Summary

Kimmel’s core argument is that aging sits in the category of problems evolution did not strongly optimize: the baseline hazard rate through primate history was so high that relatively few individuals survived to the age where longevity alleles could be selected, so “the amount of gradient signal flowing back to the genome then is not as high as one might intuitively think.” Comparatively easy to improve — which is why single-gene drugs and antibiotics deliver massive benefits — but it also means aging is not monocausal and the first medicines will add years, not immortality.
The tractable bottleneck is search, not mechanism. With roughly 2,000 transcription factors and combinations of one to six, the space is about 10^16 combinations, and brute-forcing it “would need many orders of magnitude more single-cell sequencing than the entire world has done to date cumulatively.” NewLimit sparsely samples, trains a model on perturbation→transcriptome, then generates candidate combinations in silico — pretraining plus a value head, in Dwarkesh’s framing.
Kimmel is hopeful that payload and dose will not be limiting: working combos found so far are “somewhere between one and five” factors, and clinical flu/COVID vaccines already deliver 20 unique transcripts, making a few TFs “seemingly quotidian.” TFs also sit near the bottom of the expression rank-order, so low doses suffice. Durability data today is “several weeks after a dose” — Kimmel explicitly refuses to overstate the one-shot-lasts-decades upper bound.
Delivery is currently ahead of the biology, not behind it: “They’re still winning the race against us right now” — there are nucleic-acid medicines and no reprogramming medicines. LNPs and AAVs are enough work for decades, but Kimmel’s controversial call is that neither is probably how medicines get delivered in 2100; engineered T/B cells that engraft, sense an AND gate and release payload could be the answer, because “you’ve got billions of base pairs to play with.”
The reason biotech never got a scaling-law narrative: Eroom’s Law delivers the same logarithmic curve as ML but with neither of the two properties that make ML investable — super-exponential returns to scale and cross-task transfer. “Whereas in biotech, modulo NewLimit’s new round, it has driven down valuations.” The fix on the revenue side is TAM: “All of us will one day get sick and die. So arguably, the TAM for any really successful medicine could be everybody on planet Earth.”
The claimed moat is data, not molecules: NewLimit says it has more combinatorial-TF-overexpression data than anyone “full stop,” and more cell-age reprogramming data than the rest of the world combined, all in human cells rather than “cancer cell lines which have 200 chromosomes. Is that human? I don’t know.” Analogy offered: it’s 2018 Cursor Tab, not a frontier lab — a deliberately carved subregion of the virtual-cell problem, vertically integrated because no internet-scale biological corpus exists.
Reimbursement is the underpriced risk for durable therapies: US patients churn insurers every three to four years, so a drug whose savings land in year five means “no insurer is technically economically incentivized to cover that.” Kimmel’s route is pay-for-performance plus direct-to-consumer (LillyDirect as the template), and he argues net healthcare spend falls — a third of Medicare cost lands in the final year of life, and pharma is “the one place where… you’re able to get more benefit per dollar.”

Deep dive

1. Evolution did not strongly optimize longevity — so it should be comparatively easy to improve

Kimmel decomposes the puzzle into three questions: was there positive selective pressure for longer healthy life, was there anti-selective pressure, and what were the constraints on the optimizer? His framing is explicitly ML-shaped — “if we think about the genome as a set of parameters and the optimizer is natural selection, then you’ve got some constraints on how that actually works.”
The positive-selection leg dies on the hazard rate, “simply, ‘What is the likelihood you’re going to die on any given day?’” — integrating disease, “getting eaten by a tiger,” and “scraping your foot on a rock and getting an infection.” Best evidence says it was very high, so relatively few individuals ever reached the ages where aging is the binding constraint, and “the amount of gradient signal flowing back to the genome then is not as high as one might intuitively think.”
The payoff of the whole argument is a research heuristic he applies everywhere in biology: “Did evolution spend a lot of time optimizing this? If yes, my job is going to be insanely hard. If not, potentially there are some low-hanging fruit.” Aging lands in the second bucket — and the existence of crude modern drugs that “target a single gene in the genome and turn it off everywhere at the same time” and still deliver massive benefit is his supporting evidence.

