The eater that does not exist
Walk into any school cafeteria, any hospital nutrition program, any government dietary guideline, any clinical trial. Look at what is being prescribed and to whom. The advice is the same. The food pyramid. The plate. The serving sizes. The macronutrient ratios. The general population.
For a hundred years we have been writing nutritional advice for a person who has never lived. The average eater. A statistical fiction assembled from a population, fitted to no one, prescribed to everyone.
The grandmother with type 2 diabetes and the marathon runner with iron deficiency are told to eat the same way. The teenager whose microbiome is still settling into its adult composition and the patient whose metformin and metabolic medication shape every meal are told to eat the same way. Four bodies. Four entirely different stories. One advice. We have known this is wrong for decades. We have not had the architecture to do anything about it.
That is what has changed. Not the recognition that individuals differ. Recognition has been around since Hippocrates. What has changed is that the difference is now computable.
This brief essay is about the architectural shift that makes that statement true. It is a shift across three layers. The biology of the eater. The chemistry of the food. The mathematics that holds them together. And it is a shift that has been quietly assembling in research labs, regulatory filings, and production systems for the better part of fifteen years. The cumulative effect is large enough now that a generation of physicians and nutritionists will look back at the population-level prescription that defined their training and ask how anyone ever thought it was good enough.
The argument runs as follows. The average eater was a placeholder we used because the individual was not measurable. The individual is now measurable, in many dimensions, at consumer scale. The food on the individual’s plate is now describable in many thousands of dimensions, far beyond the macronutrients and micronutrients that have anchored a century of nutrition science. And the mathematics for high-dimensional decoding and human-like reasoning over both representations together has finally matured. What this opens is the most hopeful chapter in the science of food and health in a century. It also opens a difficult question about how to govern, regulate, and adopt advice that is genuinely individual rather than generically prescriptive.
The body becomes legible
Consider what is now measurable about a single person, at consumer scale, without entering a hospital.
The genome. Roughly twenty thousand protein-coding genes, decoded for under two hundred dollars, with the relevant variants in taste perception, lipid metabolism, drug processing, and disease susceptibility now annotated and clinically interpretable.1 The TAS2R38 variant that determines whether someone tastes phenylthiocarbamide as bitter is a real fact about a real person, not an aggregate.2 The FTO variant associated with appetite regulation, the APOE variants associated with lipid handling, the MC4R variants associated with satiety. Each is a piece of the eater’s architecture that the population average could not capture.
The microbiome. The gut microbial community of a single person can be sequenced from a stool sample and characterized at species and strain resolution, in studies that began with the Human Microbiome Project in 2007 and have since expanded into the American Gut Project, the British Gut Project, and a generation of consumer-scale services.3 The microbiome is not a fixed identity. It shifts with diet, antibiotics, stress, sleep, and travel. But at any given moment it is measurable, and its composition correlates with metabolic, immune, and cognitive outcomes that the population average cannot predict.
The metabolic state. Continuous glucose monitors, once the exclusive province of insulin-dependent diabetics, are now available over the counter and worn by athletes, peri-menopausal women, and curious knowledge workers tracking how their breakfast affects their afternoon. They produce a continuous time series of a single person’s glucose response to a single meal in a single context, which is an entirely different kind of data from a fasting blood glucose number drawn at an annual physical.4 The same wearable platforms are extending into cortisol, lactate, inflammation markers, and a small but growing list of additional analytes that until very recently required a phlebotomist and a centrifuge.5
C-reactive protein, the canonical marker of systemic inflammation, can now be tracked from sweat through wearable biosensors in research-grade prototypes published in Nature Biomedical Engineering.6 Cortisol, the adrenal hormone whose chronic elevation correlates with insulin resistance and metabolic syndrome, has been demonstrated in wearable form in independent academic work going back to 2018 and has since matured into multi-analyte platforms that simultaneously track stress and metabolic response.7
What this looks like, in aggregate, is a body becoming legible to itself. Not at the population scale where it has always been visible to actuaries and epidemiologists, but at the resolution of one person on one day, eating one meal, in one biological context. The genome, microbiome, metabolome, inflammation signals, glucose response, sensory phenotype, medication regimen. None of these are abstractions any longer. They are facts about a particular eater, available in real time, at a price that puts them within reach of a consumer rather than only a patient.
