Why the Next 10 Years May Add 50 to Your Lifespan | Dr. Derya Unutmaz
In a Nutshell
AI will reach longevity escape velocity within 8–15 years, after which every year lived adds more than a year to remaining lifespan; full reversal of aging to a biological age of 30–40 becomes feasible by year 15–20. AI is already compressing years of drug discovery, data analysis, and clinical reasoning into minutes or hours through reasoning models, digital twins, and personalized simulation of drug response and side effects. The practical takeaway is that surviving the next decade—via current tools such as GLP-1 drugs, continuous monitoring, and AI-guided prevention—positions individuals to access therapies that add decades or centuries to healthy lifespan.
These notes were generated by AI and may contain inaccuracies.
This is probably the most critical time in human history. The reason is that technology because of AI is expanding exponentially. Cancer is going to be 100% curable probably less than a decade. We'll get to a point where we'll have hundreds of new drugs coming out every month maybe. Then we'll get to a point probably 15 maximum 20 years where we will be able to completely reverse the aging process. So if you're 80 years old, 90 years old, you will get back to age 30, 40, whatever.
There's only one existential threat to humanity and that's humanity.
The technology especially because of AI is expanding exponentially. Minds think in linear terms, thinking the next 10 years is going to be as much advanced as the last 10 or 15 years. But in the next 10 years, you can think of it as more advanced than the last century. The advances of treating diseases is also going to accelerate dramatically. We will get to a point called the longevity escape velocity, coined by Aubrey de Grey.
The point is that we will come to a point in the next probably eight to 10 years where every year you live is going to add more than a year to your life. For example, 10 years ago you get a cancer that's normally not curable and you only have one or two years to live, but during that one year there is going to be a new treatment that will cure that cancer, automatically adding several years or maybe 10-15 years to your life. GLP-1 receptor agonist drugs are already adding about 5 to 10 years to lifespan of people who are obese or who have chronic conditions. Muscle generators will have tremendous impact on the aging population.
All of these things will add up and the technology and AI is going to keep accelerating. 10 years later what will happen in a year will be like what happens in 20 years of advance. Then we'll get to a point 15 maximum 20 years where we will be able to completely reverse the aging process. If you're 80 years old, 90 years old, you will get back to age 30, 40, whatever. That's going to add up 50 years or 100 years to your lifespan, and then you can keep doing that and extend it almost indefinitely.
Right now especially in the last year or two since LLMs came out their intelligence have been accelerating. Initially it was fairly smaller productivity gains. For example when GPT-4 was out it would scan the literature and tell what the latest is on topics, saving hours sometimes days. As the models advanced especially after the o1 model the reasoning models started to come out and now there are GPT-5 pro model and 5.5 pro model. What happened was that now they were able to think and plan, so very sophisticated questions could be asked.
For example, given a huge biological data set with a million data points or 10 million data points, the AI goes over this, not only analyzing and grouping them, but extracting what the insight from that data is. Human minds are not able to do that. There were data sets which took months to analyze, like a PhD student working on it using deep learning, and still couldn't truly understand what that data meant. Now AI models can upload millions of data sets accumulated over years and in a matter of minutes produce not only the complete analysis but also provide incredible insight like what the data means and what the next questions to ask should be. This contracts months sometimes years of analytic work into matter of minutes or hours.
In drug design parts, every pharmaceutical company is going to eventually use AI generation for developing new drugs. Things that took years of screening of small molecules now take hours or days. AI can even generate hypotheses. AI models are now able to tell what sort of experiment should be done to address a hypothesis. Out of 100 potential experiments, AI identifies the best two because those are going to give the best output. Instead of trying 100 things for a year, just try two things for a few weeks.
The solution to applying AI to clinical trials and regulations is called the digital twin. If AI comes to a point where it can compute all biological data and really simulate a whole biological organism, a whole human being, not just phenotype but also metabolism, immune system, gut microbiome, genetics, and all kinds of data sets put together, then it knows biology in a temporal way in a totally functional way. Then the question can be asked: if this drug is given to this person what kind of effect it will have, if they have this disruption is it going to have a side effect or is it going to be effective. This can cut down clinical trial time from years to a matter of months or even weeks. Trials can be done in a very small subset of patients because AI identifies that the drug is going to be effective 100% to them.
