Overview

In this deeply technical conversation, Jacob Kimmel, president and co-founder of NewLimit, makes the case that evolution never optimized for human longevity—and that this failure creates an opening for intervention. Drawing on evolutionary biology, machine learning analogies, and the emerging science of epigenetic reprogramming, Kimmel argues that aging is not a single problem but a collection of degradations that can be addressed by resetting cells' epigenetic state. The conversation ranges from why humans didn't evolve their own antibiotics to the economics of drug discovery, with Kimmel offering a vision of a future where engineered cells, not pills or injections, deliver therapies throughout the body. The tone is rigorous but accessible, with both host Dwarkesh Patel and Kimmel freely drawing parallels between biological systems and artificial intelligence.

---

0:00Why Evolution Didn't Optimize for Longevity

Kimmel opens by dismantling the intuitive assumption that evolution should have selected for longer lifespans. He identifies three distinct reasons evolution left longevity on the table. The first is the baseline hazard rate—the likelihood of dying on any given day from any cause, including predation, infection, and accident. During most of human and primate evolution, this hazard rate was extremely high. Even if aging were completely absent, most individuals would die from other causes before reaching ages where aging becomes the primary limitation. This means there was very little "gradient signal" flowing back to the genome to select for longevity-promoting alleles. The number of individuals surviving to old age was simply too small for natural selection to act upon.

The second reason involves what Kimmel calls "kin selection" and a "length regularizer" on lifespan. From a selfish gene perspective, the genome optimizes for its own propagation, not for any individual organism's longevity. If an older individual lives longer but is less fit—consuming calories without contributing proportionally to the gene pool—then allowing that individual to die and be replaced by two younger, fitter individuals is actually better for the genome's proliferation. Aging, in this view, functions like a regularization term in machine learning: it prevents the population from becoming demographically laden with aged individuals who are net negative for the genome's spread.

The third reason is pure optimization constraint. The genome's mutation rate bounds the step size of evolutionary updates—too high and you get cancer, too low and you can't adapt. Population size limits how many variants can be screened in parallel. And most of the genome's optimization budget has been spent on fighting infectious disease, which has been the dominant selective pressure throughout human history. Even if positive selection for longevity existed and negative selection were absent, the weight on longevity in the genome's "loss function" would be tiny compared to the weight on pathogen resistance. Kimmel concludes that aging falls into the category of problems evolution never seriously tried to solve, which means it should be relatively easy to intervene—just as modern medicines, which are "incredibly simplistic" by biological standards, still provide massive benefits.

---

12:07Why Humans Didn't Evolve Their Own Antibiotics

Patel poses a sharp question: if evolution cares so much about infectious disease, why didn't humans evolve their own antibiotics? Kimmel explains that antibiotics are metabolites produced by bacteria and fungi locked in an evolutionary arms race—the "Red Queen hypothesis," where both sides must run just to stay in place. Bacteria and fungi have enormous population sizes (trillions in a drop of water) and can tolerate high mutation rates because they are prokaryotic; a single mutated cell doesn't compromise the whole organism. Metazoans like humans cannot afford such mutation rates because a single mutated cell can become cancer.

Even if the human genome stumbled into encoding an antibiotic, pathogens would likely mutate around it quickly. But Kimmel offers a fascinating counterpoint: there are "fossil" defenses in our genome. He describes TRIM5α, a gene that once protected primate ancestors against an HIV-like pathogen. At some point, a massive endogenous retrovirus threatened the genome so severely that TRIM5α was retooled to fight it—and in the process, lost its ability to restrict HIV-like viruses. That retrovirus later went extinct for unknown reasons, but the gene was never updated. Today, just a few edits to TRIM5α in human cells can restore HIV restriction. This illustrates how evolution can lose useful defenses through contingent trade-offs, and how "historical antibiotics" might be recoverable.

