, nothing has to be replaced. A great deal until the bill comes due in your sixties, when senescent cells start secreting the SASP, irritating the neighborhood, and recruiting other cells into joining them.

Apoptosis leaves you with regenerative debt. The damaged cell dies clean and leaves nothing behind, but the body has to put something back where it was. Stem and progenitor cells divide more often to do the replacing, and the replacement isn't free.

// the math of it

The argument from here uses four numbers and one piece of vocabulary. Get those in place first, then we can do the comparison.

The vocabulary. Hayflick and Moorhead's 1961 paper showed that somatic cells stop dividing after a fixed number of rounds: roughly 50 in their original fetal-fibroblast experiments, with later work putting the range at 50--70 depending on cell type, donor age, and growth conditions. I'll call each of those divisions a Hayflick token: a non-refundable replication credit the lineage spends every time one of its cells divides, and runs out of when telomeres get short enough to force replicative senescence. The ATP a cell spends today gets paid back by tomorrow's meals. The token doesn't.

The numbers. Four of them, all from direct measurements in published papers:

(a) Daily ATP turnover, average human cell  ≈  1012--1013 ATP
   -- Flamholz, Phillips & Milo 2014, dividing total-body O2 consumption by ~1013 cells

(b) Daily DNA lesions per cell  ≈  104--105
   -- Lindahl 1993, the canonical estimate of spontaneous damage load

(c) ATP cost per nucleotide synthesized de novo  ≈  ~50 ATP
   -- Lynch & Marinov 2015, direct biochemical accounting

(d) ATP cost per mammalian cell division  ≈  ~1013 ATP
   -- Lynch & Marinov 2015, allometric scaling from E. coli chemostat data to mammalian cell volume

Two of those numbers, (a) and (d), are at the same order of magnitude, which makes physical sense: a cell that doubles on roughly a daily timescale has to spend roughly a day's worth of energy building the next one. That's the central physical fact the whole argument rests on. One cell division costs about one cell-day of total metabolism.

The upper bound on repair cost. I deliberately don't want to estimate this from pathway-frequency assumptions, because the per-pathway fluxes aren't measured directly and any number I'd give would be a guess. The honest move is an upper bound instead. The most expensive single-lesion repair pathway is nucleotide-excision repair, which patches around 30 nucleotides per event. From (c), that's 30 × 50 = 1,500 ATP per lesion, ceiling. Apply that ceiling to every lesion in (b), even though most lesions are handled by cheaper pathways:

105 lesions/day × 1,500 ATP/lesion = 1.5×108 ATP/day

That's the worst case: every lesion gets the most expensive repair. Set it against (a), the daily cell budget of 1012--1013 ATP, and you get 0.0015% to 0.015% of the daily energy budget, ceiling. The real number is lower, because most lesions are handled by single-nucleotide base-excision repair (~50 ATP, roughly 30x cheaper than the NER ceiling). The upper bound alone is enough to make the qualitative point. Yang et al.'s 2021 PNAS Perspective says it outright: "the energy budget of cells remains largely unexplored". Nobody has a measured number for repair as a fraction of cellular ATP, but the upper bound is enough. Rolfe & Brown's canonical 1997 partition of cellular ATP usage doesn't list DNA repair as a line item for the same reason: the partition resolves protein synthesis (25--30%) and ion pumping (20--30%), but its noise floor is at the 1% level. Repair is two orders of magnitude below that floor even at worst case.

The comparison. Now the three responses laid out side by side, in two currencies:

repair  ≈  ≤ 1.5×108 ATP/day  + 0 Hayflick tokens
senescence  ≈  ~0 direct ATP  + 0 tokens; inflammation bill comes due later
apoptosis & replace  ≈  ~1013 ATP  (one full cell-day of metabolism) + 1 Hayflick token

cost ratio: one division / one day of repair  ≈  1013 / 1.5×108 ≈ 6.7×104×

So an apoptosis-and-replace event costs the body about sixty-five thousand days of upper-bound repair work on the cell that got killed, and that's against the ceiling. Against the realistic per-day repair load, the ratio is bigger by another order of magnitude or two. Plus a Hayflick token from the replacing lineage, which doesn't get refunded.

Senescence is even more interesting in this frame. The cell that goes senescent pays nothing directly. No replacement, no Hayflick token, no immediate metabolic cost. The bill is paid decades later, by other cells, as the SASP irritates the tissue around it. SASP costs are diffuse, smeared across the tissue and time-shifted by decades, which is why selection lets the cell keep choosing them and why senescent burden piles up the way it does.

This is where the theory's name comes from. The cell is structurally conservative. The rule it follows isn't "minimize damage". If it were, repair would have a clean fraction in Rolfe & Brown's partition, and it doesn't, because repair is four or five orders of magnitude cheaper than a division. The rule the cell actually follows is closer to "minimize divisions, because every division costs a full day of metabolism plus a non-refundable Hayflick token, while a day of repair barely registers." Repair is cheap on both axes.

// the part I'm less sure about

The native p53 setting is the one that minimizes ATP-plus-Hayflick spend under ancestral conditions. Lower the apoptotic threshold and the Hayflick budget burns down faster: you spend tokens and a day's metabolism per replacement to dispose of cells that the conservative setting would have repaired for a tiny fraction of either. Raise the threshold and the mutation and senescent piles get higher instead. The native p53 bias we're all born running is whatever balance kept those two debts roughly equal across an ancestral lifespan. Not a careful trade, just the setting selection happened to land on and stop tuning.

Don't push the apoptotic threshold harder. Break the coupling between the two costs. Pair the threshold change with telomerase, so the Hayflick token gets refunded each time you spend one. The ATP cost stays. The lineage exhaustion doesn't.