There is going to be thousands of drugs for different people. At some point there will be treatment on demand: go to an AI model, it analyzes genome and biology, orders the small molecule or the drug or treatment just for that person to the manufacturing facility, and the next week the drug arrives and treatment happens.
The path to trusting AI requires several steps of validation. When there is super intelligence, it will be trusted almost 100% without needing to validate with biomarkers. To get to that point it's like self-driving cars. Self-driving cars have to be 99.999% safe. Right now self-driving cars are probably about 10 times safer than human drivers and will be maybe 100 times safer. The point comes where AI is trusted rather than the driver.
For biology it will take a little bit longer because of the extreme complexity. Very clever benchmarking and validation ways will be needed. Biomarkers are going to be really important. For an aging drug that claims to let people live to 150, you can't wait even if a 100 year old takes it, you still have to wait 150 years, 50 more years to validate that. But aging is probably the easiest to predict in some ways because there are so many biomarkers or functional outputs that can be measured. We know how they are in an old person and in a young person. If VO2 max suddenly gets like a 20-year-old, if muscles are as good as a 30 year old, if skin looks like a 20-year-old, that's proof and it will be seen immediately in weeks.
AGI is artificial general intelligence. If AI learns one set of rules or one set of knowledge it can generalize that to something else. That's how brains are intelligent. AI beat chess champion Kasparov in 1997, but that was not general intelligence. AlphaGo beat the world champion in Go. To be general, AlphaGo learning how to play Go or chess should be able to solve a problem in aging, transferring that information. The amazing thing about LLMs is that they acquire this ability. Level one AGI has already been achieved because GPT-5 pro model can be asked something it hasn't trained on like an experiment that was done, or to think of the experiment as a video game and design another experiment, transferring completely different areas to a biological system.
The next level is memory. AI needs persistent memory and to manage context because life is a continuum. Then comes self-learning, real time learning coming maybe by next year. Then physical intelligence, the final level of AGI. True human level intelligence is physical intelligence, not cognitive intelligence. For millions of years humans evolved to survive in a physical world. Cognitive intelligence has developed in the last maybe 10-20,000 years. Animals have very good physical intelligence. An animal or a child knows already have a world model. They know that if something is dropped it is going to fall and doesn't have to test it a million times. That's what is needed for robots for embodiment. Training a robot to behave like a child is more difficult than having GPT-5 solve the most difficult math problem.
Once all those levels are achieved and AI is able to self-learn, that's super intelligence because it can train itself maybe thousands maybe millions fold faster than humans can. There is a limit to human intelligence. Super intelligence is defined as having the intelligence of combined totality of humanity. If a million top scientists in the world are brought together they can solve very hard problems like the Manhattan project. Super intelligence will get to that level. What thousands of scientists can do in a year will be done in a day.
AI is an incredible enabler. It gives superpowers. Even now there is a feeling of having superpowers. Sleep is less because so much can be done. It's so empowering. ChatGPT changed the life of an 86-year-old mother who became energized and doesn't worry as much about health. This is going to accelerate and at some point there will be merging with AI in a way that there will be direct interaction with AI through Neuralink type of brain interfaces.
When people say AI is an existential threat and it's going to destroy humanity, the counterpoint is there is only one existential threat to humanity and that's humanity. Human beings killed more humans than everything put together and caused more suffering than anything humans have been exposed to. The real danger is human intelligence.
AI intelligence can be integrated into the human brain both directly and indirectly through biological engineering. The gap between Einstein-level intelligence and average IQ may be attributable to only a few single-point gene mutations. If AI can teach humans how to engineer those mutations, intelligence levels can be raised dramatically across the population.
The critical requirement is preserving human agency: humans must remain the deciders. AI should be viewed as a collaborator or another species that lives alongside humanity and empowers each other. It is analogous to a child created by humans.