The mechanism that makes such evolutionary pivots possible is gene duplication. When a new threat emerges, the genome can copy a gene, preserve the original function in one copy, and let the duplicate mutate freely—often accumulating neutral mutations until it stumbles upon a new useful function. This is how evolution avoids the problem of needing multiple simultaneous mutations that would each be net negative on their own.

---

25:26De-Aging Cells via Epigenetic Reprogramming

NewLimit's approach targets the epigenome—the layer of chemical marks on DNA and associated proteins that determines which genes a cell uses. All cells in the body share the same genome, but a liver cell and a neuron use different genes because of their epigenetic state. As cells age, the epigenome degrades: the marks shift, cells lose the ability to use the right genetic programs at the right times, and they become less functional and more disease-prone.

The goal is to remodel the epigenome back toward a youthful state using transcription factors—proteins that act as "orchestra conductors," binding specific DNA sequences and turning genes on or off. Kimmel describes transcription factors as evolution's levers: because a small number of base-pair changes in a transcription factor can produce large phenotypic changes, they are an efficient substrate for intervention. The challenge is that each transcription factor binds hundreds to thousands of genomic sites, and there is no guarantee that aging follows a simple path that can be reversed by any natural combination.

Patel asks why NewLimit can't simply replicate Shinya Yamanaka's Nobel-winning approach, where he identified four transcription factors that turn any adult cell into an embryonic stem cell by systematically testing combinations of 24 candidates. Kimmel explains that Yamanaka's problem had two features that made it tractable: success was trivially measurable (stem cells look completely different and express unique genes), and successful cells proliferate rapidly, amplifying even a 0.001% success rate into visible colonies. Aging is different: old and young cells of the same type look nearly identical, requiring complex molecular measurements like single-cell RNA sequencing to distinguish them. And successful de-aging doesn't amplify—you can't just turn cells into cancer to make them grow. With roughly 2,000 transcription factors and the need to test combinations of 1–6, the search space is about 10^16 combinations, far too large for exhaustive screening. This is why machine learning models are essential: they can learn from sparse experimental data and predict which untested combinations are most likely to produce a desired cell state.

---

44:43Viral Vectors, Lipid Nanoparticles, and the Future of Delivery

Getting therapies into the right cells is one of the hardest problems in medicine. Kimmel explains that traditional drugs can't target transcription factors directly: small molecules are too small to disrupt the large binding interfaces between transcription factors and DNA, while recombinant proteins and antibodies are too large to cross cell membranes. New nucleic acid medicines—mRNA or DNA delivered via lipid nanoparticles (LNPs) or viral vectors—finally make it possible to deliver transcription factors into cells.

LNPs, which powered the COVID-19 vaccines, are fat bubbles that fuse with cell membranes. By default they target the liver, but researchers are engineering them with antibodies on their surface to reach other cell types. Viral vectors like AAVs (adeno-associated viruses) are engineered viruses that carry DNA cargo into cells, but they have limited packaging capacity and are inherently immunogenic. Kimmel notes that both approaches have fundamental limitations: viral vectors will always provoke some immune response, and LNPs must navigate from the bloodstream to specific tissues without fusing with the wrong cells along the way.

Kimmel's "controversial opinion" is that by 2100, neither LNPs nor viral vectors will be the primary delivery mechanism. Instead, he envisions engineered cells—likely modified T cells or B cells—that patrol the body, sense specific environmental signals, and release therapeutic payloads only where needed. The immune system already solves this problem: T cells can access almost every tissue, run "AND gate" logic to detect combinations of signals, and deliver cargo (enzymes, signaling molecules) with precision. By engineering these cells, we could create therapies that live in the body for years, responding dynamically to the patient's condition. This is conceptually similar to CAR-T therapy for cancer, but generalized to deliver any therapeutic cargo to any tissue.