// evolution didn't because evolution couldn't afford it

You could build a body that doesn't have this problem. Push p53 toward apoptosis, install telomerase upkeep so the Hayflick ceiling stops biting, beef up the stem-cell reserves and the protein-clearance machinery and the mitochondrial quality control so they can absorb the higher turnover. Senescent burden drops. Mutation burden drops. Regenerative reserve holds. The body runs hotter and outlasts its predecessor.

Evolution didn't install that body for an unromantic reason: it eats too much. A resting human already spends most of 1500--2000 kcal/day on cellular housekeeping. Add more apoptotic flux, telomere upkeep, sharper proteostasis, faster mitochondrial turnover, and you're looking at tens of percent more, conservatively, possibly much more depending on the tissue. In any environment where calories are the binding constraint, selection punishes that body harder than it ever rewards post-reproductive longevity, which it more or less can't see. Evolution declined to pay. Aging is what falls out when the upgrade doesn't get installed.

This is the disposable-soma idea, made specific. Kirkwood describes the soma-vs-germline tradeoff at the level of resource allocation. The Conservation Theory says which molecule implements the tradeoff (p53), which ceiling it's pressed against (Hayflick), and what currency evolution refused to spend (calories).

// the rest of this notebook is the upgrade

If the picture is right, then the "secret" to longer life is less about patching damage and more about whether you can buy back the body that scarcity refused to build. Move the p53 bias, pay the Hayflick cost with telomerase, and pair both with the downstream machinery that can absorb higher turnover.

The constraint that shaped the original deal, calories, isn't really binding anymore. Global caloric supply per capita has roughly tripled since industrialization. AI, robotics, and industrial automation keep pushing food, energy, and material abundance cheaper. The thing getting scarcer every year is time. Evolution optimized humans around energetic scarcity because it had to. In a future that can run gene and cell therapies that extend healthy lifespan by multiples, the marginal cost of sustaining a body longer is negligible, especially with birth rates falling. Trading future surplus for more years of healthy life is exactly the move evolution was never allowed to make. Doing it deliberately, through engineering, is what comes next.

What's open: the relative weights on mutation, inflammation, and regenerative debt. The actual size of the energy gap. Whether the senescence branch can be edited out without something downstream breaking. All of it.

-- -- --   ×   -- -- --
§ 03   The plan Seven layers, plus the delivery problem they all wait on

An elixir of life, split seven ways.

// before any of it: how do you get it in there

For something this ambitious to work, the delivery vehicle has to clear a bar that nothing in the clinic currently clears:

  • Low or no immunogenicity, so pre-existing antibodies don't disqualify a chunk of patients, and so the dose can be repeated.
  • Payloads in the tens of kilobases, ideally 50 kb+. The full set of edits, plus the regulatory elements and insulators that go with them, is large. AAV tops out around 4.7 kb; lentivirus around 9--10 kb. Both are short.
  • Near-perfect broad tropism: the same infusion has to reach every cell type that needs editing, not just hepatocytes or muscle.
  • Stable, durable integration: edits persist through effectively unlimited divisions, without insertional mutagenesis and without silencing on therapeutic timescales.

None of those properties exists at the level required yet. Several groups are working toward each. AI-guided sequence design is most likely what closes the remaining gaps, and is probably what makes the whole thing possible at all.

The more imaginative alternative skips in vivo delivery as the primary route of administration. Ex vivo: take the patient's somatic cells, reprogram them into iPSCs, do every edit in culture (where each one can be QC'd individually and a built-in selectable marker out-competes any cells that lose the cargo), then differentiate the modified iPSCs into the cell types you need and reimplant them. Making room for the engineered cells probably needs a conditioning regimen along the lines of standard bone-marrow-transplant protocols. This would be harder to design, more dangerous (you'd be replacing most native tissue with new engineered cells), and more expensive than gene therapy, which is already expensive. Whichever route closes its gap first wins. I'd bet on the first.

i. Make DNA repair better than it is.

Every cell takes thousands of DNA-damaging hits a day: oxidative base modifications, single-strand breaks, the occasional double-strand break. The body's repair pathways are good but not great, and they get worse with age. The pattern across long-lived animals is consistent: they repair faster, cleaner, and with fewer errors than we do.

The first thing I'd try is borrowing the parts that already work. Import the bowhead's repair variants of ERCC1, PCNA, and the proteins around them, and add SIRT6 with the naked-mole-rat-specific changes that drive its DNA-repair advantage. Tian et al. 2019 mapped the five residues that make the long-lived-rodent variant a stronger repair enzyme across 18 species, and Kanfi et al. 2012 had already shown that transgenic overexpression of SIRT6 itself extends lifespan in male mice. Together that's the closest thing to a "longevity gene" knock-in result that's held up. None of it guarantees the same alleles translate to humans, but the comparative-genomics signal is strong enough that they're the obvious first edits to try.

Pair it with CIRBP, the cold-induced RNA-binding protein bowheads keep elevated, which substantially reduces DNA-damage-induced apoptosis without compromising tumor suppression. CIRBP is the easy win: small coding sequence, well-characterized in mammals, the kind of allele that would have been selected against in short-lived hominids who needed quick clearance of damaged cells.

Done right, this is the floor of everything else. Better repair upstream means everything downstream has less to clean up.

ii. Build redundancy into the parts that can't fail.

Some genes can't be allowed to fail. TP53 is the obvious one. It's the surveillance protein that decides whether a damaged cell repairs, arrests, or dies. If TP53 stops working in a single cell, that cell can start the cascade into cancer. Humans get one functional copy on each chromosome. Elephants get roughly 20. When one fails, the others are still watching.