This is the first of the three architectural shifts. It is not a hypothesis. It is observable in the regulatory record, the consumer market, and the published peer-reviewed literature.
The food becomes legible
The second shift is on the other side of the fork.
For the entire history of nutrition science as a regulated discipline, food has been described in roughly one hundred fifty dimensions. Calories, protein, carbohydrate, fat, the standard panel of vitamins and minerals, the small handful of additional compounds that have been promoted to nutritional concern by epidemiology. Fiber, sodium, saturated fat, added sugar. The United States Department of Agriculture’s national nutrient database, on which most nutrition research and most dietary policy rests, tracks a list in this neighborhood.8 When a study reports that a diet rich in vegetables is associated with better cardiovascular outcomes, this is the resolution at which the diet was characterized.
The chemistry of food, however, is not in this resolution. It is much higher.
The Human Metabolome Database, in its current public version, contains structured records on more than two hundred thousand metabolites, including the food components that pass through the human metabolic system.9 Its companion food chemistry database, FooDB, catalogs roughly seventy thousand chemical compounds across about one thousand foods, each with detailed information on chemical structure, concentration ranges, sensory properties, and known biological activity.10 The Annual Review of Nutrition, in a 2024 piece on the foodome, summarized the current state of food chemistry by saying that the macronutrient and micronutrient framework that anchors nutrition policy captures perhaps one or two percent of the chemical compounds present in food.11
The New England Journal of Medicine, in a May 2025 review by Albert-László Barabási and colleagues, described the situation in starker terms. More than 139,000 distinct food molecules have been catalogued, of which only a small fraction are tracked in any national nutrition database. The remainder, which the authors call the "nutrition dark matter," exerts measurable effects on human physiology through interactions with the same protein networks that pharmaceuticals target.12 Modern chemical language models represent each food molecule as a vector embedded in a high-dimensional space, on the order of seven hundred and sixty-eight dimensions for one widely used model.13 This is the resolution at which food can now be characterized. The hundred-and-fifty-dimension picture that anchors a century of dietary policy is, against this, a low-resolution sketch.
The implication is uncomfortable for the field as it has been practiced. If food is a hundred-and-fifty-dimensional object, dietary advice can be standardized. If food is a hundred-and-thirty-nine-thousand-dimensional object, dietary advice has to engage with the chemistry that the older framework ignored. And once the chemistry is engaged, the interaction with the eater’s biology becomes the actual subject. The genome, the microbiome, the metabolic state. The advice is no longer "eat more vegetables." The advice is what specific compounds, at what concentration, will interact with this person’s biology in this context to produce the outcome the person is trying to achieve.
That is a different kind of advice, and it requires a different kind of architecture to deliver.
These two shifts, in the body and in the food, would not be sufficient on their own. A description of an eater in many thousands of biological dimensions, paired with a description of food in many tens of thousands of chemical dimensions, paired with the question of how the two interact, is a problem that ordinary statistical methods cannot hold. The third shift is that the mathematics for reasoning and non-linear optimization over very high-dimensional, structured problems has matured in parallel with the measurement and the chemistry. The technical detail of how this is done belongs in another piece. What matters here is that the joint problem of food and biology can now be held in a single computational frame, where forty years ago it could not.
What this opens
When the three shifts are placed alongside each other, measurable body, computable food, mathematics that can hold both, what becomes possible is not incremental. It is generational. Consider, first, personalized nutrition that means something. The history of personalized nutrition as a commercial category is a history of overpromising. Quizzes and questionnaires have been sold as personalization for decades, with the recommendations they produce drawing on the same population averages dressed in different language. What is possible now is materially different. A person’s glycemic response to oats, to rice, to a specific brand of bread, can be measured directly through a continuous glucose monitor over a period of days and weeks. The microbial community that mediates that response can be characterized from a single sample. The genetic variants that shape lipid handling and appetite regulation can be read once and referenced against every meal thereafter. A recommendation engine that holds these three streams in a single representation, alongside the chemistry of the food, can produce advice that is genuinely fitted to the person, rather than advice that is fitted to the population and labeled with the person’s name.
