← Future of Life Institute Podcast
Future of Life Institute Podcast
AI vs Cancer - How AI Can, and Can't, Cure Cancer (by Emilia Javorsky)
02:43:13
Loading summary
Transcript2 lines
- [00:00]
A
This is a special episode of the Future of Life Institute podcast featuring an audio narration of the just released essay from my colleague emilia Javorsky titled AI vs. How AI can and Can't Cure Cancer. This narration is provided by Eve Paseltiner, an award winning voice actor and member of SAG AFTRA. You can read the essay at CureCancer AI. At the end of the essay there is a section titled the Roadmap Forward which details exactly what must be done to accelerate cancer cures using AI. You can find timestamps for each chapter in the show Notes. Note that the original essay contains some images and lots of links that can only be accessed on the website, so view it there. For the full experience, please enjoy
- [00:55]
B
AI vs cancer how AI can and can't Cure Cancer Every year, cancer kills over 600,000Americans, our friends, colleagues and family members. It can strike suddenly and fatally. Even in those who have done everything right for their health. We know woefully little about how to prevent it. We have few effective weapons. Once it spreads, the statistics are worsening as cancer is increasingly killing young people. We must move beyond hollow promises, light on specifics which simply promise a cure for cancer. Fundamentally, a promise without a plan is a lie. We need a plan as urgent and unrelenting as the disease itself, a plan with the scale, coordination and resolve to end it. Major tech companies are racing to create artificial superintelligence. Current artificial intelligence already operates beyond human capabilities in specific domains with well defined boundaries, such as in chess or image classification. By contrast, super intelligence would be AI that substantially exceeds human cognitive capability. Across the vast majority of domains, tech companies have pitched super intelligence as the answer, making cancer cures their flagship promise. The pitch is seductive, summoning super intelligent AI genies to grant unlimited wishes like unimaginable economic growth, breakthrough treatments for devastating diseases, and reversing climate change. Think of the children they implore, pulling heartstrings by invoking illness, suffering and hope. Curing cancer is a big promise, and one that is universally considered one of the most noble and good things we can fund. But rarely is the substance of their promises examined. The logic appears airtight. If we can create systems of superior intelligence to humans across all domains, surely they will solve what has eluded our brightest scientists for decades. In 2024, Anthropic CEO Dario Amodei suggested that AI enabled biology and medicine will allow us to compress the progress that human biologists would have achieved over the next 50 to 100 years into five to 10 years, calling this the compressed 21st century. The potential for infinite benefit, we're told, justifies near infinite risks and infinite investment. Any regulatory or resource constraint that might slow this race becomes unconscionable when cancer cures hang in the balance. This approach makes a dangerous assumption that insufficient intelligence is the primary barrier to new cancer therapies. Somehow, raw computational power alone is the thing that can overcome the complex landscape of data gaps, biological complexity, regulatory constraints, and misaligned incentives that have caused billions in previous healthcare investments to fail. In fact, exponential growth in biomedical knowledge is already here. The doubling rate of medical knowledge was 50 years in the 1950s and by some estimates was down to every 73 days by 2020. Yet this intelligence explosion has not significantly moved the needle on cancer mortality or greatly increased annual new drug approvals. An abundance of knowledge and an oversupply of brilliant scientists have not moved the needle on more cures. The superintelligence narrative shapes capital allocation, policy priorities, and public expectations. The National Cancer Institute's 2025 budget, which funds most fundamental cancer research in the United States, was $7.2 billion, a mere 1.3% of the $540 billion projected total spend by private markets to build out superintelligence in 2026. The opportunity cost for medical progress is significant. The share of biotech funding is at a 20 year low. While unprecedented sums of VC dollars are directed to the development of superintelligence, the markets at least believe the hype. The obsession with the pursuit of multi billion dollar super intelligence obscures what AI can already do nearly for free. Today's AI capabilities already deliver real medical value not through big tech's pursuit of godlike machines, but through targeted solutions for specific problems. Even Google's deep minds AlphaFold succeeded by focusing on one well defined protein folding. Across pharma companies, biotech startups, and academic labs, AI is already cracking concrete bottlenecks in cancer treatment, drug discovery, toxicity