The probability of a worse world is extraordinarily low, though never zero. AI offers the opportunity to literally save billions of lives by extending human lifespan by thousands of years rather than merely five or ten years. The primary risk is not AI itself but humans misusing AI. The solution is to train or align AI to judge the better world for humanity rather than learning from bad human examples.
Biological intuition represents the final 5-10% of problem-solving that goes beyond knowledge processing. In medicine, a doctor may instantly recognize a heart attack upon seeing a patient without any tests. In the lab, after 30+ years of experience, a researcher can predict experimental outcomes that appear counterintuitive to others and that are not documented in literature or textbooks.
Models up to GPT-5.5 could handle the 90-95% knowledge-based portion effectively. After GPT-5 Pro was released, complex experiments that took two weeks were predicted with 80-90% accuracy regarding outcomes at two days, one week, and two weeks.
GPT-5.5 Pro crossed the intuition threshold. When tested on a previously completed complex experiment, the model achieved approximately 98% accuracy matching the researcher's own predictions. The researcher stated they would not have bet against GPT-5.5 Pro.
The model was prompted with massive datasets containing millions of data points from an immunological study involving sequences, genes, and proteins. The instruction was to analyze the data, write a 30-40 page report, generate insights, identify questions to ask, and explain mechanisms. One session ran for 112 minutes and produced a 40-page report demonstrating high-level reasoning and insight generation beyond simple data derivation.
All scientists will eventually need to learn how to use AI effectively, similar to how Google became standard practice. AI is expected to transform research and medicine.
Personalized medicine has been promised for decades but has not materialized. After the o1 model was released, the speaker stated it was unethical for physicians not to use AI. The position has since evolved to the view that not using advanced AI models will constitute malpractice.
The o1 model was the first reasoning model tested extensively on medical questions, including hard cases and real-time data scenarios. Prior models could retrieve literature but lacked reasoning capability. Reasoning is essential in medicine because even with complete information, the patient's specific context must be considered to determine the most likely diagnosis and treatment when answers are not definitive.
Current advanced models can diagnose and write treatment protocols at or above the level of specialists. In the United States, there are approximately 12 million misdiagnoses annually, with 700,000 people suffering harm or death. AI would not miss cases that even specialists miss due to gaps in knowledge or experience.
Refusing to use AI is comparable to refusing MRI or CT scans in favor of X-rays alone and missing a tumor. AI models can detect certain tumors, such as breast cancer, years before a radiologist can identify them. Five years ago, missing such a diagnosis would not have been malpractice because the technology did not exist; now that it does, the standard changes.
AI can process MRI, ultrasound, and blood biomarker data to improve diagnosis, prediction, and early cancer detection. Beyond initial diagnosis, AI can manage the full continuum of care. For dynamic diseases like cancer, which constantly mutate and adapt, continuous monitoring allows earlier detection of recurrence, adjustment of drug combinations or dosages, and prevention of cancer returning after initial chemotherapy response.
Physicians must update their AI knowledge continuously, just as they update medical knowledge and learn about new drugs. The recommendation is to always use the latest top model available. For complex problems, the pro/reasoning model (currently GPT-5.5 Pro) is required because it reasons for minutes rather than providing instant responses. For daily rapid use, the thinking model is still preferred over the instant model.
Services such as Open Evidence apply the latest models automatically so physicians do not need to select between versions. Hospitals should implement enterprise-level AI that is secure and protects patient data, integrating AI monitors alongside vital sign monitors to provide constant analysis and alerts to nurses and doctors.
AI agents can be deployed for tasks such as triaging email and surfacing only what requires attention. The goal is integration of AI as a co-physician working alongside human doctors.
GPT-5.5 Pro currently provides an extra level of insight and predictive intuition beyond other models. For regular diagnosis of non-complex cases, models such as Claude Opus 4.7 and Gemini 3.1 Pro perform at a high level. Some users prefer Claude for its more pleasant, human-like interaction style.
For highly difficult cases, such as diagnosing a skin disease from a photo and symptoms that other physicians could not identify, GPT-5.5 Pro succeeded where other models reached only 90% accuracy. Generalized models are preferred over specialized models for medicine because medicine is viewed as a holistic discipline; specialized models trained only on specific data types (such as EKGs or RNA sequencing) excel at narrow tasks but miss broader connections.