Consider, second, drug formulation that fits the patient. The pharmaceutical industry has known for decades that the population average is a fiction. The same drug, in the same dose, produces materially different outcomes across patients, and a substantial fraction of the difference traces to genetic variation in metabolism, microbial composition that affects absorption, and the food consumed alongside the drug. Pharmacogenomics has matured into a clinical specialty, and the FDA now publishes genotype-based dosing guidance for hundreds of drugs.14 The next frontier is the joint optimization of formulation and food. Designing the drug, the dose, and the dietary context together for a specific patient. This becomes computable when the same architecture that characterizes the eater can characterize the molecule and the meal.
Consider, third, the GLP-1 generation of metabolic medicines, which is the most significant pharmacological development in food and metabolism in fifty years. Semaglutide, tirzepatide, and the next generation of agents under development have demonstrated that the human relationship with food is now pharmacologically modifiable at population scale.15 Tens of millions of people have taken these agents. The variation in response, in side effect profile, in optimal dose, and in the food, exercise, and stress context under which the agent works best is enormous. The hopeful claim of personalized medicine is no longer a promise about the future. It is an immediate need. The architecture that can hold the patient’s biology, the drug’s mechanism, the food on the plate, and the outcome the patient is trying to achieve, and reason over the full picture as one system, is the architecture that will decide whether the metabolic medicine generation is used well or used badly.
Consider, fourth, the older patient on multiple medications, eating in the context of declining appetite, shifting microbiome, and the slow inflammation of aging. Geriatric nutrition has been one of the weakest areas of clinical practice precisely because the population average is least informative there. The variation across older patients is dramatic. The consequences of getting it wrong, in falls, in cognitive decline, in metabolic dysfunction, are immediate. An architecture that holds the individual ontology can do for this patient what the average could never do.
These are not speculative. The components are deployed. The question is no longer whether the architecture is possible. It is how quickly it diffuses through the systems, clinical, regulatory, commercial, educational, that have been built on the average eater.
The case study, in public
A decade ago, McCormick and Company, the global flavor and seasoning firm, launched a system called FlavorPrint.16 The architecture was, at the time, ahead of its category. McCormick worked with a technology partner to build a multidimensional representation of flavor. The system was initially described in public materials as a framework of thirty-three flavor dimensions, later extended to a far higher-dimensional representation including thousands of dimensions of taste, aroma, texture, and lifestyle context.17 Every recipe in McCormick’s catalog, and every consumer who interacted with the system, could be expressed as a unique fingerprint in this space. The system was used both consumer-side, to recommend products and recipes that would land for a specific person, and inside the firm’s research and development organization, where the same representation accelerated the development of new products by mapping flavor white space and consumer preference onto the same ontological representation.
FlavorPrint was, in its mature form, a working production instance of the architecture this essay is about. It treated flavor as a high-dimensional object. It treated the consumer as a high-dimensional object. It used symbolic AI inference reasoning and ontological representations of food and flavor science to create the joint representation. It produced recommendations and product designs that were fitted to the individual rather than the average. It was deployed publicly. It was discussed in the trade press for half a decade.18
The reason it matters is what it demonstrated. In a public, traceable, on-the-record manner, the architecture worked at scale on the food side. The food was the first thing we made readable. Everything in this essay since the section on the foodome is the consequence of taking that demonstration seriously and asking what happens when the same architecture is extended to the body.
What still has to happen
The architecture is real. The deployment is uneven, and the institutional barriers are large. Three deserve naming.