prediction, and clinical trial efficiency. This is where investment belongs. But developing practical AI tools to solve problems and remove friction is fundamentally different from chasing super intelligence genies that assume raw computing power will magically cure all ills. Every complex problem demands we ask, is this bottlenecked by insufficient intelligence? If not, what's actually blocking progress? If so, can current AI solve it? Today's AI already addresses real intelligence and efficiency gaps in drug development. The deeper question Is intelligence truly the fundamental barrier to curing cancer, or are we misdiagnosing the problem entirely? In this essay, I leverage my background in both medicine and AI to examine the battle of AI versus cancer and understand who will win. AI excels at problems that are rules based or data rich. Biology has neither. AI excels in domains with complete rules, objective win conditions, instant feedback, and no physical constraints. Medicine has none of these. It offers incomplete information, stochastic outcomes, physical constraints, and delayed consequences. Wing conditions are subjective and can take years. AI can help, but history shows us that intelligence alone only suffices in domains specifically structured to reward it. Silicon Valley's mental model for progress is fundamentally shaped by Moore's law. This creates an entire culture that expects plans and budgets for exponential improvement. When tech leaders look at medicine, they instinctively assume similar dynamics should apply. But biological systems are not semiconductors. Evolution optimized tradeoffs in human biology over billions of years for robustness and redundancy. Safety mechanisms that prevent harmful mutations can also slow beneficial changes. Unlike software, where updates change systems instantly, biological interventions operate on systems that evolve to resist rapid change. In the human body, rapid cellular changes are more likely to be cancer than progress. Further, biology imposes fundamental limits on the compressibility of time. You cannot speed up a pregnancy with more engineers, and you cannot compress clinical research beyond the rate at which disease progresses in human bodies. Human biology is bounded by the timescales of cells, organisms, and populations. Even as computational power and medical knowledge increased exponentially, life expectancy gains have been linear at best, and FDA drug approvals have remained flat for decades. AI companies hang the promise of medical progress on super intelligence, amounting to a country of geniuses in a data center. Yet there is already an oversupply of human genius scientists. Now let's investigate cancer mortality rates and how they have changed over time. The Global Cancer Observatory has tracked 10 major cancer types from 1975 to 2023 and charted the rate of deaths due to cancer per 100,000 people standardized by age. In the US the patterns vary dramatically by cancer type. Lung cancer mortality peaked in the 1990s and has declined steeply since. Hodgkin lymphoma has plummeted from roughly 1 per 100,000 in 1975 to nearly zero by 2020, a genuine success story. However, most other cancers show stagnation or modest declines. Pancreatic, breast, and colorectal cancers have remained stubbornly flat or declined only marginally. This data indicates that medical progress against cancer has been selective and uneven, not comprehensive. Now let's talk about the annual number of new drugs approved by the FDA in the US from 1985 to to 2024, broken down by new chemical drugs, new vaccines, and other new biologics like antibodies and gene therapies. The numbers from our world data reveal a striking absence of growth. While there are year to year fluctuations, the overall approval rate remains relatively flat with only a slight incline, hovering between roughly 20 and 70 new drugs per year across the entire four decade period. Despite exponential increases in computational power and medical knowledge, the rate of new drug approvals has not increased systematically. Fundamentally, superintelligence does not equal super solutions grand societal challenges are rarely intelligence limited problems, but systems limited ones. The uncomfortable reality is that the primary bottlenecks to curing cancer are systemic problems and misalignments within our current power to change. Even if super intelligent AI genies existed today, their wish granting ability in medicine would be severely constrained by factors no amount of compute can overcome. Who am I and why did I write this? I I'm Amelia Javorsky, MD, MPH and I am the Director of the Futures Program at the Future of Life Institute. Throughout my career I've had the opportunity to work bench to bedside from basic science, co founding startups conducting clinical trials, medicine regulatory compliance, to public health. In 2017, I became motivated to work on how to ensure AI advances human progress and to ensure that we're not taking on risk without benefits. At this point, many in the biosecurity community, an area that overlapped with my work in public health, were thinking about