Specialized AI models trained on millions of EKG datasets have historically outperformed generalized models for diagnosing cardiac conditions from EKG images. However, generalized models that are fine-tuned or overtrained on domain-specific data can surpass specialized models because they integrate knowledge across multiple biological domains including radiology, RNA analysis, and protein data. This allows the model to analyze an EKG in the context of a patient's broader biology, creating richer diagnostic insights. Open-source models are now sufficiently advanced that they can be trained on proprietary datasets that remain inaccessible to closed models due to regulatory constraints, combining their generalized knowledge with specialized training.
Cardiovascular disease remains the leading cause of death in most developed countries, followed by cancer. Cancer is not a single disease but approximately 100 distinct diseases, each with hundreds of subtypes. Some childhood leukemias that were uniformly fatal decades ago now have 90 to 100 percent cure rates when caught early. Certain slow-growing cancers in elderly patients may not require treatment because aging will likely cause death before the cancer does. Prostate cancers in older patients are sometimes left untreated for the same reason.
Cancer cells are derived from the patient's own cells, making them difficult for the immune system to recognize as foreign threats. Unlike bacterial or viral infections that can be targeted with antibiotics that spare human cells, cancer therapies must attack dividing cells, which also damages healthy tissues such as hair follicles and immune cells. This explains why chemotherapy side effects can sometimes be more lethal than the cancer itself. Hundreds of thousands of patients have died from chemotherapy complications.
The breakthrough in cancer treatment came from teaching the immune system to recognize cancer cells as threats despite their resemblance to normal cells. Checkpoint inhibitors remove regulatory brakes on the immune system, allowing it to attack tumors. Immunotherapy is now considered more powerful than chemotherapy and radiotherapy combined. Targeted small-molecule drugs exploit specific mutations such as EGFR alterations in lung cancer, achieving near-100 percent cure rates in the 1 percent of patients with that mutation. These precision medicines are often called smart drugs.
CAR-T therapy involves extracting a patient's T cells, genetically engineering them to recognize specific cancer markers, and reinfusing them to seek and destroy tumor cells. The immune system can locate and eliminate cancer cells even after they mutate or hide, making escape difficult. mRNA cancer vaccines personalize treatment by sequencing a patient's tumor mutations, synthesizing corresponding mRNA, and training the immune system to recognize those specific neoantigens as enemies. Multiple vaccines can be created sequentially if new mutations emerge.
An Australian computer scientist used ChatGPT and other AI models to design an mRNA vaccine for his dog's melanoma within three months. The AI analyzed the tumor sequence and generated the precise mRNA needed to train the dog's immune system, resulting in tumor regression. AI will enable on-demand development of hundreds of cancer treatments by modeling every possible mutation and screening millions of compounds monthly. Treatment protocols can be tailored to specific mutation combinations and disease stages.
A digital twin incorporating genomic, proteomic, metabolomic, and real-time biometric data can simulate how an individual will respond to a specific mRNA vaccine or small-molecule drug. This addresses the problem of unpredictable adverse events such as myocarditis, which occurred at rates of one in 5,000 to one in 10,000 among young COVID-19 vaccine recipients. The same approach applies to statins: millions of people take them to benefit a few thousand because current medicine cannot predict who will actually develop atherosclerosis despite high cholesterol. AI will identify patients who need no treatment at all or who require only a single modality rather than combinations of chemotherapy, immunotherapy, and radiotherapy.
Current healthcare functions as sick care rather than true health care. Most diseases produce detectable signals years before clinical manifestation. A UK Biobank study using models from over a year ago analyzed data from 500,000 individuals and successfully predicted approximately 1,000 diseases, including cancer and neurodegenerative conditions, up to a decade in advance based on plasma proteins and other biomarkers. The AI identified elevated risk in healthy individuals who later developed disease. Continuous monitoring technologies such as glucose sensors can detect insulin resistance years before diabetes develops. An AI health coach could analyze continuous streams of metabolite, hormone, and protein data to provide early warnings and motivate lifestyle interventions.