The first is regulatory. The agencies that govern dietary guidance, medical devices, and pharmaceuticals have been built on the population average. The frameworks they use to evaluate evidence (the randomized controlled trial, the population endpoint, the standard of care) assume that the right comparison is between an intervention and a placebo across a population, rather than between an architecture and a population across many individuals. Adapting the regulatory framework to evaluate personalized recommendations is a substantial unfinished project, with real progress in pharmacogenomics and real friction in nutrition.
The second is clinical. The training of physicians, nutritionists, and pharmacists is built on the average. The standard of care reflects the average. The reimbursement codes reflect the average. Moving any of this to accommodate individual ontologies is institutional change, and institutional change moves slowly.
The third is interpretive. A high-dimensional, individualized recommendation, even when it is correct, is not always actionable. A person who is told that a specific bread interacts unfavorably with their microbiome and elevates their inflammation marker, but only when consumed alongside a specific medication and after a poor night’s sleep, has more information than they can act on without help. The architecture’s outputs have to be rendered in ways that fit the cognitive bandwidth of the person making the decision, which is a problem of user experience and clinical communication as much as it is a problem of mathematics.
These three barriers are why the architecture’s diffusion will be uneven and why the early years will look more like research and development than like consumer adoption. The science is ahead of the institutions. The mathematics is ahead of the regulators. The deployment is ahead of the training. The integration will take a decade.
The end of the average
The average eater was a placeholder. We used it because we had no other option. We did not have the measurement to characterize the individual at the resolution that mattered. We did not have the chemistry to characterize the food at the resolution that mattered. We did not have the mathematics to hold both representations together and reason over them. So we made do with a fictional person, prescribed advice to that fictional person, and accepted the population-level outcomes.
What the architectural shift means is that the placeholder is no longer necessary. The individual is now legible. The food is now legible. The reasoning that connects the two has finally matured. What follows is a generation of work in which the prescription stops being statistical and starts being architectural. Fitted to the specific eater, against the specific food, in the specific biological context, toward the specific outcome the person is trying to achieve.
The end of the average is not the end of nutrition science. It is the moment at which nutrition science becomes serious about the individual it has always claimed to serve.
The food was the first thing we made readable. The body is what we are making readable now. What follows is a chapter that the discipline has been waiting on for a hundred years.
Personalized food and nutrition driving health and longevity. The science is here. The architecture is here. What remains is the institutional work of making it the standard of care.
Notes
1. Mudge, J.M., et al. "GENCODE 2025: reference gene annotation for human and mouse." Nucleic Acids Research, 53(D1), D966 to D975, 2025. DOI 10.1093/nar/gkae1078. GENCODE Human Release 47 (October 2024) reports 19,433 protein-coding genes. Over the past two decades the count has been steadily revised downward as transcript-level evidence has improved. Live statistics at gencodegenes.org/human/stats_47.html.
2. Bufe, B., et al. "The Molecular Basis of Individual Differences in Phenylthiocarbamide and Propylthiouracil Bitterness Perception." Current Biology, 15(4), 322 to 327, 2005. The TAS2R38 gene encodes a bitter taste receptor whose two common haplotypes (taster and non-taster) account for most of the variation in PTC perception, with population-level frequencies that vary substantially across ancestral groups.
3. NIH Human Microbiome Project, launched December 2007 (commonfund.nih.gov/hmp). See the launch press release at genome.gov/26524200. The American Gut Project, established by Knight and colleagues at UCSD, provides additional consumer-scale microbiome data and is one of the largest open citizen-science microbiome datasets (americangut.org).
4. The FDA cleared the first over-the-counter continuous glucose monitor (Dexcom Stelo) on March 5, 2024, followed by Abbott’s Lingo and Libre Rio later that year. See FDA Press Announcement (fda.gov).
5. For a review of multi-analyte wearable biosensors, see Yang, Y., and Gao, W. "Wearable and Flexible Electronics for Continuous Molecular Monitoring." Chemical Society Reviews, 48, 1465 to 1491, 2019. See also the 2025 review in Nano-Micro Letters on noninvasive on-skin biosensors for diabetes management (pmc.ncbi.nlm.nih.gov/articles/PMC12314300).