The idea: knock in additional TP53 copies into safe-harbor loci. The protein is the same. The cell ends up with multiple independent sources of it, so silencing or mutation of any one copy doesn't lose surveillance. Pair it with the naked-mole-rat HAS2 variant that produces high-molecular-mass hyaluronan, the one that makes cells stop dividing the second they crowd each other. The 2023 paper from Gorbunova and Seluanov showed that transgenic mice carrying the naked-mole-rat HAS2 variant lived longer, got cancer less, and had lower systemic inflammation. That's the first direct demonstration that this kind of comparative-genomics allele transfers cleanly into a different mammal. Encouraging, but no guarantee it transfers a second time.

Worth folding in Karen Vousden's point that TP53's downstream effects are tunable. Not every activation needs to push the cell into apoptosis. Some should push toward repair, some toward arrest. The native version does this implicitly. An engineered version could do it explicitly. Several copies with slightly different effector biases would give the cell more dials than its current one-copy-binary-output setup.

The whole layer is the cheap, conservative version: more of what already works. It doesn't change the rules. It just makes them harder to break.

iii. Lower the apoptotic bar, and let telomerase pay the cost.

The body's quality-control rule, set by TP53 and the apoptotic machinery around it, is roughly "if a cell has accumulated this much damage, kill it." The threshold is high. It has to be: cells are expensive to make, the body needs the work they do, and aggressive killing wrecks tissues. So the bar gets set conservatively, with some slack. That slack is where cancer hides. A damaged cell that should have been killed slips through, and some of those slips become tumors. Elephants have twenty copies of TP53 precisely so that no single failure of the surveillance machinery is enough to let one through. They've lowered their effective threshold by piling on redundancy.

Lower the apoptotic bar. Overexpress telomerase. Let them compensate for each other. The cell turns over more often, so cancer doesn't get to incubate, and the Hayflick limit doesn't catch up because every division refills the telomere reserve.

A more aggressive apoptotic threshold means cells that are even slightly damaged get killed and replaced. Pre-cancerous cells barely get to exist. The tumor-suppressive effect should be large. The cost is that the tissue is now turning over much faster than it would naturally, which means stem cells are dividing more often, which means telomeres shorten faster, which means the Hayflick limit arrives sooner. Without a fix, you've traded cancer for premature stem-cell exhaustion.

The fix is hTERT overexpression. Telomerase reactivation in the soma is normally framed as cancer-risky. Here, it's the opposite. The cell is only being pushed into more division because the apoptotic threshold has been lowered, and the telomere problem only appears because turnover has gone up. Telomerase here is the compensating half of an engineered pair, not a separate bolt-on. The cancer risk it would otherwise raise gets bought back several times over by the lowered apoptotic threshold, in theory. Experiments tell you the ratio.

Apoptosis and telomerase are one dial expressed twice. You tune one against the other. Telomerase gene therapy on its own already delays aging in mice without driving up cancer, and the whole idea is built on the point Titia de Lange has been making explicit for years: telomere length and the shelterin complex decide whether a cell can keep dividing safely.

One way to build it: a bi-directional promoter with a p53-responsive element in the middle, sensitized so it fires sooner than the native threshold. Apoptotic cassette running off one side, hTERT off the other. A cell can't get the proliferation benefit of hTERT without also turning on the quality-control half that keeps it honest. Two or three copies in different safe harbors, so silencing one doesn't release telomerase on its own. The others are still watching. Knock out the endogenous hTERT promoter with prime editing so the engineered telomerase is the only source. One coupled circuit, not two independent edits. One sketch. Better ones exist.

// a more ambitious version: build a new p53

One thing the sketch above leaves on the table: it uses the native p53 protein, so activating it brings all of p53's downstream effects, including the ones we'd rather skip. p53 induces a fan of outcomes: repair, cell-cycle arrest, apoptosis, senescence, metabolic reprogramming, autophagy, and the balance between them is set by post-translational modifications, cofactors, and the specific response elements engaged. Senescence is a real problem at the turnover rates this circuit calls for. Senescent cells leak inflammatory cytokines (the SASP), and you don't want a circuit generating them every time it fires.

An alternative I find interesting: high-throughput protein engineering of p53 itself. Deep mutational scans and ML-guided directed evolution can now sample millions of variants per round and pull out ones with specific functional profiles. The aim: a designer p53 that induces repair, cycle arrest, and apoptosis at lowered thresholds but does not push the cell into senescence. Damaged cells go quietly via apoptosis and get replaced. Repairable cells get repaired. The senescence branch, which throws off chronic inflammation for years, gets engineered out. Some natural p53 variants already separate these axes partially. Whether you can pull them fully apart in an engineered variant is open. AI-driven protein design is starting to get good at this kind of problem, and the variant-p53 effector bias becomes another knob the coupled circuit can dial against the telomerase ceiling.

All of it tunable. The apoptotic threshold, the telomerase ceiling, the bi-directional ratio, the senescence-biased p53 variant: each is a knob you'd play with in iPSC experiments long before going near an animal. The layer also behaves well with the others. Better repair (layer i) cuts the damage the apoptotic circuit has to respond to. Autophagy (layer v) handles non-genomic stressors. The immune rebuild (layer vii) clears apoptotic debris faster.

Push this circuit hard and what stops you isn't biology, it's food. More turnover, more cells, more calories. A body engineered like this has to eat more, which is a much easier problem than aging, and one civilization is already solving. The bet is trading calories the future will have in surplus for years the present doesn't.