2. The same hazard-rate logic implies intelligence was under-optimized too

Dwarkesh extends the argument to AI forecasting: high hazard rates cap adolescence length, because “if you’re just hanging out learning stuff for 50 years, you’re just going to die before you get to have kids yourself” — and if long adolescence is unaffordable, there’s little point selecting for a bigger brain. Conclusion: “maybe intelligence is easier than we think, and there’s a bunch of contingent reasons evolution didn’t churn as hard on this variable as it could have.” Kimmel: “I entirely agree with that particular thesis.”
Dwarkesh’s aside on how strange the search itself is — “It’s a long-horizon RL problem, a 20-year horizon length, and then there’s a scalar value of how many kids you have” — with the observation that given how hard RL is on hour-long tasks, “it’s surprising that any signal propagates across a 20-year horizon.”

3. Kimmel’s pet hypothesis: fluid intelligence peaks at 25–30 because that’s where the population was

The puzzle he takes seriously: great mathematical discoveries cluster “roughly before 30” (he flags he may have the exact age wrong). Societal explanations — becoming staid, teachers restricting your thinking — fail his cross-cultural test: “that’s true across centuries? Is that true across many different unique cultures around the world?… That seems unlikely to me.”
His simpler explanation is demographic: fluid intelligence is maximized at the age where selection pressure was densest. “If you had to pick an age at which fluid intelligence was selected most strongly for, it’s probably around 25 or 30. That’s probably about the age of the adults in the large populations that were being selected for during most of evolution.”
Dwarkesh sharpens it to the annus mirabilis — Newton with “optics, gravity, calculus at 21” — then offers Alexander von Humboldt, whose single South American expedition up Chimborazo, observing ecological layers repeating across latitude and altitude, “was the basis of his entire career.” The tell for how famous he was: when you see something named Humboldt, “it’s this one guy.”

4. Aging as a length regularizer — the kin-selection case against long life

Kimmel hedges this leg hard: “I don’t know how strong some of the mathematical models that people put together here are. You can find people using the same idea to argue for and against.” On a selfish-gene view, extending lifespan without eliminating decline creates “this nasty regularization term.”
The mechanism: a longer-lived but less fit individual contributes fewer net calories per marginal year than “two 20 year olds that follow behind them,” so a population demographically laden with aged individuals is net negative for the genome’s proliferation. “A genome should optimize for turnover and population size at max fitness.”
Dwarkesh’s reframing is the line that sticks: “I love this idea of aging as a length regularizer” — the biological analogue of penalizing chain-of-thought length during training, with lifetime calorie consumption as the penalty term.

5. Even granting positive selection, optimizer constraints also limit longevity

Mutation rate bounds step size in both directions: “if you dial your mutation rate up too high, you probably get a bunch of cancers, so you’re selected against. If you have it too low, you can’t really adapt to anything.” Population size bounds how many variants screen in parallel.
And one dominant source of selection is something else entirely — infectious disease, which “is actually what shaped a lot of our population demographics.” The weighted-loss framing is the point: even if longevity is in the objective, “if you simply imagine that the lambdas are dialed toward infectious disease resilience more effectively, then you can construct an argument for yourself.”
Stacking all three, the counterfactual where humans are longevity-optimal requires “an incredibly contingent scenario” — positive selection present, negative selection absent, and substantial evolutionary pressure allocated to an incredibly hard problem.

