Dwarkesh PodcastJacob Kimmel on Dwarkesh Patel: Why Evolution Ignored Aging
Why the high ancestral hazard rate cut longevity's gradient signal; editing TRIM5alpha shows epigenetic reprogramming can outpace evolution.
CHAPTERS
- 0:00 – 0:51
Why evolution barely optimized human longevity (and why that’s good news)
Kimmel lays out a core heuristic for bioengineering: if evolution strongly optimized a trait, improving it will be hard; if not, there may be “low-hanging fruit.” He argues aging sits in the under-optimized bucket, making intervention more plausible than many assume.
- •Heuristic: check whether evolution spent much time optimizing a trait
- •Aging framed as comparatively under-optimized vs. traits like pathogen defense
- •Early examples of biological “misses” that could be engineered around
- •Sets up longevity as an engineering problem rather than a fixed fate
- 0:51 – 3:09
Reason #1: High baseline hazard rates blunted selection for long life
The conversation develops the idea that most of evolutionary history featured high day-to-day mortality from injury, predation, and infection. If few individuals reach old age, selection has little signal to optimize late-life healthspan.
- •Hazard rate integrates all causes of death, not just aging
- •If most die young, there’s weak gradient signal to improve late-life traits
- •Implication: longevity wasn’t strongly selected even if it would help today
- •Connects evolutionary constraints to modern intervention opportunities
- 3:09 – 8:19
Longevity, adolescence, and the selection window for intelligence
They connect high hazard rates to why long childhoods/adolescence are costly, and how that shapes selection for intelligence. Kimmel proposes fluid intelligence peaks when it mattered most for reproduction—early adulthood—potentially explaining why many breakthroughs skew young.
- •Tradeoff: longer learning periods vs. dying before reproducing
- •Fluid intelligence preservation late in life may not have been selected for
- •A cross-cultural pattern: major mathematical/scientific breakthroughs skew early
- •Examples like Newton and Humboldt as ‘single-year’ career-defining bursts
- 8:19 – 10:00
Reason #2: Longevity can be selected against via demographic ‘regularization’
Kimmel introduces kin selection-style arguments: longer-lived but less-fit elders can reduce a genome’s total resource throughput. If extended lifespan doesn’t also preserve high functional fitness, selection may favor turnover.
- •Genome-level view: selection optimizes propagation, not individual welfare
- •Aging as a ‘length regularizer’ on calories consumed vs. contributed
- •If health declines persist, replacing elders with younger adults can be favored
- •Framing: longevity requires both longer life and sustained fitness
- 10:00 – 12:48
Reason #3: Evolution’s optimizer is constrained (mutation rate, population size, competing objectives)
The third pillar is optimization constraints: mutation rates cap step size, population sizes cap parallel search, and infectious disease often dominates the evolutionary objective function. Even if longevity were beneficial, it may have been underweighted versus immediate survival pressures.
- •Mutation rate bounds how fast genomes can explore solutions
- •High mutation rates raise cancer risk; low rates limit adaptation
- •Population size limits ‘parallel compute’ available to evolution
- •Infectious disease resilience likely took priority over longevity improvements
- 12:48 – 15:29
Why didn’t humans evolve antibiotics? Red Queen dynamics and evolutionary arms races
They explore why mammals didn’t evolve antibiotic-like metabolic programs despite heavy selection from infection. Kimmel argues microbes win arms races due to enormous population sizes and tolerance for high mutation rates, making antibiotic innovation and counter-innovation rapid.
- •Antibiotics largely come from microbial metabolites (bacteria/fungi)
- •Red Queen hypothesis: constant arms race to stay in place
- •Microbes have massive population sizes and can tolerate high mutation rates
- •Multicellular organisms face cancer risk from high mutation burdens
- 15:29 – 22:31
Lost defenses in our genome: TRIM5alpha, ancient viruses, and gene duplication as a search trick
Kimmel discusses evidence that hosts can “lose” prior defenses as pathogens change—illustrated by TRIM5alpha’s evolutionary history. The section also explains gene duplication as a mechanism that enables exploration without breaking existing essential functions.
- •TRIM5alpha: can be edited to strongly restrict HIV despite current mismatch
- •Host–pathogen co-evolution leaves fossils (e.g., endogenous retrovirus remnants)
- •Gene duplication allows neutral drift and multi-step evolutionary paths
- •Homology/gene families reveal duplicated and specialized genes across the genome
- 22:31 – 25:51
From evolution to intervention: aging isn’t monocausal, so therapies will likely be incremental
They return to longevity medicine risk: therapies may fix parts of aging without solving all decline. Kimmel argues aging is multi-layered, so the realistic path is successive treatments that add meaningful healthy years rather than a single ‘magic pill.’
- •Aging likely has multiple contributing mechanisms, not one upstream switch
- •Expect early therapies to add years of healthspan without eliminating decline
- •Evolutionary arguments align with a multi-causal aging picture
- •Sets expectations for staged progress in longevity biotech
- 25:51 – 31:32
Epigenetic reprogramming at NewLimit: transcription factors as ‘orchestra conductors’
Kimmel explains NewLimit’s approach: use transcription factors (TFs) to remodel the epigenome toward youthful states. They measure success not just by “looks-like” transcriptomic shifts, but by restored cellular function while avoiding identity drift and tumor risks.
