Dwarkesh Podcast: Why Evolution Designed Us to Die Fast, & How We Can Change That – Jacob Kimmel

Date: August 21, 2025
Guest: Jacob Kimmel, President & Co-founder, New Limit
Host: Dwarkesh Patel
Episode Theme:
A deep dive into why evolution has not optimized for long human lifespans, how new biotechnological advances—epigenetic reprogramming in particular—are poised to intervene, and how progress in AI models and biotechnology may reshape the landscape of aging, medicine, and society.

Episode Overview

Dwarkesh Patel interviews Jacob Kimmel about the evolutionary reasons humans age, what it would take to intervene in the aging process at the cellular level, and the cutting-edge work at New Limit—where AI models and massive data are being deployed to find interventions that could extend human healthspan. They discuss analogies between biology and machine learning, new drug development paradigms, challenges in delivery and regulation, and the transformative potential of general-purpose "virtual cell" models for customizing medicines.

Key Discussion Points & Insights

1. Why Did Evolution Make Us Die? (00:11–12:07)

• Hazard Rates & Evolutionary Selection Pressure
  • Naturally high death rates (predators, accidents, infection) meant most didn’t live long enough for extended health or longevity to matter from a genetic selection perspective.
  • "The baseline hazard rate was very, very high. Even absent aging, you’re unlikely to actually reach those outer limits of health." — Jacob (01:51)
• Constraints on Evolutionary Optimization
  • Genome changes occur slowly due to limitations in mutation rate and population size.
• Interplay With Intelligence
  • Length of adolescence and brain capacity are limited by risk; optimizing for more intelligence or a longer development period is only selected for if survivability supports it.
  • "If you made adolescence too big, then you would just die before you get to have kids. ...maybe intelligence is easier than we think, and there are a bunch of contingent reasons evolution didn’t churn as hard on this variable as it could have." — Dwarkesh (03:33)
• Regularization Against Longevity
  • Long-lived but less productive (aged) individuals may consume resources without contributing, imposing a cost on the genome’s propagation (akin to regularization in ML).

2. Antibiotics, Host Defense, and Evolutionary Constraints (12:07–20:51)

• Why Humans Didn’t Evolve Antibiotics
  • Pathogens and host are locked in a ‘Red Queen’ arms race—bacteria/fungi outcompete with higher mutation rates and larger genomic populations.
• Genetic Defense and Duplication
  • Host defenses often arise from gene duplication, which allows the genome to ‘experiment’ without losing core functions.
• Examples of Ancient Defensive Genes
  • E.g., TRIM5alpha (human gene) offers clues about viral resistance and the tradeoffs in immune defense.

3. Aging: Multicausality and the Challenge for Medicine (22:17–24:09)

• No Magic Bullet Theory
  • Aging is not monocausal; interventions are likely to partially improve healthspan, not “fix everything at once.”
  • "I don’t think that there is a single monocausal explanation for aging." — Jacob (23:17)

4. Epigenetic Reprogramming & New Limit’s Approach (25:38–31:42)

• What Are Transcription Factors?
  • The “orchestra conductors” of the genome, telling cells which genes to turn on/off and controlling cell identity and function.
• Epigenetic Drift and Aging
  • The epigenetic code that maintains youthfulness decays over time; idea is to restore youthful patterns to rejuvenate function.
• Why Simple Approaches Don’t Scale
  • Unlike the Yamanaka factors discovery (resetting cells to stem cells), identifying ways to restore youthfulness in all cell types is a problem with combinatorial complexity and no easy readout. AI models are required to handle the vast search space.

5. Machine Learning Analogies in Biology (31:42–43:24)

• Basis Set of Transcription Factors
  • The genome is akin to code—transcription factors are like modular, composable programmatic primitives (“basis directions”), allowing both evolution and engineers to leverage combinatorial flexibility.
• Attention Mechanisms Analogy
  • "TFs are kind of like the queries, the genome sequences they bind to kind of like the keys, genes are kind of like the values." — Jacob (42:08)

6. Can We Target Transcription Factors with Drugs? (45:23–50:07)

• Why Old Drugs Avoided Transcription Factors
  • Physical constraints: small molecules can’t easily disrupt large protein-DNA interfaces.
• New Modalities
  • Advances such as mRNA-lipid nanoparticles (as used for COVID vaccines) and viral vectors (like AAVs) now allow engineering cells at the transcription factor level.

7. The Grand Challenge: Delivery Across the Whole Body (50:07–56:35)

• Current Strategies
  • Lipid nanoparticles (LNPs): “fat bubbles” targeting certain tissues (not universal, best so far for liver/immune cells).
  • Viral vectors: effective but limited by cell-type specificity and immunogenicity.
• Long-Term Vision: Cellular Delivery
  • "I think we're probably going to have to solve delivery the way that our own genome solves delivery... The immune system." — Jacob (53:27)
  • Envisioned: engineered cell therapies (akin to CAR T cells) for targeted, persistent delivery of anti-aging interventions.

8. Partial Progress Still Meaningful (58:27–63:04)

• Tissue-specific Benefits Can Be Systemic
  • Rejuvenating a single tissue (e.g., liver via hepatocytes) can have global effects—transplant studies confirm systemic benefits from rejuvenated organs.
  • Knock-on effects already visible in current medicines (e.g., broad benefits of Ozempic/GLP-1 agonists).