the potential risks emerging from artificial intelligence. Since the launch of consumer facing LLMs and multi billion dollar AI fundraising rounds, these two previously disconnected worlds, medicine and AI, have collided with the promise of artificial superintelligence. This is the magic genie which many companies suggest will grant us cancer cures. The corporate promise is unsurprising, but what is strange is that the claim has gone largely unexamined. I have seen how a new therapy is developed, and intelligence, super or otherwise, was definitely not the bottleneck. I believe we need to break down the promise and examine both the real and exciting ways current AI tools can advance medicine and also flag what the holdups actually are to medical progress. Chapter 2 Reckoning with past failures Silicon Valley's healthcare graveyard the current promises of cancer cures must be contextualized within a long history of failed efforts by technology companies to transform health care. These failures reveal a pattern consistently overestimating the value of intelligence and technology, while greatly underestimating the complexity and data challenges and institutional complexities of medicine. IBM's Watson Health serves as perhaps the most cautionary tale. At a 2017 Health IT conference, IBM CEO Ginni Rometti told attendees that AI is real, it's mainstream, it's here. And it can change almost any everything about health care, promising to usher in a medical golden age after $5 billion invested, Watson Health was sold for parts in 2022 with minimal real world impact. As Professor Gary Marcus observed, Watson was a great system for jeopardy, but it didn't mean you could expect the same technology to solve cancer too. The fundamental problem was not insufficient AI capability, but a misunderstanding of healthcare's data landscape and clinical workflows. Microsoft's HealthVault, launched with fanfare as a platform to revolutionize health records, was shuttered in 2019 after failing to gain meaningful market adoption. Google's Life Sciences spinoff, Verily, despite $2.5 billion in investment over six years, was already being called a case study in the power and the limitations of tech industry hubris. By 2019, the company's ambitious efforts to develop comprehensive health monitoring and intervention tools floundered on the complexity of translating sensor data into clinically actionable insights. Google's longevity spinoff Calico represents perhaps the most striking example of the intelligence Calico was founded in 2013 with the explicit mission to solve aging and extend human lifespan. Radical life extension is another super intelligence promise. Superintelligence will not only cure cancer, but extend our lifespans by 1,400 years or even longer. Through some yet to be defined magic, Calico attracted top scientific talent and significant funding. The company operated in stealth for years, the assumption being that Google scale intelligence and resources applied to biology would Yield Breakthroughs. In 2025, the company publicly revealed that its lead drug candidate for ALS had had failed in clinical trials 12 years and billions of dollars later, with little to show for the investment, the failure illustrates a fundamental misunderstanding that intelligence and capital alone can overcome the biological complexity and translational challenges inherent in in therapeutic development. Haven, the joint healthcare venture from Amazon, Berkshire Hathaway and J.P. morgan, represented unprecedented corporate muscle applied to healthcare reform. If anyone could break through healthcare's ossified structures, surely it would be three of the world's most powerful companies working in concert. Yet haven collapsed in 2021, falling victim to the very perverse incentives and fragmented local structures it sought to overcome. As Professor Scott Snyder observed, given the tremendous training, process, discipline and specialization needed to deliver high quality care, the and the natural risk aversion with patient lives at stake, attempts to reinvent or circumvent the traditional care delivery system have typically failed. Even Apple's healthcare ambitions, while not outright failures, have fallen far short of CEO Tim Cook's 2019 proclamation that if you zoom out into the future and you look back and you ask the question, what was Apple's greatest contribution to mankind? It will be about health. While the Apple Watch includes health sensors and has found some clinical applications, it has not transformed health care delivery or clinical practice in the revolutionary way Cook promised. These failures share common threads Silicon Valley has repeatedly stormed into healthcare with the hubris of outsiders attempting to reinvent a system they do not fully understand, failing to learn from past mistakes. The repeated pattern suggests not mere execution challenges, but systemic misunderstanding of what's actually limiting medical progress. The AI drug Discovery Reality Check if established tech giants with unlimited resources have failed in healthcare, perhaps specialized AI companies would succeed. The story of