- •TFs regulate gene programs by binding DNA and shaping epigenetic marks
- •Aging involves epigenomic drift that misregulates cellular responses
- •Assays: transcriptome ‘looks-like’ + functional readouts (e.g., hepatocytes, T cells)
- •Key safety concern: de-aging without changing cell identity or inducing cancer
- 31:32 – 45:42
Why AI/models matter: combinatorial TF search, single-cell genomics, and Yamanaka’s special case
Dwarkesh asks why the Yamanaka-factor approach can’t just be replicated for de-aging. Kimmel explains: aging is hard to measure, success doesn’t amplify, and the TF combination space is astronomically large—so single-cell readouts plus predictive models are essential.
- •Yamanaka succeeded because success was easy to detect and colonies amplify rare hits
- •Aged vs. young cells are nuanced; need high-dimensional molecular measurement
- •Single-cell RNA-seq enables detailed state measurement and model training
- •Combinatorics: thousands of TFs → ~10^16 combos for 1–6 factors; exhaustive search impossible
- 45:42 – 50:48
Why TFs are hard to drug—and how delivery modalities are changing (LNPs, AAVs, nucleic acids)
They discuss why traditional small molecules and antibodies struggle with transcription factors, pushing pharma to ‘bank-shot’ upstream targets. New nucleic-acid delivery (LNPs, viral vectors) makes TF modulation more direct, shifting the therapeutic landscape.
- •Many drugs affect TF activity indirectly via signaling pathways
- •Small molecules can enter cells but often can’t modulate TF–DNA interfaces well
- •Antibodies/proteins are too large to cross membranes to reach nuclear TFs
- •mRNA/LNPs and viral vectors enable direct TF expression inside target cells
- 50:48 – 1:03:44
The delivery bottleneck and a provocative long-term solution: engineered cells as living drug couriers
Kimmel surveys delivery strategies and argues current methods may not be the endgame. He suggests engineered immune-like cells could provide targeted, conditional delivery throughout the body, akin to programmable CAR-T logic but for therapeutic payload release.
- •LNPs: improving targeting but face physical constraints and off-target uptake
- •AAVs: useful but immunogenic, limited payload, and not universally tropic
- •Hypothesis: future delivery via engineered cells with AND-gate sensing and payload release
- •Body coverage: immune surveillance reaches most tissues, with a few immune-privileged compartments
- 1:03:44 – 1:07:03
How much payload, how often, and how durable? Practical constraints for reprogramming medicines
They discuss realistic dosing: a small number of TFs may be sufficient, and TFs’ naturally low expression suggests low doses could work. Epigenetic marks can persist for years in principle, but current evidence supports benefits lasting weeks to months, implying periodic dosing as a near-term path.
- •Typical payload might be ~1–5 TFs; within current mRNA delivery capacity
- •TFs are lowly expressed endogenously → efficacy may not require high copies
- •Epigenetic state persistence suggests long durability is possible in principle
- •Near-term expectation: dosing every month/few months rather than daily administration
- 1:07:03 – 1:10:12
Synthetic transcription factors and ‘extra-physiological’ states: going beyond what evolution built
Kimmel addresses whether non-human or synthetic TFs could outperform natural ones, citing examples like SUPER SOX improving iPSC reprogramming. They also discuss aging phenomena that aren’t purely cellular (e.g., elastin polymer aging) and may require engineered, non-natural repair programs.
- •Natural TFs are a good basis set, but may not be optimal for de-aging goals
- •Synthetic/engineered TFs can exceed natural performance (e.g., SUPER SOX)
- •Some aging effects (like elastin fiber polymerization failure) may require de novo states
- •Long-term vision: program cells into repair-capable states not seen in normal adulthood
- 1:10:12 – 1:24:55
Eroom’s Law, the ‘virtual cell,’ and using perturbation data to reverse drug R&D decline
They define Eroom’s Law (declining drug output per dollar) and explore how a general-purpose “virtual cell” model could create compounding returns. Kimmel describes training models on perturbation→state transitions, analogous to pretraining and then task-specific optimization in LLMs.
- •Eroom’s Law: fewer new drugs per inflation-adjusted R&D dollar since the 1950s
- •Virtual cell idea: learn how perturbations shift cell state from rich molecular readouts
- •Sparse experimental sampling + in silico search over vast combinatorial spaces
- •Analogy to LLM stack: representation learning first, then objective/value heads for specific goals
- 1:24:55 – 1:32:12
Why progress has been slow: Perturb-seq’s maturation, scaling costs, and combinatorial labeling challenges
Dwarkesh asks why Perturb-seq-era ideas (since ~2016) haven’t already transformed medicine. Kimmel points to practical bottlenecks: sequencing costs, readout quality, barcode detection fidelity, and the huge jump in complexity from single to combinatorial perturbations—now improving rapidly.
- •Early single-cell perturbation assays were expensive and noisy
- •Barcode/perturbation detection errors can destroy supervision signal
- •Combinatorial perturbations amplify labeling problems (error compounding)
- •Recent cost and throughput improvements enable million-cell experiments routinely
- 1:32:12 – 1:45:20
Pharma economics for durable therapies: pay-for-performance, insurer churn, and direct-to-consumer shifts
They close on how companies get paid for long-lasting, health-preserving treatments, and whether gray markets undermine incentives. Kimmel argues reimbursement innovation (e.g., pay-for-performance annuities) and more direct purchasing channels may better fit durable therapies, while prevention could reduce costly end-of-life spending.
- •Insurer churn (3–4 years) misaligns incentives for long-horizon health benefits
- •Pay-for-performance: spread payments over time contingent on ongoing efficacy
- •Direct-to-consumer channels (e.g., ‘Lilly Direct’) may expand for health-preserving drugs
- •Prevention could reduce high-cost late-life care; drugs are a small share of total spend