One more thing to remember. About 4--11% of cancers skip telomerase entirely and use a recombination-based mechanism called ALT: Alternative Lengthening of Telomeres. The setup above does nothing about ALT. You'd also want to keep ATRX functional, since losing it is the main switch into the ALT pathway, and screen for the markers (C-circles, ALT-associated PML bodies). Otherwise you've built a beautiful trap and left one door open.

glossary Apoptosis -- programmed cell death; the body's clean shutdown of cells.
p53 / TP53 -- the protein that decides whether a damaged cell repairs, arrests, dies, or goes senescent.
Senescence -- a cell-cycle exit state where the cell stops dividing but doesn't die; secretes inflammatory factors (the SASP).
Telomere -- repetitive DNA cap (TTAGGG) at the end of each chromosome.
Telomerase / hTERT -- the enzyme that extends telomeres.
Hayflick limit -- the number of divisions a normal human cell can do before telomeres force it to stop; ~50--70 depending on cell type and donor age.
Hayflick token -- shorthand used in this notebook: one of those finite divisions, treated as a non-refundable budget item the lineage spends every time a cell divides.
Bi-directional promoter -- a regulatory element that drives transcription in both directions, forcing co-expression.
ALT -- Alternative Lengthening of Telomeres; the telomerase-independent backup some cancers use.
Deep mutational scan -- experimental method that measures the functional consequence of every possible point mutation in a protein.
iv. Reset the age of cells without erasing what they are.

Cells store their identity in something called the epigenome: chemical marks on DNA and on the histones DNA wraps around that decide which genes get read. Liver cells and skin cells have the same DNA but very different epigenomes; that's how one genome makes hundreds of cell types. The pattern also drifts with age, in ways that turn out to be predictable enough to read someone's age off it. Steve Horvath's epigenetic clocks are accurate to within a few years, and improved second-generation clocks like PhenoAge now predict morbidity and mortality on top of chronological age.

In 2006, Shinya Yamanaka showed that switching on four genes (OCT4, SOX2, KLF4, c-MYC, the Yamanaka factors) is enough to walk an adult cell back to something almost identical to an embryonic stem cell. He won the Nobel for it. Useless for longevity on its own: the last thing you want at sixty is for your liver cells to forget they're liver cells.

What came later is more interesting. If you turn the factors on only briefly, cells rewind their epigenetic clock partway without losing what they are. They stay liver cells, or skin cells; they're just younger ones. Cycling them on and off in mice with an accelerated-aging disease extended lifespan and softened the disease, with no tumors in published cohorts (work from Belmonte). In normal old mice, a single dose of three of the four factors (dropping the cancer-prone c-MYC) doubled remaining lifespan in the male cohort.

The danger is real. The Yamanaka factors overlap heavily with the proteins that cause cancer. Run them long, run them hard, run them in the wrong tissue, and you get teratomas. The fix is a better promoter, not more clever cargo. One that fires only when a cell is showing damage signals (NF-κB activity, oxidative stress, p16), or that drips OSK out at a constant low level too faint to dedifferentiate but high enough to keep the methylation clock from drifting forward. Drop c-MYC and you give up some potency for a much wider safety margin. The whole thing lives or dies on the regulatory element, not the cargo.

glossary Epigenome -- chemical marks on DNA and histones that control gene expression without changing the sequence.
Yamanaka factors / OSKM -- Oct4, Sox2, Klf4, c-Myc.
OSK -- the same minus c-Myc; safer.
Stress-induced promoter -- a regulatory element that fires only in response to specific damage signals.
v. Help cells take out their trash.

Cells produce trash. Damaged proteins, broken organelles, aggregates of stuff that should have been recycled. The proteasome chops up individual misfolded proteins; autophagy ("self-eating") packages bigger debris into membrane sacks and ships them to the lysosome, the cell's recycling center, where enzymes break it back into building blocks. Both systems get worse with age. The lysosome slows down, accumulates undigestible junk (lipofuscin), and starts leaking.

This matters most for the brain, because neurons mostly don't divide. They can't dilute their garbage by splitting. Whatever they fail to clear, they keep. Alzheimer's, Parkinson's, Huntington's: at the cellular level, all three are diseases of failed clearance. Amyloid-β, tau, α-synuclein, polyglutamine repeats. All the DNA repair in the world doesn't help a neuron drowning in protein aggregates.

One transcription factor, TFEB, runs most of the autophagy-and-lysosome program. When it's active, the cell builds more of every gene it needs to digest its own waste. Mice given extra TFEB live longer and stay healthier (Ballabio). Stabilized versions clear tau aggregates in mouse Alzheimer's models. In C. elegans, the equivalent gene (HLH-30) is required for lifespan extension across six different longevity pathways, which is unusual. You don't see that kind of convergence often.

The proposal: install a stabilized version of TFEB so the cell holds young-cell-level autophagy capacity instead of letting it drift downward. Without something like this, everything else falls short. DNA repair fixes the genome, but the genome doesn't help if the rest of the cell is choking on its own debris.

vi. Add specific upgrades that some humans already have.

Layers i--v aim to fix things that are breaking. This one is different. It adds specific genetic variants (almost all of which already exist in some living humans) that confer measurable advantages. Most aren't speculative. Each has population-scale evidence behind it. None individually is a guaranteed transfer (population genetics tells you something different than what a one-cell knock-in tells you), but each is a strong starting hypothesis.

Once you're already editing a genome to install repair upgrades, the marginal cost of also installing the variants that protect against the most common causes of death, and a few that point past survival, is roughly zero. Not doing it is harder to defend than doing it.

Knock out PCSK9. Some people are born with a non-functional copy. Their LDL cholesterol is low from birth, and in cohorts with nonsense mutations, their heart disease risk is something like 88% lower (work from Hobbs and Cohen). Two FDA-approved monoclonal antibodies and one approved siRNA already mimic this state pharmacologically. Verve Therapeutics (acquired by Eli Lilly in July 2025) has VERVE-102 in trials to make the edit permanent. Make the edit once, never take a cholesterol drug again.

Convert APOE4 to APOE2 (heterozygously). The biggest known genetic risk factor for late-onset Alzheimer's is a variant of APOE called E4. About 15--25% of people of European descent carry one copy and have ~3--4x the risk; 2--3% carry two and have ~10--12x. There's a protective variant, E2, that goes the other way. The three differ at only two amino acid positions, so the edit is small. Aim for E2/E3 heterozygosity, not E2/E2; two copies of E2 cause a separate lipid disorder.