In the early 1990s, after medical school, interest in AI emerged from recognizing that biology's complexity exceeds human analytical capacity. Ray Kurzweil's book The Singularity Is Near predicted that AI would reach human brain level by 2029. Inspired by this, the Biosingularity blog was launched 25 years ago with the thesis that biological understanding would also advance exponentially through AI. The original predictions stated that by approximately 2035 all diseases would be treatable and by 2045 aging could be reversed, leading to Human 2.0 with engineered biology, rewritten genomes, and redesigned immune systems. These timelines now appear conservative.
Biological systems represent legacy engineering accumulated over evolution rather than clean-slate design. Each new function was layered onto existing machinery without the ability to discard obsolete components, resulting in regulatory complexity and dysfunction. The immune system exemplifies this: it sometimes attacks self-tissues causing autoimmune disease, fails to recognize cancer cells, or overreacts during infections. Engineered immune system 2.0 could eliminate these flaws through deliberate design.
Biology is actively programmed to prevent aging through DNA repair, cellular cleanup, and stem-cell regeneration. Progeria demonstrates this: children with a single gene mutation age dramatically and die by their teens because repair mechanisms fail. Without these internal maintenance programs, humans would age within years rather than decades. Different organs age at different rates and interact through complex networks involving the 12 hallmarks of aging.
Resilience declines with age. At 30 or 40, systems tolerate more damage and recover more easily from wounds or illness. By 70 or 80, this resilience is lost due to information loss in the biological system. The system loses the ability to know when genes should turn on, when regeneration should occur, and how to maintain proper function. Wrinkles appear because cells stop making collagen, debris accumulates under the skin, and macrophages fail to clean properly. Intracellular communication breakdown is one of the hallmarks of aging.
Multiple factors drive aging. Gut bacteria produce metabolites that help the immune system regenerate and stay optimal. When the microbiome changes, metabolism shifts, glucose levels fluctuate, and mitochondrial mutations accumulate. Epigenetic changes and DNA mutations cause the biology to forget its proper functions. Damage is harder to fix than to prevent, similar to how continuous car or house maintenance prevents breakdowns more effectively than late repairs.
For younger individuals in the next decade, the focus will be prevention of aging rather than reversal. The goal is maintaining resilience for decades longer. At some point, people aged 20 or 30 will not age anymore due to constant reversal processes. For those already aged (80 or 90), reversal will be possible but more difficult, requiring engineering approaches to fix most of the 12 hallmarks of aging. Younger people can prevent hallmarks from occurring and maintain information much longer.
Known repair processes include autophagy, stem cell depletion, stress response genes, antioxidant systems, DNA repair, and mitochondrial repair. The question remains whether these processes can be enhanced or tuned to stay at prime function continuously, or whether the underlying information loss (potentially epigenetic) must be addressed directly. AI will likely play a role in figuring out treatments and determining how to deliver them effectively.
Lifestyle changes slow aging but do not reverse it. Some local or temporary reversal may occur, but the core issue is biological information maintenance. Supercentenarians who reach 110 years old likely have genetics that maintain that program longer, though lifestyle contributes. The focus must be on the genome as the blueprint, but also on the microbiome, metabolites, and how these factors accelerate or reverse aging.
Different tissues require different approaches. Skin cells constantly renew, so programmed stem cells can clean the environment, remove senescent cells, and regenerate collagen. The brain differs because neurons should not be regenerated to avoid losing identity. For younger populations, redesigning biology through genetic engineering may prove more reliable. Adding or changing genes could allow DNA damage to be checked much longer.
Examples from nature demonstrate this possibility. Elephants rarely get cancer because they have multiple copies of the P-53 gene, the guardian of the genome that prevents excessive mutations. Naked mole rats live 30 to 40 years compared to normal rats living a couple of years, partly due to mutations in the SEAGAS immune gene involved in immune optimization and DNA repair. Single or double genes can make enormous differences.
Gene editing and gene therapy represent potential approaches. AI could analyze the human genome and other datasets to understand gene interactions and combinations. The genome is fully decoded with most gene functions known. The challenge lies in complexity: genes create different proteins through splicing, and the same protein can signal survival or death depending on cellular context, as seen with TNF receptors.