AI Drug discovery from 2013 to present suggests otherwise. It's largely been a story of disappointment relative to the hype. The term techbio emerged to rebrand biotech, asserting the importance of the tech by putting it first. As a head of the UK Bio Industry association acknowledged in 2024, I see TechBio as a state of mind, an approach emerging in companies large and small at the interface of data, AI, life science and innovation. Companies leading the techbio movement were primarily AI drug discovery startups. Industry pioneer Recursion Pharmaceuticals, founded in 2013, centered its mission on realizing a tech bio future. It promised to use AI to discover drugs faster and more efficiently than traditional approaches. Accentia and Benevolent AI were other early leaders, attracting significant investment and generating substantial media attention with their promises of AI revolutionized drug discovery. The pitch was compelling. AI would identify novel drug targets, design optimal molecules and dramatically accelerate the path to clinical success. A decade later, the results tell a more sobering story. Following significant challenges, Accentia was folded into recursion in 2024 with the hope of creating an AI drug discovery superpower, representing an implicit admission that neither company had achieved the promised breakthroughs independently. By June 2025, Recursion had laid off 20% of its workforce. In January 2026, Recursion was trading around $4.20, far below its mid 2021 high of $40. Benevolent AI was delisted from Euronext Amsterdam in March 2025, merging with Osaka Holdings. Other long running AI drug discovery companies, including Atomwise and Relay, have experienced significant challenges. A retrospective analysis revealed that AI discovered molecules show substantially higher success rates in phase 1 clinical trials compared to historical norms. This validates AI's superior ability to find novel molecules with drug like properties. This is genuine progress and should not be dismissed. But this advantage disappears by phase two trials when drugs are tested for efficacy to see if they actually treat disease. The bottleneck is not finding promising molecules, it's predicting which will actually work in humans. As one venture Capitalist estimated in 2023, if you take the hype and PR at face value over the last 10 years, you would think it AI drug discovery goes from 5% to 90%. But if you know how these models work, it goes from 5% to maybe 6% or 7%. One recent ray of hope came from Insilico Medicine's success in advancing an AI discovered candidate for idiopathic pulmonary fibrosis, a lung disease with few good treatment options. Notably, Insilico developed AI tools to target specific bottlenecks at each phase of drug development from target identification through optimization of clinical trial design, not only at the discovery phase. Leveraging AI as a tool to manage complexity and reduce friction throughout the drug development process has emerged as more promising than the narrow focus on molecular discovery alone. While encouraging, this success is still preliminary as true success in drug development is defined by FDA approval, insurance coverage and physician confidence in favorable benefit to risk ratios for patients. Thirteen years into the AI drug discovery movement, we still lack a single FDA approved drug that cleared the full bar of regulatory approval, reimbursement and clinical adoption. The failure point is not intelligence, but the complexity of human biology, the limitations of pre clinical models, the friction of drug development and the challenges of clinical translation. Market failures for life saving Medicines the story of AI discovered antibiotics deserves special attention because it represents a controlled experiment. The science was strong enough that there should have been a bidding war to develop the drug, yet it faced significant market headwinds. In 2020, researchers published bespoke AI tools that successfully identified novel antibiotic candidates. The result was an exciting new antibiotic candidate. The scientist named Halison an homage to the original AI to go mainstream. HAL 9000 from 2001 A Space Odyssey Unlike HAL, the AI worked, the chemistry worked, the mouse studies worked. Further, compared to most drugs, antibiotics that work in mice have a high predictive value to work in humans. The clinical need is desperate, with antibiotic resistance killing an estimated 1.27 million people globally each year. But five years later, where are these antibiotics? The problem wasn't with the science, it was with the market. New antibiotics are inherently unprofitable, as they must be used sparingly to avoid promoting resistance, limiting revenue potential. Major pharmaceutical companies have largely abandoned antibiotic development programs despite depressing public health need. The antibiotic startup graveyard is instructive. Achaogen went bankrupt in 2019 despite FDA approval for plazomycin. Melinta Therapeutics filed for bankruptcy. Aradyme abandoned antibiotic development entirely. The pattern is clear. The market structure makes life