9. Engineering the Rejuvenation Toolkit (63:04–69:31)

• How Many Transcription Factors Needed?
  • "Somewhere between one and five… small enough number that you can encapsulate it in current mRNA medicines." — Jacob (63:06)
• Longevity of Effect
  • Epigenetic marks are durable; interventions may persist for years or decades, but current data is limited to several weeks to a few years.

10. Next Frontier: General Purpose “Virtual Cell” Models (71:15–89:18)

• Why ML Scaling Laws Don’t (Yet) Apply in Biotech
  • Drug development is still bespoke, with no compounding benefit from prior discoveries.
• What Will Change That?
  • A general-purpose virtual cell model: an AI that predicts phenotypic effects of perturbations, trained on huge combinatorial data (perturb-seq).
  • "This is so similar to in LLMs you have first imitation learning with pre-training... then you do RL about a particular objective..." — Dwarkesh (81:35)
• Labeling & Data Infrastructure
  • Critical bottleneck is dense, reliable, combinatorially-labeled data in relevant cell types; New Limit is vertically integrating to own the “genomic corpus.”

11. Industry & the Future of Pharma (91:32–103:56)

• Monetizing Breakthroughs in an Age of Cheap Synthesis
  • IP challenges as molecules become easy to copy, but major revenue channels (especially in the US) will still reward brand-name drugs via payers and direct-to-consumer models.
  • "Drugs are roughly 7% of healthcare spend ... the OOM is right." — Jacob (97:29)
• Who Will Build the Platform?
  • Big pharma acts more like a late-stage buyer; innovation, especially in computational/ML approaches, is driven by small biotechs that create phase I/II assets.

Notable Quotes & Memorable Moments

• On evolution’s optimization focus:
  "If evolution spent a lot of time optimizing this, my job is going to be insanely hard. If not, potentially there are some low hanging fruit." — Jacob (04:05)

• On the unexpected generality of new drugs:
  "If someone told you I’m going to find a single molecule ... and it’s going to have benefits for weight loss, cardiovascular disease, ... addictive behavior, neurodegeneration, you would have told them they were crazy." — Jacob (62:01)

• On building new biotechnological tools:
  "We don’t have the same luxury [as Yamanaka did]... We actually need to screen a much broader portion of TF space in order to be successful ... The number of possible combinations is about 10^16." — Jacob (35:12)

• On integrating biology and machine learning:
  "You can kind of think of transcription factors as evolution’s levers upon the broader architecture of the genome." — Jacob (41:42)

• Visions for future delivery:
  "I imagine ... the way we will be delivering these nucleic acid payloads is actually by engineering cells to do it ... the same way the immune system already does." — Jacob (53:27)

• On the potential impact of targeted rejuvenation:
  "Even just one tissue can benefit other tissue systems in your body at the same time." — Jacob (59:19)

• On the general model for drug discovery:
  "What would a general-purpose platform ... look like for biotech? ... You would need models that enabled you to take the success in one medicine and lead that to increased probability of success on the next." — Jacob (71:15)

• Healthcare system reflections:
  "Pharmaceuticals are the one place where, because of the mechanism of things going generic ... you’re actually able to get more benefits per dollar." — Jacob (99:01)

Timestamps for Key Segments

• 00:11 — Evolution’s lack of focus on longevity
• 06:01 — RL analogy for evolutionary selection
• 12:07 — Why humans didn’t evolve intrinsic antibiotics
• 22:17 — Multi-causality of aging and medicine
• 25:38 — New Limit’s approach: epigenetic reprogramming
• 31:42 — Importance of AI models vs. brute-force experimentation
• 41:42 — ML analogies in genome/transcription factor logic
• 45:23 — Why big drugs don’t target transcription factors
• 50:49 — The delivery challenge: nanoparticles, viruses, the immune system
• 58:27 — Systemic benefits of partial anti-aging interventions
• 63:06 — How big must a reprogramming payload be?
• 71:15 — General-purpose virtual cell models & platforms
• 91:32 — Pharma and reimbursement in the age of easy molecule synthesis
• 101:05 — Is big pharma building the general-purpose platforms?

Episode Tone & Style

• Language: Deeply knowledgeable but accessible, filled with analogies to machine learning and software engineering.
• Style: Fast-paced, geeky, curious—with an eye toward first principles and technological optimism.
• Notable Moments: Jacob’s analogies (e.g., attention mechanisms in biology), Dwarkesh’s probing on how and when the “virtual cell” might become as foundational as LLMs, frequent references to scaling laws and parallels to AI advances.

Summary

This conversation encapsulates how new biological understanding, combined with AI-driven modeling and massive data, is converging on the possibility of not just treating aging but perhaps one day reversing it. Evolution didn’t 'try' to optimize for long lives due to high external risk and constraints on genetic change, but that also means there could be low-hanging fruit for modern interventions. Transcription factors, as nature’s modular regulators, are a promising handle, but their manipulation at scale requires data, models, and delivery platforms far beyond current pharma standards. Startups like New Limit are vertically integrating to own both the biological problem and the computational stack—with the long-term hope of generalizing discoveries across diseases, mirroring progress in ML.

For further deep dives into the science, business, and future of anti-aging interventions, visit www.dwarkesh.com.