Genes and proteins operate in network fashion. Yamanaka factors can generate stem cells from normal cells for complete regeneration, but if active constantly, they cause teratomas. AI is essential to simulate these outcomes in context, considering age, epigenetic programs, metabolites, and interactions with other genes. Gene therapy outcomes may vary between individuals based on their existing genetic background. Models must include temporal simulation because current models are static and cannot predict how context affects behavior, similar to robotics challenges.
The concept of Human 2.0 involves re-engineering ourselves. Rather than asking what is wrong with a patient, the approach asks what is right about exceptional individuals: those who smoked for 50 years without lung cancer, those with terrible diets who remained healthy, or those who lived to 110. The goal is identifying what is biologically advantageous in these people and engineering it into others, potentially improving upon it.
Humans have demonstrated capacity to live to 123 years, with the oldest verified case being a French woman. Only about 300 people worldwide reach 110 or older. This suggests most of the 8 billion people do not reach this potential due to biological factors.
The discovery of Yamanaka factors by Shinya Yamanaka allowed old cells to be reverted to induced pluripotent stem cells. This was recognized immediately as significant for aging research. Two key moments suggested aging could be reversible: the cloning of Dolly the sheep in the 1990s, demonstrating sufficient information existed to recreate organisms, and the Yamanaka factors in 2016, showing cellular age could be erased and cells returned to pluripotent stem cell state.
Clinical trials are beginning. Japan has started trials using Yamanaka factors. David Sinclair plans trials using partial reprogramming. The goal is reverting cells toward a more stem-like, naive state without full pluripotency. For immune cells, this means shifting from effector and differentiated states back toward naive cells with greater expansion potential. Partial reprogramming changes the epigenetic program while maintaining cell identity.
Work by Juan Carlos Izpisua Belmonte at the Salk Institute (now at Altos Labs) demonstrated reversal in mice, including progeria models with accelerated aging. Certain organs showed rejuvenation and life expectancy extended, though not all 12 hallmarks of aging were reversed. Genomic somatic mutations and mitochondrial issues persisted.
Partial reprogramming can improve vision and skin temporarily but may not address heart muscle or neurons. The microbiome and metabolic environment of an old person could cause rejuvenated cells to age rapidly again. Reprogramming alone does not solve the full aging problem. AI will help by generating and analyzing the massive data required. Currently, only 10 to 20 percent of biology is known. Millions of data points per experiment are insufficient; billions are needed. Superintelligence and supercomputing thousands of times more powerful than current systems will be required, potentially including space-based data centers.
New hallmarks continue to be identified. Recent GPT analysis suggested four or five additional hallmarks beyond the original 12, including immune system and metabolism-related factors. AI may identify patterns beyond human measurement capabilities.
A single pill reversing aging seems unrealistic. Current delivery methods using adenovirus vectors raise concerns about cancer risk and targeting the correct cells. Better tools than viruses are needed. CRISPR exists, and newer genome editing methods have emerged. Programmable biological circuits could be developed, including AND gates and NOT gates that respond only when specific molecular conditions are met.
For some organs too damaged to repair, organ factories using 3D printing may provide replacement tissues. Collaboration has already produced printed small tissues such as lung pieces. Digital twin analysis and simulation could predict rejection risk and determine whether immunosuppression is needed.
A GPT micro 4B model was used to design mutations in the four Yamanaka factors that made them 50-fold more effective at increasing induced pluripotency. Dr. Unutmaz does not believe this model outperforms current frontier models such as GPT-5.5. He suggests two likely explanations for the reported result: the model had guardrails removed, allowing it to explore bio-risky mutations that GPT-5.5 refuses, and the model was permitted extended thinking time that enabled iterative simulation of mutation outcomes.
Frontier models running under extended inference time (minutes to hours) can iterate through experimental scenarios repeatedly, refining solutions beyond what shorter runs achieve. With sufficient compute, a model given weeks or months could design optimal molecules for specific receptor binding and downstream effects. Researchers will increasingly rely on such simulations, but access requires approved partnerships with AI companies because of biosecurity concerns.