6. Why humans never evolved their own antibiotics — the Red Queen and raw parallelism

Dwarkesh’s question stops him: “Why didn’t humans evolve their own antibiotics?” Kimmel: “It’s an excellent question that I haven’t heard posed before.” There’s no prima facie barrier — “you couldn’t imagine encoding an antibiotic cassette into a mammalian genome” is not obviously true.
The blocker is competitive dynamics — the Red Queen hypothesis, from Through the Looking Glass, where “the Red Queen is running really fast just to stay in place.” Bacteria and fungi hold the advantage because of “trillions of copies of the genome. Massive analog parallel computation,” plus tolerance for high mutation rates as prokaryotes. For metazoans, one over-mutated cell “might turn into a cancer and eventually kill off the organism.”
The tradeable implication Dwarkesh draws out: there should be “millions of ’naive antibiotics’” abandoned across evolutionary history. Kimmel — flagging “I’m going a bit beyond my own knowledge here, but my strong hypothesis would be yes” — points to ancient CRISPR spacers in bacteria as a readable record of “what the warfare was like.”

7. TRIM5alpha: we traded away HIV resistance, and duplication is how evolution affords multi-step edits

The best-specimen example: TRIM5alpha binds an endogenous retrovirus that no longer exists, fitting “around the capsid of the virus like a baseball in a glove.” Tracing it back through monkeys, a previous iteration inhibited SIV — Old World monkeys can’t get SIV; New World monkeys and humans can. The genome apparently gave up HIV-like restriction to fight a retrovirus that then went extinct, “reasons unknown, no one knows why.” A couple of edits in human cells today “can edit it back to actually restrict HIV dramatically.”
Dwarkesh pushes on plausibility — with per-base mutation rates around one in a billion, how does a multi-base binding sequence ever assemble? Kimmel’s answer is gene duplication: intermediate edits that break a gene are net negative for fitness, so evolution can’t traverse them, but with a backup copy “those first two edits are totally tolerated.” The TRIM5alpha edit count is “in the tens,” not kilobase rearrangements.
The practical tell he offers anyone: look at gene names. “Very often you’ll see something where it’s like gene one, gene two, gene three” — sometimes coincidence of discovery, often homology. So evolution “doesn’t have to start from scratch… let’s do a copy paste on that and then iterate and fine-tune on those parameters.”

8. Aging is not monocausal — and the first medicines will only fix part of it

Dwarkesh spots that the evolutionary argument constrains the product: if evolution declined to fix aging because it isn’t traceable to one cause, then partial fixes are the default, contra longevity proponents who say “there’s going to be a source that explains all of aging and we’ll get it.”
Kimmel accepts it flatly: “I don’t think that there is a single monocausal explanation for aging… it’s not that there is some upstream ‘bad gene X’ and all we have to do is turn that off and suddenly aging is solved.” He has dedicated his career to epigenetics because he thinks it explains a lot, not everything.
The forecast, stated as a hedge and a categorical at once: “There’s not going to be a singular magic pill.” Rather, medicines that “add multiple healthy years to your life, years you can’t otherwise get back” — while “you are still going to experience, for the first medicine, some amount of decline over time.”

9. Transcription factors as the orchestra conductors — and the assays that keep it honest

The mechanism: TFs “don’t perform many functions directly themselves, but they bind specific pieces of DNA and then they tell which genes to turn on, which genes to turn off,” writing the epigenetic marks that answer why “your eyeball and your kidney have the same code, and yet they’re performing different functions.” That epigenome can degrade with age, so cells can’t invoke the right programs at the right times.
Asked whether the intervention is clean, he’s blunt: “How I wish it were straightforward. No, it’s very likely” deleterious side effects — each TF binds hundreds to thousands of genomic sites, and “there are no guarantees that aging actually involves moving perfectly along any of the vectors in this particular basis set.”
So NewLimit runs two layers of measurement: a “looks like” assay (“Can I make an old cell look like a young one based on the genes it’s using?”) and function — can a hepatocyte process metabolites, alcohol and caffeine; can a T cell respond to antigens. Plus explicit pathology checks, because Yamanaka-style reprogramming also changes cell type, and in a body that “would probably cause a type of tumor called a teratoma.”