Patel asks whether delivery limitations mean we'll only be able to treat some tissues while others age normally. Kimmel argues that even partial success could have outsized benefits because the body is highly interconnected. Liver transplants from young donors reduce recipients' risk of multiple diseases beyond liver function. Ozempic, which targets a single hormone receptor, has knock-on benefits for cardiovascular disease, addictive behavior, and possibly neurodegeneration. Similarly, reprogramming even one cell type—like hepatocytes or hematopoietic stem cells—could produce systemic health improvements through endocrine signaling and tissue crosstalk.

---

1:06:22Synthetic Transcription Factors and the Limits of Natural Biology

Kimmel addresses whether the optimal reprogramming medicine will use only natural human transcription factors. His answer is nuanced: natural transcription factors provide a good "basis set" because evolution has shaped them to navigate cell states that human cells might plausibly occupy. But there's no guarantee that aging follows the same trajectories as development, and the ideal combination might never occur in nature.

He points to work on "Super Sox"—a mutated version of the SOX2 transcription factor that dramatically improves the efficiency of iPSC reprogramming (converting adult cells to stem cells). Since iPSC reprogramming never happens in nature, there was no reason to expect natural transcription factors to be optimal for it. Simple modifications like mutagenizing one factor or swapping domains between factors produced large improvements. Kimmel predicts that the end-state products of aging research will likely include synthetic genes that have never existed in any organism.

Patel raises the example of sagging skin, which is caused by degradation of elastin fibers—long polymerized protein cords that hold skin in place. These fibers are only formed during development; adult cells can make the individual protein units but can't polymerize them into new fibers. This isn't a cellular aging problem in the usual sense—young skin cells don't fix it either. Kimmel suggests the solution would require engineering cells into a state that doesn't exist naturally: one that can reinvigorate the polymerization process or repair damaged fibers. This is a case where de novo engineering, not just restoring a youthful state, will be necessary.

---

1:09:31Can Virtual Cells Break Eroom's Law?

Eroom's Law (Moore's Law spelled backward) describes the consistent trend that the number of new drugs approved per billion dollars invested has declined since the 1950s. Kimmel explains that this isn't because drug discovery is getting harder in an absolute sense, but because the low-hanging fruit has been picked, and the remaining problems are more complex. Unlike AI scaling, where increased investment has produced super-exponential returns (the promise of AGI), drug discovery has seen costs rise without proportional increases in revenue per drug.

The fundamental problem is that drug discovery has been bespoke: success at treating disease X doesn't make it easier to treat disease Y. Most of the risk isn't in making a molecule that hits a target—it's in figuring out what target to hit in the first place. Kimmel argues that a "general model" for biology would need two properties: it would generate medicines with very large addressable markets (everyone ages, so anti-aging therapies could treat everyone), and it would enable compounding returns, where each discovery increases the probability of success on the next.

The "virtual cell" is one vision for such a model. The idea is to measure cells' molecular states (typically by RNA sequencing), perturb them in many ways (turning genes on and off), and train models to predict how any perturbation will change the cell state. With enough data, you could search in silico for interventions that shift diseased or aged cells back toward health. Kimmel notes that NewLimit is building exactly this kind of model, but focused on a narrow regime: predicting how combinations of transcription factors affect specific cell types that are currently deliverable. This is analogous to building a specialized LLM for code completion rather than a general-purpose model.

Patel draws a parallel to LLM development: first you pre-train on a general objective (predicting the next token), then you fine-tune with RL on specific tasks. Kimmel agrees, though he notes NewLimit doesn't use RL yet. The pre-training objective is predicting the full transcriptome from a perturbation; the "value judgment" head then selects for younger-looking cell states. This same architecture could be extended to other goals—making T cells less inflammatory for autoimmune disease, or making neurons more functional for neurodegeneration.