saving medicines commercially unviable. The market failures are so significant, researchers had to start a social venture, fairbio, with philanthropic and government funding to generate a route to move AI discovered antibiotics forward. While AI can and already is reducing friction and cost in drug development, the biggest expenditures, clinical trials and market launch remain controlled by regulators and payers with no incentive to compromise. Market forces may block antibiotics, but surely not promising cancer therapies. Tenospamycin, which showed positive results in late stage trials for multiple myeloma and breast cancer, tells a different story. Very few promising molecules reach late stage trials and combined with encouraging data, the drug seemed prime for approval. Instead, the study was closed prematurely for resource based reasons. Bristol Myers Squibb offered no official explanation, but a lead investigator suspected drug most likely was halted because tenospamycin is difficult to produce and there's only limited time remaining before the drug's patent expires. Combined, these factors would make further investment in the drug difficult to justify. Financially killing the program was not an act of corporate malice, but a necessary balancing act. Weighing the value of a portfolio of promising drugs against market forces. Without extended patent exclusivity, companies cannot justify the enormous costs of late stage trials, FDA approval and market launch, especially when other patented drugs compete internally for investment. Tenospamycin isn't alone. By some estimates, 268 off patent drugs and 1,574 natural compounds have shown safety and potential anticancer effects. Yet regulatory and market constraints prevent the testing needed to get these drugs to patients. Current systemic constraints are already curtailing the potential of AI drug discovery capabilities and that of biology's progress more generally. The intelligence exists, the science works, the medical need is dire, and yet the drugs don't reach patients. If we can't solve this problem for current AI discovered antibiotics or promising cancer drugs, what reason do we have to believe more compute is the solution? Chapter 3 Misguiding Myths and Errors Cancer is the most complex of all diseases. Inherent barriers to a universal cure when scientists underestimate complexity, they fall prey to the perils of unintended consequences, wrote Siddhartha Mukherjee in his Pulitzer Prize winning history of Cancer, the emperor of all maladies. Cancer cure sits atop the list of superintelligence panaceas. Yet it may be the complex disease least likely to yield to any universal curative breakthrough, superintelligent or otherwise. First, most of us already have potential cancers in our bodies. Our bodies continuously produce cells with mutations that could become cancerous. Yet our innate mechanisms of cell removal and immune surveillance eliminate these threats. Cancer is not a foreign invader, but a breakdown in our own regulatory systems. What we call cancer is not a single disease, but thousands of different diseases, each a product of distinct molecular drivers, cell types of origin, evolutionary pressures, and microenvironments. As Mukherjee eloquently put it, cancer's life is a recapitulation of the body's life, its existence a pathological mirror of our own. In fact, even in a single tumor, there can be significant biological heterogeneity, meaning cells within the tumor have different pathways gone awry. Early research into long lived cancer resilient mammals such as elephants, naked mole rats, and the 200-year-old bowhead whales sparked hope of finding a singular mechanism for cancer resistance. However, again nature proved more complex. Rather than a shared compound or a universal protein target, evolution discovered entirely distinct genetic pathways to achieve resilient phenotypes. As our understanding of biology advances, cancer continues revealing greater complexity rather than simplification. The landmark 2000 paper Hallmarks of Cancer by Douglas Hanahan and Robert Weinberg identified six key capabilities that drive cancer development. By 2011, the authors expanded this to include the roles of metabolism, immune evasion, and the tumor microenvironment. Cancer was no longer just about rogue cells, but about their relationship with surrounding tissues and the immune system. The 2022 update added new dimensions, including cellular plasticity and the microbiome. The framework had grown substantially over two decades, not because the original understanding was wrong, but because cancer kept revealing new complexity and as our ability to study it improved. This trajectory represents a shift from understanding cancer as a single cell gone rogue to recognizing it as an individualized state of continuous evolution, presenting an inherent barrier to any universal a I discovered panacea. The disease adapts, develops resistance, finds alternative pathways, and fundamentally is an adversarial co evolutionary process, not a static problem. Progress in oncology has come from graduating from a futile search for universal