10. Yamanaka’s problem had two key advantages

Dwarkesh’s challenge is the sharpest of the episode: Yamanaka narrowed 24 embryonic TFs down to four by elimination, “That doesn’t require any fancy AI models. Why can’t we do the same things” for age? Kimmel’s preface: “Most of science is problem selection. You don’t actually get better at pipetting or running experiments after a certain age, but you do get better at picking what to do.”
Feature one was a trivial success criterion — fibroblasts (literally “cells that stick to glass”) converting to embryonic stem cells, read out by a reporter that turns blue. Feature two was amplification: original efficiency was “a basis point or a tenth of a basis point, so 0.01%, 0.001%,” but over ~30 days the rare successes proliferate into colonies you find “by holding the dish up to the light.”
Aging has neither. “An old liver cell and a young liver cell, prima facie, actually look pretty darn similar,” with no single gene serving as a binary classifier — single-cell genomics was the enabling technology, since “I don’t think our approach would really have been possible until it emerged.” And success doesn’t amplify, so “in some ways, the bar for a medicine is higher than what Yamanaka achieved.” Dwarkesh’s deadpan workaround — “Give the young cells cancer… Just make the old ones cancer, and then they’ll grow” — earns “Dwarkesh, you’ve solved it for me.”

11. 10^16 combinations are why models are needed, and why TFs are the right handle

The arithmetic: somewhere between 1,000 and 2,000 TFs (“developmental biologists love to argue about this over beer, but let’s call it 2000”), combinations of one to six, gives about 10^16 possibilities — exhaustively screening “would need many orders of magnitude more single-cell sequencing than the entire world has done to date cumulatively.” Hence sparse sampling, learned interaction terms, then in-silico generation toward “some target destination in state space.”
His contention that TFs behave like a usable basis set rests on development: hundreds of cell types are specified by TF groups that “are actually pretty similar to one another,” so evolution can “swap one TF in or swap one TF out of a combination and get pretty different effects.” Because mutations are small and random, biology is forced into a regime where small edits yield large phenotypic changes — “a relatively favorable regime for generic, gradient-like optimizers,” which he likens to evolution strategies rather than true gradients.
The analogy he warns requires “a very special audience”: “TFs are like the queries. The genome sequences they bind to are like the keys. Genes are like the values.” Change one embedding vector and outputs shift dramatically. Elsewhere he calls TFs “evolution’s levers upon the broader architecture of the genome”; Dwarkesh notes Trenton Bricken’s brain-attention work and Eddie Chang’s neuropixels evidence for positional-encoding-like firing across sentences.

12. Why there aren’t more TF drugs: both classical modalities are the wrong size

Most existing drugs already act through TFs — receptor blockers, cytokine inhibitors, signaling-pathway agents all terminate in a TF being switched on or off. “We’re kind of taking these crazy bank shots because we can’t hit the TFs directly.”
The physical reason is a Goldilocks trap: small molecules are small enough to cross the membrane but “TF’s binding DNA is a pretty darn big surface” — small molecules “can get all the way into the nucleus, but they can’t do much once they’re there.” Recombinant proteins and antibodies are potent enough but “too big to get through the cell membrane.”
What changed is nucleic-acid medicine: lipid nanoparticles — “You wrap them in a fat bubble” that fuses and drops mRNA into the cytosol — plus viral vectors mean “we’ve only very recently actually gotten the tools we need to start addressing transcription factors as first-class targets rather than treating them as maybe some ancillary third-order thing.”
On whether the natural TF set is even the right search space: probably yes as a prior, but Super-SOX (Sergiy Velychko’s mutated SOX2, outperforming the canonical Oct-4, Sox2, KLF4, Myc) suggests otherwise — “iPSC reprogramming never happens in nature, so there’s no reason to necessarily believe that the natural TFs are optimal.” His 2100 guess: “synthetic genes that have never existed.”