6. Tu, J., et al. "A Wireless Patch for the Monitoring of C-Reactive Protein in Sweat." Nature Biomedical Engineering, 7, 2023. Caltech research (W. Gao group) demonstrating wearable CRP sensing without phlebotomy. Open access at pmc.ncbi.nlm.nih.gov/articles/PMC10592261.
7. Parlak, O., Keene, S. T., Marais, A., Curto, V. F., and Salleo, A. "Molecularly Selective Nanoporous Membrane-Based Wearable Organic Electrochemical Device for Noninvasive Cortisol Sensing in Sweat." Science Advances, 4(7), eaar2904, 2018. Stanford research demonstrating wearable cortisol measurement in sweat.
8. USDA FoodData Central, the consolidated public database for U.S. food composition data, tracks roughly 150 nutrients across most catalogued foods (fdc.nal.usda.gov).
9. Wishart, D. S., et al. "HMDB 5.0: The Human Metabolome Database for 2022." Nucleic Acids Research, 50(D1), D622 to D631, 2022. The published HMDB 5.0 release records 217,920 metabolites. The live database has grown beyond 250,000 entries since publication (hmdb.ca).
10. FooDB, version 1.0, hosted by The Metabolomics Innovation Centre at the University of Alberta (foodb.ca). Roughly 70,000 food components and food additives across about 1,000 raw and minimally processed foods.
11. Menichetti, G., Barabási, A.-L., and Loscalzo, J. "Decoding the Foodome: Molecular Networks Connecting Diet and Health." Annual Review of Nutrition, 44, 257 to 288, 2024 (pmc.ncbi.nlm.nih.gov/articles/PMC11610447).
12. Menichetti, G., Barabási, A.-L., and Loscalzo, J. "Chemical Complexity of Food and Implications for Therapeutics." New England Journal of Medicine, 392(18), 1836 to 1845, May 8, 2025. DOI 10.1056/NEJMra2413243 (pmc.ncbi.nlm.nih.gov/articles/PMC12674684).
13. Ross, J., et al. "Large-Scale Chemical Language Representations Capture Molecular Structure and Properties." Nature Machine Intelligence, 4, 1256 to 1264, 2022. The MoLFormer-XL configuration uses a 768-dimensional hidden representation across 12 transformer layers.
14. Pharmacogenomics Knowledge Base, PharmGKB (pharmgkb.org). The FDA Table of Pharmacogenomic Biomarkers in Drug Labeling currently lists more than three hundred FDA-approved drugs with pharmacogenomic information in labeling (fda.gov).
15. For a review of GLP-1 receptor agonists in metabolic disease, see Drucker, D. J. "GLP-1 Physiology Informs the Pharmacotherapy of Obesity." Molecular Metabolism, 57, 101351, 2022. Dr. Drucker, at the Lunenfeld-Tanenbaum Research Institute, is among the founding figures of GLP-1 endocrinology.
16. McCormick and Company introduced FlavorPrint publicly in 2013, with the consumer-facing platform launching in beta in March of that year. See Consumer Goods Technology, "McCormick Leaves a Mark on Consumers," November 2013 (consumergoods.com/mccormick-leaves-mark-consumers), and "McCormick Taps Big Data Platform," November 2013 (consumergoods.com/mccormick-taps-big-data-platform).
17. The thirty-three flavor dimension framework was the original public framing of FlavorPrint. See Consumer Goods Technology, "McCormick Leaves a Mark on Consumers," November 2013 (consumergoods.com/mccormick-leaves-mark-consumers), and Boston Consulting Group, "IT Innovation Leaders in Consumer Packaged Goods," 2015 (bcg.com/publications/2015/it-innovation-leaders-consumer-packaged-goods). The system’s underlying representation was extended significantly over the years that followed, encoding thousands of dimensions of taste, aroma, texture, and lifestyle context.
18. For additional contemporaneous coverage of FlavorPrint, see Consumer Goods Technology, "McCormick Taps Big Data Platform," November 2013 (consumergoods.com/mccormick-taps-big-data-platform), covering the rollout and in-store integration of the system.
Share this post