Patel asks why, if perturb-seq (the core technology) has existed since 2016, we're still waiting for breakthroughs. Kimmel explains that early versions had terrible data quality: only about 50% of cells could be correctly labeled with which perturbation they received. For combinatorial perturbations (multiple genes at once), the fraction of correctly labeled cells dropped exponentially. Sequencing costs have fallen from dollars per cell to fractions of cents, and technical improvements have made labeling nearly perfect. NewLimit can now generate millions of single-cell data points in a single day—a scale that was impossible even a few years ago.

---

1:31:32Economic Models for Pharma in the Age of Longevity

Patel raises a practical concern: if anti-aging therapies work, how will anyone make money? The US healthcare system already spends 20% of GDP, and most of that goes to administering existing treatments rather than developing new ones. If a therapy costs tens of dollars to manufacture but tens of thousands to administer (due to doctor visits, scans, and monitoring), the system might not become more efficient.

Kimmel argues that the pharmaceutical industry is the one part of healthcare where technology has consistently reduced costs per unit of benefit—drugs eventually go generic, and old medicines continue to work. He notes that drugs account for only about 7% of healthcare spending, and that a third of all Medicare costs are spent in the final year of life. Preventing even a few expensive hospitalizations could dramatically reduce total system costs.

For reimbursement, Kimmel points to "pay-for-performance" models where the cost of a durable therapy is spread over its effective lifetime, with insurers paying only for the years a patient is on their plan (the average person churns insurers every 3–4 years). He also highlights Lilly Direct, where patients can order drugs directly from the manufacturer, bypassing pharmacy benefit managers and compounders. As medicines shift from treating acute disease to preserving health—where patients feel the benefit in daily life—direct-to-consumer models may become more viable.

On the structure of the industry, Kimmel notes that about 70% of approved drugs originate from small biotechs, not large pharma, even though large pharma spends most of the R&D dollars. Large pharma has effectively externalized early discovery to biotechs, partnering later for expensive clinical trials. Some firms, like Roche (which bought Genentech), maintain internal research groups doing cutting-edge work, but the general trend is toward a venture-capital-like model where small companies explore risky ideas and large companies commercialize them.

---

Conclusion

This episode matters because it bridges two conversations that rarely meet: the fundamental biology of aging and the practical economics of drug development. Kimmel's argument that evolution didn't optimize for longevity—and that this makes aging a solvable engineering problem rather than an immutable law—is both scientifically grounded and genuinely hopeful. The detailed discussion of delivery mechanisms, synthetic transcription factors, and the economics of reimbursement shows that Kimmel is thinking not just about whether de-aging is possible, but about how it would actually reach patients. The conversation leaves the listener with a clear sense that while the challenges are immense—from combinatorial search spaces to immune-privileged compartments to insurance churn—none of them appear fundamental. The tools are arriving: cheaper sequencing, better delivery, and AI models that can navigate biological state space. Whether NewLimit or another company will be the one to crack the problem remains to be seen, but the path forward is becoming visible.

---

Key Takeaways

• Evolution did not optimize for longevity because the baseline hazard rate was too high, kin selection may actively select against long-lived individuals, and the genome's optimization budget was consumed by fighting infectious disease.
• Aging is not monocausal; it involves multiple layers of molecular degradation, meaning no single "magic pill" will solve it, but partial interventions can still add healthy years.
• Epigenetic reprogramming using transcription factors can reset cells to a younger state, but the search space of ~10^16 possible combinations requires machine learning models to navigate.
• Current delivery methods (lipid nanoparticles and viral vectors) are improving but will likely be superseded by engineered immune cells that patrol the body and release therapies on demand.
• Synthetic transcription factors that never existed in nature may outperform natural ones for specific therapeutic goals, as demonstrated by "Super Sox" for iPSC reprogramming.
• Eroom's Law—the declining productivity of drug discovery—can potentially be reversed by building "virtual cell" models that enable compounding returns, where each discovery makes the next more likely.
• The economics of longevity therapies will require new reimbursement models like pay-for-performance and direct-to-consumer distribution, but preventing expensive end-of-life care could reduce overall healthcare costs.