13. Delivery is currently ahead of the biology — and by 2100 it probably won’t be LNPs or viruses

Asked whether delivery lags his understanding of aging, Kimmel corrects the premise: “Just to give the delivery folks credit, they’re currently ahead. There are currently no reprogramming medicines for aging, and there are medicines that deliver nucleic acids. They’re still winning the race against us right now, but to your point, I hope the lines cross.”
Today’s two modalities both have ceilings. AAVs are “like a very small delivery truck,” and sequence engineering only adds a NOT gate — “you can start broad with your delivery vector and then use sequence to narrow down to make it more specific, but not the other way around.” Viruses are irreducibly immunogenic; LNPs face physical constraints escaping the bloodstream without fusing en route.
His stated controversial opinion — offered after Dwarkesh’s jab, “You have just one? You’re trying to solve aging and you have only one?” — is that delivery gets solved “the way that our own genome solved delivery”: engineered T and B cells that patrol, run “an AND gate logic,” and release payload locally. They’d engraft and persist for years, dosing you when your body dictates rather than when you see a physician, with “billions of base pairs to play with in terms of encoding all your logic.” “If I could clone myself and work on an even riskier endeavor, that’s probably what I would do.”
CAR-T is the partial precedent — it engineers the detection half and leaves the payload alone. The complementarity argument: virus people target immune-privileged compartments (knees, shoulders, eye, brain, probably ear) precisely because their drugs are immunogenic, so “the shadow of all the diseases you can’t address with viruses is what you can address with cells.”

14. Partial fixes cascade — one tissue’s rejuvenation shows up body-wide

Dwarkesh’s worry is a patchwork body — great liver, everything else aging normally. Kimmel: you don’t get “this strange, Frankensteinian benefit in health in some aspects and lack of benefit entirely in others,” because rescuing one cell type produces knock-on benefits.
The controlled evidence is transplants: old humans receiving young livers show reduced risk of several other diseases and better overall survival, not merely tolerance for fatty food and drink — because liver and adipose are endocrine organs “sending out signals to many other places in your body.” He credits Frederick Appelbaum’s book (Appelbaum trained with Don Thomas, who invented human bone marrow transplants) for cases where HSC replacement incidentally cured an unrelated disease.
The inverse exists too: break TFAM, a mitochondrial transcription factor, in one specific subset of T cells in mice and “you dramatically shorten the lifespan.” Ozempic is offered as the existence proof in the other direction — GLP-1/GIP-1 incretin mimetics with benefits in weight, cardiovascular disease, “possibly for addictive behavior, and maybe even preventing neurodegeneration.” “If someone told you” all that from one molecule, “you would have told them they were crazy.”
On dose and durability he is careful: one-time dosing is possible “in principle,” and the existence proof for decade-scale persistence of epigenetic states is that “my tongue doesn’t spontaneously turn into a kidney” — plus Luke Gilbert’s single-locus edits surviving 400-plus divisions, and similar editors lasting years in monkeys. But: “We don’t have data like that today. I don’t want to overstate. We do have data that these positive effects can last several weeks after a dose.” Not everything is cellular, either — sagging skin traces to elastin fibers that only polymerize during development, so the fix likely requires programming cells to “extra-physiological” states.

15. Eroom’s Law is a scaling law that ran the wrong way — and biotech lacks ML’s investment advantages

Eroom’s Law, a portmanteau from his friend Jack Scannell, inverts Moore: a consistent decline in new molecular entities per billion dollars invested, starting in the 1950s and persisting “through many different technological transitions.” Structurally it looks like ML scaling — “You throw in more inputs and you get consistently diminishing outputs.”
Dwarkesh’s contrast is the section’s spine: in ML the same curve raised exponentially more capital, “Whereas in biotech, modulo NewLimit’s new round, it has driven down valuations, driven down excitement and energy.” His diagnosis is generality — one model absorbs $100M, then $1B, then $10B, versus “we made money on this drug and now we’re going to use that money to invest in 10 different drugs in 10 different bespoke ways.”
Kimmel offers two possible explanations for that difference. Returns may not scale super-exponentially — biotech’s outputs “haven’t necessarily scaled in their potential revenue,” so cost increases aren’t counterbalanced by ROI. And cross-task transfer is limited: treating disease X “doesn’t necessarily engender you to be able to then treat ‘disease Y’ more readily,” because the durable expertise sits in making molecules, which “isn’t actually reducing the largest risk in the process.”
His fix on the revenue side is TAM, and he says it categorically: “All of us will one day get sick and die. So arguably, the TAM for any really successful medicine could be everybody on planet Earth” — an argument against the industry’s drift from broad categories toward “narrower and narrower genetically-defined diseases that have small patient populations.”