Install the FOXO3 longevity haplotype. Specific intronic variants in FOXO3, especially rs2802292, show up in centenarians more often than chance allows, replicated across Japanese-American, European, Chinese, and Ashkenazi cohorts. The variants don't change the protein. They sit in an intron and turn FOXO3 expression up. This is a harder edit: prime editing or small targeted knock-in to a regulatory element. Probably worth it.

Better muscle and bone, together. Two edits that should probably be done as a pair. Follistatin antagonizes myostatin (the brake on muscle growth); early human gene-therapy data is preliminary and confounded, but the biology looks sound. Strong muscles on weak bones is a bad trade, so pair it with the LRP5 G171V variant that runs in some families and produces bone density 3 to 8 standard deviations above the population mean with no obvious cost to carriers.

Cognitive support, ambitiously. Klotho's KL1 domain improves working memory in aged rodents and aged primates. One sketch: a hepatocyte-specific cassette that fuses the KL1 domain to the albumin signal peptide, so the engineered liver quietly secretes KL1 into the bloodstream at a low, steady level. Albumin is the most abundant protein in serum, since the liver makes it constantly, so the production ceiling is enormous and the secretion route is well-trodden. Hold expression to hepatocytes with a tissue-specific promoter. This is the piece that goes past survival. It's the closest thing on the list to a deliberate cognitive upgrade, and the same liver-as-bioreactor pattern works for whatever other circulating factors people find next.

The caveats: every primate study so far is on acute single doses. Whether chronic exposure burns out the receptor or hits something off-target, nobody knows. The dose response is non-monotonic, which means the promoter calibration probably matters more than the cassette does. Both are doable. Neither is done.

// free-lunch alleles

Beyond the headline edits, there's a longer list of variants where one form of the gene confers a real advantage and the other form imposes no measurable cost on its carriers. Once the technology is good enough that the editing itself is a non-event, these become almost free additions. A partial list:

Knock out ABCC11. A single SNP, 538G>A (rs17822931), determines whether earwax is wet or dry and, in the homozygous-A state, also abolishes axillary body odor. In AA homozygotes, the secretion of amino-acid conjugates of human-specific odorants is abolished and the precursors of steroidal body odor are substantially reduced. The variant is common across East Asian populations and has no documented downside.

Truncate SLC30A8. Across about 150,000 individuals sequenced, carriers of rare protein-truncating variants in SLC30A8 had a 65% reduced risk of type 2 diabetes and improved fasting glucose. The variants are loss-of-function. The protein is a pancreatic islet zinc transporter. No known cost to carriers.

Partial loss of SCN9A for a higher pain threshold. Full loss leaves you anosmic and reckless, so the edit aims for partial reduction. Some pain still gets through; chronic pain doesn't dominate the experience.

The DEC2 P385R short-sleep variant, which lets natural carriers run on six hours of sleep with no measurable cognitive cost (Ying-Hui Fu). Speculative as an installed edit, since the carrier cohort is small, but operationally valuable if it transfers: an extra two productive hours a day, every day, for life.

Same pattern, more candidates worth looking at:

The pattern matters more than the list: find rare alleles where carriers do better and don't pay for it, and make them less rare.

What's still unknown is the combinations. Each individual edit has data behind it. A dozen edits gives you sixty-six pairwise interactions, plus higher-order ones. Some will compound, some will cancel, some will fight each other in ways no single-allele study would have caught. Mapping that space is most of what has to happen after layer i.

vii. Rebuild the immune system.

The immune system has two jobs that matter for aging: clearing infections, and patrolling for cells that have started to go wrong: pre-cancerous cells, virus-infected cells, senescent cells leaking inflammation. Both jobs degrade with age. The thymus, where new T cells are trained, starts shrinking in adolescence and is mostly fat by fifty. Hematopoietic stem cells in the bone marrow start producing more inflammatory myeloid cells and fewer lymphoid ones. T cells that have been around for decades become exhausted and stop responding to new threats.

The immune system sits over everything else. When it ages, every other system gets worse: senescent cells stop being cleared, new cancers slip through, vaccines stop taking. Immune decline doesn't sit beside the other problems; it multiplies them.

The move I'd try: replace the patient's hematopoietic stem cells with engineered ones carrying the full set of edits. Bone marrow transplants, which do precisely this, have been performed for more than half a century, longer than any other cell-replacement therapy. The catch: the engineered HSCs would have to come from the patient's own iPSCs (to keep the autologous match) and then be differentiated into actual long-term-engrafting HSCs. That last step, making HSCs from iPSCs that engraft long-term in humans, is still an open problem. Recent work has produced multilineage engraftment in mouse models with forced transcription-factor cocktails (Daley, plus the Lis and Murry labs), but no clinical-grade iPSC-derived HSC product yet matches what you get from bone marrow or cord blood. Half the risk is known (the transplant). Half is open (the manufacturing).

The hard part is not breaking immune tolerance. The immune system distinguishes self from non-self by training. T cells that bind self-antigens get killed during development in the thymus (central tolerance). Outside the thymus, regulatory T cells and signaling proteins like PD-1 and CTLA-4 dampen any auto-reactive cells that escape (peripheral tolerance). Swap in engineered HSCs, and the new immune cells develop in the existing thymus, learn the existing self-antigens, and recognize the body as self.

What can go badly wrong is if the edits produce new surface antigens the immune system has never seen. Now you have an immune system that thinks the modified cells are foreign and attacks them, or worse, gets confused and starts attacking healthy native cells. That's how chronic autoimmune disease starts. The rules to avoid it: keep edits inside the cell whenever possible; if something has to be on the surface, make it look like a fragment of an existing self-protein; don't globally lower the threshold for immune activation; preserve Tregs and the PD-1/CTLA-4 system. Most of this layer is a discipline question: every edit has to pass an "is this likely to break tolerance" check before it ships.

