AI needs task-shaped datasets. By default it will accelerate progress at the biological layers we already measure well. Solutions for longevity live at the layers we don't. We can fix that.
I like this because it treats AI as a tool, not magic. If the biology is poorly measured, even a powerful model may only accelerate the parts we already understand. What kinds of longevity data do you think we most urgently need?
The ideal: Longitudinal human datasets with repeat measurements over at least a decade, and rich readouts (omics not cholesterol checks), and ideally with interventions.
There was not too much data on solving open Erdos problems in the web. But AI solved some of them. I think AI can learn a lot on the abundant types of data a lot, and then show great results after fine-tune on sparser data.
Erdos problems surely have easy feedback via formalizers, but e.g. trying to grow large vascularized organs in a robotized lab can also have good automatic feedback.
(4) Oracle Health Learning Health Network — 123M+ patients.
(5) Ping An — ~400M registered users, the Ping An Medical Master multimodal model trained on ~1.44B online consultation records and six large medical knowledge bases.
(6) UK Biobank — ~500,000 adults, deeply phenotyped, exomes complete, whole genomes in progress, multi-organ MRI on a ~100K subset, accessible to credentialed researchers worldwide.
(7) All of Us — ~1M target.
(8) FinnGen — 500K+ Finns linked to national registries.
(9) Estonian Biobank — 200K in a country of 1.3M.
(10) Million Veteran Program — 1M+ US veterans.
(11) China Kadoorie Biobank — 500K.
(12) deCODE — roughly half of Iceland, for thirty years.
(13) Recursion — 100B+ cells per year, millions of phenomics images per week.
(14) JUMP-Cell Painting — ~75M single-cell profiles in the public benchmark, ~1.6B cells in larger releases.
(15) Illumina — Billion Cell Atlas, CRISPR perturbations across 200+ disease-relevant cell lines, with a stated plan for 5B cells in three years.
They are not all accessible to any aging researcher, although probably not any aging researcher has the necessary computational power and knowledge to process it.
Much of what you discuss at the biological/cellular level, I tackle at the organism level in my book Life for Health, about solving chronic and age-related disease with life insurance, instead of health insurance.
The basic idea is to marry a long-term holistic approach to reversing disease and maintaining a better health trajectory (including medical and non-medical interventions), with a universal life insurance policy that rewards individuals as they achieve clinical outcomes, set with clinicians.
Of course, validating clinical outcomes requires regular blood draws (or other monitoring), creating the incentive to build a longitudinal dataset that's even richer than UK Biobank (including medical records, participant-provided app data, biosamples & scans, and narrative overlay about events and motivations.
I'd give participants a financial stake in their data, a micropayment just for allowing it to be reused (deidentified, of course), along with participation in future uses, including pre-qualification for future clinical trials or products relevant to their situations. Life insurers -- more than health insurers -- will want to ensure people have access to predictive tests including MCEDs, proteomics, etc.
The bottom line is that the effort to generate actionable data at all levels has to be built into the overall clinical delivery and financial/coverage structures by design. That creates the economic alignment that enlists participants into broad, longitudinal research that will benefit themselves and others like them.
(Note: for those interested in learning more about Life for Health, the book's available on Amazon, or contact me, and I'm happy to send a free e-book.)
This was great bro! Good job!
Much appreciated.
I like this because it treats AI as a tool, not magic. If the biology is poorly measured, even a powerful model may only accelerate the parts we already understand. What kinds of longevity data do you think we most urgently need?
The ideal: Longitudinal human datasets with repeat measurements over at least a decade, and rich readouts (omics not cholesterol checks), and ideally with interventions.
How much might it cost?
I think maybe both ask both maybe better FYI
https://undark.org/2026/09/25/opinion-ai-biology-data-gap/
https://undark.org/2026/09/25/opinion-ai-biology-data-gap/
Martin, have you seen this? → https://orabiomedical.com/
As a retail investor, what would u recommend we put money in to back quality AI x Biotech companies? I see many quality companies are all private atm.
Yeah nothing to recommend except consider who will be the constraining suppliers if what the private companies try to do happens.
There was not too much data on solving open Erdos problems in the web. But AI solved some of them. I think AI can learn a lot on the abundant types of data a lot, and then show great results after fine-tune on sparser data.
Erdos problems surely have easy feedback via formalizers, but e.g. trying to grow large vascularized organs in a robotized lab can also have good automatic feedback.
What do you think about these sources of data?
(1) Truveta — de-identified EHR for 120M+ patients, linked to claims at ~200M-patient scale.
(2) Tempus — 40M+ clinical records, ~9M de-identified and ingested, ~1.1B medical documents, ~4M sequenced samples, ~350 PB of multimodal data.
(3) Epic Cosmos — ~300M patient records, ~16B encounters.
(4) Oracle Health Learning Health Network — 123M+ patients.
(5) Ping An — ~400M registered users, the Ping An Medical Master multimodal model trained on ~1.44B online consultation records and six large medical knowledge bases.
(6) UK Biobank — ~500,000 adults, deeply phenotyped, exomes complete, whole genomes in progress, multi-organ MRI on a ~100K subset, accessible to credentialed researchers worldwide.
(7) All of Us — ~1M target.
(8) FinnGen — 500K+ Finns linked to national registries.
(9) Estonian Biobank — 200K in a country of 1.3M.
(10) Million Veteran Program — 1M+ US veterans.
(11) China Kadoorie Biobank — 500K.
(12) deCODE — roughly half of Iceland, for thirty years.
(13) Recursion — 100B+ cells per year, millions of phenomics images per week.
(14) JUMP-Cell Painting — ~75M single-cell profiles in the public benchmark, ~1.6B cells in larger releases.
(15) Illumina — Billion Cell Atlas, CRISPR perturbations across 200+ disease-relevant cell lines, with a stated plan for 5B cells in three years.
They are not all accessible to any aging researcher, although probably not any aging researcher has the necessary computational power and knowledge to process it.
Terrific essay!!!
Much of what you discuss at the biological/cellular level, I tackle at the organism level in my book Life for Health, about solving chronic and age-related disease with life insurance, instead of health insurance.
The basic idea is to marry a long-term holistic approach to reversing disease and maintaining a better health trajectory (including medical and non-medical interventions), with a universal life insurance policy that rewards individuals as they achieve clinical outcomes, set with clinicians.
Of course, validating clinical outcomes requires regular blood draws (or other monitoring), creating the incentive to build a longitudinal dataset that's even richer than UK Biobank (including medical records, participant-provided app data, biosamples & scans, and narrative overlay about events and motivations.
I'd give participants a financial stake in their data, a micropayment just for allowing it to be reused (deidentified, of course), along with participation in future uses, including pre-qualification for future clinical trials or products relevant to their situations. Life insurers -- more than health insurers -- will want to ensure people have access to predictive tests including MCEDs, proteomics, etc.
The bottom line is that the effort to generate actionable data at all levels has to be built into the overall clinical delivery and financial/coverage structures by design. That creates the economic alignment that enlists participants into broad, longitudinal research that will benefit themselves and others like them.
(Note: for those interested in learning more about Life for Health, the book's available on Amazon, or contact me, and I'm happy to send a free e-book.)