16. The hard part is knowing what to target — and Dwarkesh pushes back on that

Kimmel’s claim: “Figuring out what to make an antibody to target is the hard thing about drug discovery.” Dwarkesh doesn’t buy it cleanly, citing the small-molecule Goldilocks problem: “There it seems like getting the hook is the big problem.”
Kimmel’s concession and rebuttal: yes, if you bind yourself to small molecules, many targets are undruggable — but other modalities exist. His thought experiment: lock 10 smart drug developers in a room and have them write down high-conviction target–disease pairs blocked only by a missing molecular hook. “It’s a fairly small list. It would probably fit on a single page” — against an astronomical number of possible indications and gene combinations.
The falsification test he offers: if lack of a hook were the only barrier, transgenic animals — where you can turn on arbitrary gene groups in arbitrary cells at any time and dose — should let us cure these diseases in the best animal models. “For the majority of pathologies, we just don’t have many of those examples.”

17. The virtual cell as pretraining — and why NewLimit is Cursor Tab, not a frontier lab

The virtual cell (“a very nebulous idea, sometimes numinous”) is the candidate general model: perturb, measure transcriptomes, learn the map, then search in silico for interventions that shift diseased cells toward healthy. NewLimit does both heads — one predicting every gene the cell expresses, “an objective rather than a value judgment on the cell,” the other a value judgment that a transcriptome “looks like a younger cell.” Inputs include TF representations derived from protein foundation models, so “the model’s starting from a pretty smart place.”
Dwarkesh’s mapping, which Kimmel accepts with a caveat: pretraining is learning how cells work, then “there’s another afterward layer of these value judgments.” “That makes me more optimistic on this. LLMs work and RL works.” Kimmel: “the conceptual analogy is very apt. We don’t actually use RL at the moment, so I don’t want to overstate the level of sophistication we’ve got.” On who writes the labels, Dwarkesh jokes about “labelers in Nigeria clicking different pictures of cells”; Kimmel: “It’s more like developmental biologists locked in a room, as my friend Cole Trapnell would say.”
Dwarkesh’s structural critique: this is Cursor in 2018 building its own LLM — “It seems like you’re combining two different layers of the stack.” Kimmel’s answer accepts the frame and shrinks it: “imagine we’re trying to, in 2018, create Cursor Tab, but we’re not trying to create a full LLM.” NewLimit works on only a few cell types — the ones with credible delivery today — because “if we solve the problem of what TFs to use, we can make a medicine pretty quickly.”
Why vertical integration is forced rather than chosen: LLM data was “a common good… produced as a function of the internet,” while biological data isn’t. “Imagine we’re in the early 1980s and we are just now thinking about trying to create some of the first web pages.” So they build the high-quality corpus themselves — “the Wikipedia that you might train your overly analogized- LLM on.”

18. Perturb-seq existed since 2016 — why it took time to scale

Dwarkesh’s skeptical prompt: three near-simultaneous 2016 papers (Ido Amit’s lab at Weizmann, Aviv Regev’s at the Broad with Atray Dixit, Jonathan Weissman’s at UCSF with Britt Adamson) — “We’re still waiting, I guess, for the big breakthroughs it’s supposed to cause.”
The unglamorous answer: readout cost fell from dollars per cell to cents and fractions of cents, and early barcode detection worked only “about 50% of the time.” His analogy makes the stakes concrete: “imagine you hired someone who every other tube labels it wrong… you’ve just randomized all your data labels.”
Combinatorial perturbation compounds it multiplicatively — at 50% detection, correctly labeled cells scale as 1/2^n, so “very quickly more of your data is mislabeled than is labeled.” Six or seven years ago the reagent makers published the first million-cell dataset as a proof of concept only they could run; “Now two scientists in our labs can generate that in an afternoon,” at millions of cells per day.