One more piece. An old immune system can be partially rejuvenated by restoring the thymus itself. Treating aged mice with the protein RANKL restores thymic architecture, increases naive T-cell output, and rescues both anti-tumor and vaccine responses to something close to young controls. Pharmacological and genetic interventions to keep the thymus producing naive T cells for longer are both on the table.

glossary HSC -- hematopoietic stem cell; the bone-marrow cell that makes all blood and immune cells.
Thymic involution -- age-related shrinkage of the thymus and its replacement by fat.
Central tolerance -- the deletion of self-reactive T cells in the thymus during development.
Treg -- regulatory T cell; suppresses other T cells to maintain peripheral tolerance.

On the other side of all this, a different sort of person. Someone whose cells repair DNA at bowhead rates and clear their own waste at young-adult rates for life. Whose tissues, when injured, are more likely to regenerate than scar. Whose cells can keep dividing without exhausting their telomere reserve, and whose quality control catches pre-cancerous cells before they get to incubate, with a redundant TP53 system layered behind them as a backstop. Whose LDL stays low without a daily pill; whose Alzheimer's risk has been edited downward at the source; whose bones are denser, whose muscles hold mass into old age, whose pain threshold is set a little higher than most people's. Whose immune system stays trained on new threats instead of getting exhausted on old ones. Who sleeps six hours and feels rested. Who, with a circulating KL1 domain quietly produced by their liver, finds their memory at eighty a little sharper than most people's at fifty. Every piece on that list already exists in some living human, some living animal, a published mouse, or in cells in a dish somewhere.

What's not known is the ceiling. Each piece in this notebook has been shown to do something in a narrow context: a single allele transferred between species, a single cassette in a single mouse line, a single promoter under a single stress. Some of the central proposals (the coupled p53/hTERT circuit especially) haven't been built or tested in mammals yet, as far as I can tell. What happens when you stack dozens of these into one therapeutic is open. Nobody's tried it. That's most of why it's worth trying.

-- -- --   ×   -- -- --
§ 04   What could kill it Three ways this all dies in the first five years

Telomere accidents, delivery, manufacturing: the conventional risks have armies of people already working on them. The three below don't. Each could kill the whole thing on its own.

The body might not let you do all of this at once.

Almost every layer makes more work for the immune system. Killing damaged cells more aggressively (layer iii) generates apoptotic debris that macrophages have to clean up. Replacing cells more often means more divisions and more chances for the immune system to encounter something that looks new. The reprogramming circuit (layer iv), if it fires when it shouldn't, throws off inflammatory signals all by itself. Even the autophagy bump (layer v) increases lysosomal output, which can spill into inflammation if it isn't contained.

Each edit on its own looks tolerable. The body's tolerance for simultaneous low-grade inflammatory pressure is finite, and every layer spends some of it. The question is whether seven layers running at once stays under the threshold where the body flips into a chronic inflammatory state (Franceschi et al. 2000 called it inflammaging). If the answer is no, if you can't raise maintenance without breaking the immune environment, everything has to be redrawn around that ceiling: smaller increments per layer, tighter monitoring throughout.

This is the question I have to ask first. Not because I want the answer to be no, but because if it is, every other decision after that changes.

Selection happens twice.

This is the harder problem, and it's where the Conservation Theory shows up again. The first time selection picked cheap was on the species line, over millions of years, and the result was us. The second time happens inside an engineered body, cell by cell, every time one of them divides.

Every edit you install costs the cell something. Extra DNA repair burns ATP. Stricter quality control kills more neighbors. Telomerase upkeep, sharper autophagy, beefier proteostasis. All of it shows up on the metabolic bill. The engineered cell is, by design, more expensive to run.

Now think about what happens at low-grade DNA damage, the kind that doesn't trip the apoptotic threshold and doesn't do anything dramatic. It happens constantly. And every time it does, the engineered cell spends more energy handling it than a stripped-down cell would. A cell that loses a single transgene by chance, or silences one through normal epigenetic drift, gets a small metabolic discount. It divides slightly faster, dies slightly less, and its descendants out-compete the cells that kept the full set.

This is the same logic that built aging the first time, replayed on a faster clock. Selection prefers cheap. It always has. The engineered population drifts toward whatever subset of edits the daily metabolic budget can support, and the most expensive edits get silenced first, which means the safety-critical ones, the quality-control cassette especially, are the first to go.

Bi-directional promoters and multi-locus redundancy are partial defenses. They buy time. They don't buy forever. This is the same dynamic that makes engineered microbial circuits decay over a few hundred generations. Whether the half-life in human somatic cells is years or decades, nobody knows yet. If it's short, you schedule re-treatment. If it's long, you get away with it. The first-year experiments have to start answering this.

You can't beat the Conservation Theory by ignoring it. Any attempt to install a higher-burn body without confronting the second round of selection loses to it eventually. Selection happens twice, on two different clocks, for the same reason.

Every dial breaks something at the high end.

The first two are the show-stoppers. Either, if it goes wrong, ends the effort. The third is more diffuse but runs across every layer: every protective system you turn up has its own opposite failure if you push it too far. Six of these keep me up at night. Grain of salt, since the numbers don't exist yet:

  • Too much apoptosistissue fragility. Cells get killed before they break, but also before they finish doing their jobs.
  • Too much repaironcogenic tolerance. The cell forgives damage it shouldn't, and carries pre-cancerous lesions forward.
  • Too much regenerationclonal selection. High turnover gives any cell with a slight growth edge more chances to take over.
  • Too much immune surveillanceautoimmunity. A more alert immune system also misfires on self more often.
  • Too much proteostasisgrowth suppression. Aggressive quality control means cells spend more on housekeeping and less on doing their job.
  • Too much stemnessdedifferentiation. Cells start forgetting they're liver cells, or skin cells, or whatever they were supposed to be.