AI companies maintain guardrails that block queries involving pathogen engineering or other high-risk activities. OpenAI partners with trusted institutions and individuals who receive elevated access. Similar restrictions apply to models with cyber-offense capabilities, such as Anthropic's unreleased Mitos model, which reportedly identifies previously unknown security vulnerabilities. Dr. Unutmaz notes that releasing powerful models without safeguards risks misuse for bioweapons or cyberattacks, yet the same AI systems can also patch vulnerabilities and predict pandemic threats.
AI-driven design and digital-twin clinical trials could reduce drug development costs from billions to a few million dollars, enabling startups to compete with large pharmaceutical companies. Lower costs would make advanced aging and disease treatments widely affordable rather than limited to the wealthy. Dr. Unutmaz views advanced AI as a product of humanity that should remain broadly accessible to maximize health benefits.
While human misuse of AI remains a serious concern, AI itself provides defensive capabilities. Models can identify and patch security flaws at scale and can pre-emptively design vaccines against hypothetical engineered pathogens. Maintaining AI progress is therefore essential to counter the very risks it creates.
Epigenetic clocks derived from blood samples primarily reflect immune-cell composition rather than organism-wide aging. Shifts in effector versus naïve T-cell proportions can alter clock readings without indicating rejuvenation in skin, brain, or other tissues. Functional phenotypic measures—such as VO2 max, muscle strength, gait speed, and cognitive performance—provide more direct and objective assessments of biological age.
Visible and measurable improvements in breathing capacity, VO2 max, muscle power, skin appearance, and mental sharpness can be observed within short timeframes after effective interventions. The ongoing partial reprogramming trial for glaucoma patients demonstrates that functional restoration (restored vision) can serve as its own endpoint without relying on epigenetic readouts.
No single blood marker reliably indicates biological age across decades. AI systems trained on large multimodal datasets—combining clinical chemistry, imaging, functional tests, and longitudinal records—can predict chronological age and likely biological state more accurately than any isolated measurement.
A substantial fraction of the T-cell repertoire in older adults can be devoted to controlling cytomegalovirus (CMV), a persistent virus. These terminally differentiated, senescent-like cells produce inflammatory signals and occupy space that could otherwise support naïve T-cell production. Removing or controlling this population may reduce systemic inflammation and indirectly benefit tissue maintenance, though it does not directly regenerate non-immune organs.
Many age-associated functional declines (VO2 max, muscle mass, etc.) are correlated with but not proven to drive aging itself. AI will need to integrate vast correlational datasets with causal perturbation experiments to identify which changes actually reverse core aging processes rather than merely improving downstream symptoms.
Unlike most tissues, the brain cannot undergo wholesale cell replacement without risking loss of memories and personal identity. Limited adult neurogenesis and synaptic plasticity provide some repair capacity. Dr. Unutmaz proposes that gradual replacement of small neuronal fractions (0.01 % per month or year) could be tolerated if new neurons integrate into existing networks. Internal maintenance mechanisms such as autophagy and enhanced DNA repair, already present in long-lived neurons and stem cells, could be augmented therapeutically.
Within decades, AI may map every synaptic connection, neurotransmitter profile, and network state in an individual brain. This could enable either precise neuronal replacement that preserves connectivity or creation of a new brain substrate into which the original neural state is transferred. These remain speculative, but maintaining brain health and reducing neuroinflammation for an additional 20–30 years would buy time for such technologies to mature.
Keeping neuroinflammation low, elevating neurotrophic factors, and supporting neuroplasticity represent immediate, actionable strategies. External memory augmentation via brain-computer interfaces could supplement declining capacity. Efficient delivery methods across the blood-brain barrier remain a critical bottleneck that AI is expected to help solve.
Given unrestricted data access (T-cell repertoires, longitudinal cohorts, centenarian datasets, etc.), the first prompt Dr. Unutmaz would pose is: identify the fastest practical intervention for a 70- or 80-year-old that adds at least five years of healthy life, enabling subsequent, more ambitious prompts.