19. Who pays for a durable medicine — insurer churn, pay-for-performance, and DTC

The gray-market question (Chinese GLP-1 peptides, isomorphic molecules) he mostly declines — “that’s IP enforcement at a geostrategic level that I’m not qualified to speak to” — but argues the payer system protects incumbents: given a “sketchy vial off of some website from some company in Shenzhen” versus a low co-pay for real Tirzepatide, most patients choose Tirzepatide. And a candid caveat: “You and I probably live in a milieu of people who are much more comfortable with ordering the vials from Shenzhen than most people might be.”
The real structural problem is duration versus churn: the average person in the US system churns insurers every three to four years, so for a medicine whose savings arrive in year five, “no insurer is technically economically incentivized to cover that.” His candidate fix is pay-for-performance — reimburse a tenth per year of a ten-year drug, contingent on measured effect — while conceding the measurement problem is “a big challenge in this industry” outside gene therapies where you can simply assay expression.
His directional call is direct-to-consumer for health-preserving medicines, with LillyDirect as the live template: prescription straight to “Lilly, the source of the good stuff,” bypassing the pharmacy–PBM chain “and not involve some intermediary compounder in the middle that might not even make your molecules properly.” Once patients feel the benefit daily, “that model will start to dominate,” and financing over time becomes as ordinary as any large purchase.

20. De-aging should bend healthcare spend down, and pharma is the only deflationary piece

Dwarkesh’s framing of why this matters: healthcare is already 20% of GDP and growing, and “the overwhelming majority of this is going towards administering treatments that have already been invented” rather than inventing new ones. Kimmel confirms his order of magnitude: “drugs are roughly 7% of healthcare spend.”
Kimmel expects net spend to fall, with caveats that Baumol’s cost disease and disintermediation of customer from provider are “a larger economic problem” biotech can’t solve alone. The load-bearing statistic: “something like a third of all Medicare costs are spent in the final year of life, which is shocking when you realize that the average person on Medicare is probably a decade-plus covered by it.” Preventing even a couple of inpatient visits shifts burden from administration to pharmaceuticals.
The reason that shift is deflationary is the generic bargain: “The pharmaceutical system is the only piece of healthcare where technology has made us more efficient.” Hence his closing test — asked when you’d want to be born as a patient, “you always want to be born as close to today as possible,” because per dollar “you can access more pharmaceutical technology today than has ever been possible in history.”

21. Big pharma has become a venture firm, and an oligopsony of buyers

Asked what Lilly’s or Pfizer’s head of R&D thinks about general-purpose platforms, Kimmel first narrows his own moat claim: more data than anyone “full stop” on combinatorial TF overexpression, and “very, very confident we have more data than anyone looking at trying to reprogram a cell’s age” — but other groups hold large single-cell perturbation datasets. His differentiation is substrate: human cells “with the right number of chromosomes, whereas it’s very common to do things in cancer cell lines which have 200 chromosomes. Is that human? I don’t know.”
The market-structure read: “you can think about some of the modern pharmas a bit like venture capital firms” — external innovation divisions functioning as corp dev, letting nimble biotechs explore pioneer ideas and partnering later. From memory and flagged as approximate, “something like 70% of molecules approved in a given year come from originally small biotechs rather than large pharmas, even though you look at the actual dollars of R&D spend on the balance sheet and it’s largely in big pharma” — much of that spend being trials, which biotechs partner out.
The clearing price sits in one place: biotech startups selling into “an oligopsony of pharmas,” where “there’s a very liquid market for the phase one, phase two assets.” Not universal — Roche, which “bought Genentech back in 2013,” has R&D run by Aviv Regev, “one of the scientists I admire most in the world,” who helped invent this very technology and runs a large group doing it in-house.
Disclosure, from Dwarkesh at the close: “I am a small angel investor in NewLimit now, but that did not influence the decision to have Jacob on.”

E204 | The Growing Pains of Sportswear: lululemon’s Slump and Alo Yoga’s Rise Perfecting the Investing Craft with Caro-Kann’s Artem Fokin