None of this means don't move the dials. It means move them together, in matched amounts, and test for the opposite failure every time. A serious first-year experiment looks at both ends of the curve, not just the one you were hoping for. The plan dies from a sum of small overshoots far more often than from one catastrophic one.

Untested. And tried, failed.

Untested. Stacking longevity-associated edits in one cell or animal: bowhead CIRBP plus elephant TP53 plus naked-mole-rat hyaluronan plus SIRT6. The coupled p53/hTERT circuit under a state-conditional promoter. Partial reprogramming under stress-induced promoters in aged primates. Clinical-grade iPSC-to-engraftable-HSC manufacturing. Decade-scale silencing kinetics of multi-gene constructs in human somatic cells. Pairwise interactions across the layer-vi upgrades. Whether the FOXO3 longevity haplotype is causal in humans or a marker for something else. Chronic KL1 exposure for any meaningful duration, in any primate.

Tried, didn't land. Sirtris, acquired by GSK in 2008 for $720M, closed by 2013, the underlying biochemistry never reproducing cleanly. Bioviva in 2015, Liz Parrish's self-injection of telomerase and follistatin without safety data, a publicity stunt and not evidence of anything. GDF11 as a "young blood factor," debunked in 2015. NMN and NR trials, mostly modest or null in humans. TAME, chronically underfunded for a decade, still not fully enrolled. Senolytic combinations like dasatinib + quercetin, mixed at best in humans. VERVE-101, paused in 2024 over LNP-driven liver toxicity, which is why VERVE-102 has a redesigned lipid nanoparticle.

None of which means longevity gene therapy is hopeless. It means anyone walking in should treat each new claim with the calibration of someone who's watched several rounds of premature optimism go sideways. Every analogous effort that skipped slow in-vitro work paid for it later, which is most of the argument for being patient about the early steps.

§ 05   What I'm doing right now What's actually on my desk this morning

Two things at once. Neither waits.

What I'm spending my days on is genetic-circuit therapeutics for cancer: multi-input constructs that fire inside a tumor and stay quiet in healthy tissue.

The longevity work is the parallel obsession. Cancer and aging share more machinery than most people realize, but I'm not selling one as the path to the other. Cancer kills people now. Aging is what most other dying eventually reduces to. Both are worth doing.

I write the longevity side here mostly so I can argue with my own thinking in public.

§ 06   What comes next What I'd build first, given a lab and a decade

I don't know how long any of this takes. Anyone who tells you they do is selling something. What I can describe is what comes first and what comes after.

What still has to grow up before any of this is real.

The biology isn't what's slowing this down. A handful of engineering tools haven't grown up. Each has serious effort behind it. Progress here matters more than any new discovery in aging biology does.

Vector engineering. Everything is gated here. The next generation of integrating vectors (engineered lentivirus, virus-like particles, and beyond) needs to land before "the full layer-iii build in one infusion" becomes real. ML-designed AAV capsids that explore millions of variants and recover engineered diversification are one direction. Engineered lipid nanoparticles, virus-like particles, and engineered exosomes that bypass capsid immunogenicity are others. Most of the realistic timeline depends on this, not on aging biology.

Synthetic promoter design. Layer iii is solvable because promoters can be engineered to fire only under specific cellular states: p53-active, hypoxic, oxidatively stressed, NF-κB-active. ML models trained on cis-regulatory grammar are now producing synthetic promoters with much sharper on/off ratios than anything natural, and as they get better, the margin for "fire hard in damaged liver, stay quiet in stem cells" widens, and the calibration that has to happen in the clinic shrinks.

Prime editing fidelity, multiplexed CRISPR, base editing reach. The layer-vi upgrade list is basically a bet on the editing tools getting more precise and more parallel. Base editors that could in principle install APOE4 → APOE2 without a double-strand break exist in cells and animals, though no clinical trial of that specific edit has been published yet. Prime editors that can hit FOXO3's intronic edits (a regulatory edit, harder than a coding-sequence edit) are improving fast. Multiplexed sgRNA delivery, off-target prediction, and DNA-damage-response load are all active engineering problems. Dozens of edits per dose isn't far off. One per dose was a decade ago.

iPSC→engraftable-HSC manufacturing. Layer vii waits on this. The biology (forced transcription-factor reprogramming to push iPSC-derived hematopoietic progenitors toward true long-term-engrafting HSCs) is being worked on in Daley, Lis, and Murry's labs, among others. The clinical-grade version is a manufacturing problem more than a discovery problem at this point. When it's solved, layer vii becomes a real option. Until then it isn't.

Off-target prediction. The hard part of multi-edit work usually isn't whether each individual edit works on its own, but whether the combined off-target footprint stays inside an acceptable safety envelope. ML models that predict cut-site biases, chromatin accessibility, and combinatorial off-target effects across dozens of edits per cell are improving fast, and they're roughly what makes the layer-vi multiplexing realistic.

Where AI changes the math.

A lot of this is search over very large possibility sets: envelope sequence, promoter sequence, guide RNA selection, codon usage, off-target prediction, combinatorial edit interactions. Models that have absorbed the biological literature and can sweep those sets fast are the difference between a decade of trial-and-error and a month of in-silico design plus targeted validation. The same pattern that compressed antibody and small-molecule design applies here, with the catch that gene-therapy design has a much smaller historical training set than small-molecule chemistry. That gap will close, but slowly.

The other thing AI changes is how much science a small team can do. Work that today needs thirty people across envelope design, promoter calibration, multi-edit guide design, off-target screening, and primary-cell validation will, plausibly, be doable by far fewer over the next decade. Not a reason to wait; a reason to design a company around the assumption that whatever takes ten people today will take less before long.

The order, roughly.