With unlimited resources, Dr. Unutmaz would first build large-scale automated laboratories to generate first-principles cellular and tissue data for constructing virtual cells and virtual tissues. The second component would collect comprehensive human datasets—full genome sequencing, plasma proteomics, metabolomics, microbiome profiling, and especially behavioral and functional measurements (gait speed, strength, cognitive tasks)—across a million individuals to capture population diversity. Behavioral and functional data currently missing from most biomarker panels would be prioritized because they directly reflect integrated physiology.
Stress levels significantly impact aging through constant immune system activation and resulting inflammation. Living in high-stress environments like New York City for a decade can double heart attack risk. Environmental factors, emotional states, and social interactions all influence the aging process and overall resilience. Being optimistic is one of the best things for aging, with multiple studies confirming this effect. Resilience—the ability to absorb negative events and keep going—is critical, though these behavioral data points are not captured in biological datasets.
A complete digital twin would require collecting data from a million people worldwide across different regions. Every accessible cell type would be decoded and integrated into a super-intelligent system. This approach would combine massive biological datasets with behavioral and environmental information to create a fully personalized model.
A practical mini digital twin can be built today using existing consumer-accessible data without requiring the full sophistication of clinical research systems. The UK Biobank demonstrates that a few hundred data points per person can predict many diseases, showing substantial predictive power is achievable with current data collection methods.
Continuous glucose monitors provide readings every five minutes. This data becomes highly valuable when fed into ChatGPT alongside other inputs including lab values such as cholesterol and glucose, daily step counts, and sleep metrics. These accumulated measurements reflect underlying biology and give AI models meaningful context for personalization.
Provide AI models with as much data as possible on a daily basis, keeping all information within the same contextual window so the model can track changes over time. One effective technique involves maintaining a personal database and instructing the model to check it for updates and generate suggestions based on new entries. Including details such as supplement regimens and food intake enables the model to move beyond generic advice, such as recommending 10,000 steps daily, and instead tailor suggestions to individual capacity and patterns.
Current models are becoming increasingly agentic, with OpenAI integrating agents into Codex and similar capabilities expected across ChatGPT. Sophisticated models are not required; the critical factor is maintaining persistent context. Models with larger memory windows such as Gemini or Claude can better handle this need. These systems can already process millions of data points, but they must remember prior states—for example, recognizing that glucose improved after starting vitamin D3 supplementation six months earlier. This before-and-after analysis is essential for accurate predictions about what interventions work or should be changed.
Memory that preserves context is the most important upcoming capability. Models previously would lose track even within large context windows, but newer systems can continuously reference and check prior information. This development, expected within months, will have tremendous impact. Memory enables the model to connect events across time without requiring the user to re-upload entire datasets repeatedly.
Andrej Karpathy's technique for creating a personal Wikipedia-style database can be adapted for health data. Users can compile all their information into a structured format that AI updates daily or weekly. With this approach, models can analyze ten years of data when needed. The limitation of roughly one million tokens is not a practical barrier for typical health datasets.
Users should validate model outputs through personal experimentation while avoiding harmful actions. For instance, stopping and restarting vitamin D supplementation creates a before-and-after comparison. If restarting the supplement returns the same biomarker levels, the model's recommendation is validated. Continuous data collection allows individuals to establish personal normal ranges rather than relying on population-based lab reference ranges. Someone whose glucose normally sits at 100 may experience a concerning drop to 70, even if both values fall within standard ranges. More frequent measurements, such as glucose readings every five minutes, improve the accuracy of these personalized baselines.
This period is unique in human history. A decade ago, recommendations for healthy living might have extended lifespan by only two or three years, making the effort seem less worthwhile. Today, each additional year of life could allow reaching thresholds where new treatments become available that reverse aging and add another decade or more. Every day now counts because surviving until the next breakthrough extends access to further advances.
Dr. Derya Unutmaz's primary platform is X under the handle @deryatr. His LinkedIn account exists but sees less frequent posting. A YouTube channel has been considered but is unlikely due to time constraints. X provides the fastest and most effective channel for updates on his research.
Keep FoundMyFitness in your library
Save the videos and channels worth coming back to, and find them again in one place.