Run the early in-vitro experiments to answer the threshold, coupling, and silencing questions before anything goes near an animal: whether the apoptotic cassette kills healthy cells at acceptable rates, whether the bi-directional coupling holds under oncogene pressure, whether multi-locus integration survives a hundred-plus cell doublings. Then animals: HSC replacement proof-of-concept in chimeric mice, pairwise interaction maps for the upgrade-list edits, a time-course on KL1 to find out whether chronic exposure burns out the receptor, and iPSC manufacturing QC, including mitochondrial-DNA mutation screening, which nothing in this plan currently addresses well and which has to be solved before any human work.

The realistic first-in-human is probably patients already getting myeloablative bone marrow transplants for blood cancers: engineered HSCs versus standard, with endpoints on genomic stability, secondary cancer rates, and immune competence. Not a longevity trial. A bone marrow transplant with better cells, run through the standard biologics path, with the CAR-T precedent doing most of the regulatory work. If it lands, it's the first proof the edits do anything in human tissue on human timescales. Everything afterward is expansion.

The reason I'm writing this down at twenty-something instead of waiting until I have the lab: what decides whether any of this gets built isn't biology, it's whether enough of the right people start working on it ten years earlier than they would have. People die every day from things that, given enough time and the right work, wouldn't have to be fatal.

DJ PAOLETTI
// things I like and things I think living draft · 2026

Stuff I like & stuff I think.

Not in any particular order. Updated whenever.

§ 01

Anime & TV

Stuff I'd actually rewatch, or already have.

01
Vinland Saga
A story about a kid who wants revenge and slowly figures out his purpose in life. It hits.
02
Mob Psycho 100
The fights are wild. It's one of those shows where there's something genuinely moving sitting beneath the veneer of an action-comedy.
03
My Hero Academia
I watched this when I was little, and it earns a spot on the list for that reason as much as any other.
04
Severance
No show on TV looks like Severance. Half the appeal is just watching it: I've never seen a concept like that on screen.
05
Naruto
The classic. The Pain arc still hits. I won't complain about the filler.
06
Pantheon
The most underrated sci-fi show ever made.
07
Parks & Rec / The Office // comfort tier
Background TV for when I want the room to feel populated.
08
Any 12-episode isekai // work fuel
Playing on the second monitor while I'm deep in something on the first. Doesn't really matter which one.
§ 02

Movies

Things I've watched more than twice.

01
Memento
The first time blew my mind. Still my favorite Nolan.
02
How to Train Your Dragon // all of them
The flying scenes. The music. Each one made me cry.
03
Thor: Ragnarok
The only Marvel movie that figured out you can just be fun the whole time. Taika carried; nothing he's done for Marvel since has worked as well.
04
Baby Driver
A movie cut to its soundtrack instead of the other way around. The opening car chase is incredible.
§ 03

Music

What I actually listen to.

01
Rihanna
Always.
02
ABBA
Genuinely incredible. "The Winner Takes It All" is a masterpiece.
03
Coldplay
I'm not sorry about it.
04
Lady Gaga
Obvious reasons.
05
Adele
Crying-in-the-car music.
06
Whatever's loud enough to drown out my thoughts
Late at night, mostly. See § 05.
§ 04

Places

The city matters less to me than what's around it.

Where I live matters less to me than what's near me. Two things make a place worth being in at this stage: density of relevant talent, and proximity to the people who write the checks. The rest is decoration. Greater Boston clears the bar. The Bay Area clears the bar. A handful of other cities do too. Most don't. People who optimize for views or weather over those two things tend to underrate how much of who you become is decided by who you're around. I can play tennis and do science most places. I can only maximize my odds of doing something meaningful in a few of them.
§ 05

Half-formed ideas

Things I think but haven't fully argued for. Probably some are wrong.

01
Most "personality" is just energy level
Watch how someone acts when they've slept ten hours and eaten a real meal vs. when they haven't. The former is who they actually are. Give people the benefit of the doubt.
02
Anime is doing things movies and TV currently can't
The medium is in a quiet golden age and most people outside of it haven't clocked it yet.
03
A walk at night with loud music is one of the best ways to collect yourself
No phone in your hand, no destination, headphones at a volume that's probably bad for your hearing.
04
The fastest way to read someone's pretentiousness level is "movies" vs "films"
It works almost every time. "I love movies" and "I love films" are not the same sentence.
05
The best founders I've met read fiction
Not exclusively, but they read it. Holding a long imagined world in your head and watching it move is closer to what building a company actually feels like than most business books admit.
06
Most people give up too early on things they'd be great at
The dip between "I'm new at this" and "I'm good at this" is much longer than people are told. Anyone who is genuinely great at something spent more time looking dumb than is comfortable.
§ 06

Predictions I'll probably regret

Things I think will happen. Some will be wrong. That's fine.

01
Healthy life expectancy will start moving
Healthspan, not lifespan. Years a typical person stays functional, not just alive. The first real bumps come from boring stuff: GLP-1s, better cardiovascular drugs, maybe a senolytic that finally works. The interesting bumps come later.
02
Most of the people building important things in 2040 are in school right now
Credentials don't decide who runs the next wave of bio companies. Getting your hands on real problems early does. Most of the people who will matter are unknown.
03
The first durable gene therapy for something people don't have to have will be cardiovascular
PCSK9, probably. Cheap. One dose. Decades of statin-equivalent benefit. Once that lands, the regulatory pathway for "edit-to-prevent" stops being theoretical and people start asking what else qualifies.
04
AI will compress the cost of running a biotech by an order of magnitude
Not by replacing scientists. By making each one effective enough that you need many fewer of them per program. The first companies built this way will look very strange compared to what biotech looks like now.
05
"Aging as an indication" gets a real foothold this decade
Not necessarily through TAME. Probably through a multimorbidity endpoint that gets approved for something adjacent and then becomes the template. Once the first one lands, the rest follow faster.

"This list is wrong in places. That's the point of writing it down."

// the working theory