solutions towards leaning into complexity and individual variation. Through the 20th century, cancer care focused on surgical removal of tumors and nonspecific elimination of rapidly dividing cells with chemotherapy. The genomic revolution revealed that specific genetic mutations of cancer cells drive outcomes regardless of the organ in which a tumor develops. This led to targeted therapy. Gleevik, approved in 2001, was the first drug targeting a specific genetic mutation. Recognition of the immune system's vital role unlocked immunotherapy drugs. More recently, CAR T cell therapy, personalized cancer vaccines, epigenetic reprogramming, and individualized Drug development have all emerged as our understanding of cancer has advanced. Finally, given the time for an idea in the lab to become an approved therapeutic, the science in the clinic today actually reflects discoveries from over a decade ago. AI tools are already accelerating the era of precision oncology, a fact that gets overshadowed by big tech's promise of an absolute cure as the work is often decentralized across academia, small startups and internal development programs at pharmaceutical companies, AI is helping to discover new drug targets, biomarkers predict toxicity and has helped expand the design of potential new biologics beyond their traditional set of druggable targets. AI is identifying which treatments a patient is likely to respond to, predict resistance in advance and minimize toxicity. AI is even helping surgeons in the operating room better identify tumor margins to make sure they get all of the cancer and don't need an additional operation. But all of this progress raises the question, why aren't cancer survival rates dramatically increasing? One of the strongest prognostic factors is early detection, and outside of mammograms and colonoscopies, there has been little material progress that has gone mainstream. This is partly because new diagnostics struggle to find viable business models, but also because early detection itself is medically complex. Not all early detected cancers are destined to become life threatening, creating risks of over diagnosis and over treatment. The South Korean screening program for thyroid cancer resulted in a 15 fold increase in thyroid cancer over two decades, but mortality remained stable as most of the cancers detected were ultimately small tumors unlikely to cause problems, putting patients through the risks of treatment without benefits. Conversely, broad screening can save lives, with mammography being a prime example and one recently improved by AI tools. In fact, leading physician Eric Topol has argued that the largest clinical trial ever conducted on AI in medicine now supports mandatory AI assisted mammograms to improve early detection of breast cancer. FDA requirements mandate cancer therapies are studied in populations, not individuals, imposing limits on leaning into the complexity of disease. The narrower you define a patient population, the longer the study takes burning precious patent time. Worse are the economic incentives driving pharmaceutical firms towards late stage trials. Bluntly, advanced stage cancer patients die faster, reaching survival endpoints sooner, meaning faster cheaper clinical trials. Consequently, new therapies most often meet patients in advanced stages of disease when the body is worn down and the cancer is aggressive. Medical, regulatory and financial institutions continue to grapple with financial and ethical frameworks for deploying such therapies earlier then studied in trials. But given the life or death stakes and high costs, there are few clear answers. These dynamics help to explain why we have so few treatments for the prevention and early treatment of some of the most pressing unsolved diseases. The clinical trials are long and often require a large study population to detect a significant effect. Together, these factors equate to prohibitive development costs. Let's look at the data on the average amount of time in years it takes to detect a primary clinical outcome in phase three trials broken down into different treatment areas such as cancer prevention, Alzheimer's prevention, kidney disease, stroke prevention, diabetes and more. The data shows that cancer prevention demands five to 10 years of follow up watching for cancer incidence in asymptomatic populations. Alzheimer's prevention requires four to five years tracking cognitive decline. Early stage cancer treatment requires three to five years for disease free survival, but most notably, late stage cancer treatment requires just 1.5 to 2.5 years to measure overall or progression free survival. The contrast is brutal. Preventing a disease takes years of patient follow up. Treating the same disease after it has metastasized takes months. This key difference, combined with patient clocks that tick regardless of trial length, creates an institutional structure that starves treatments focused on prevention and early intervention of resources and instead channels innovation toward end stage interventions where trials can be completed before patients expire. Now, when we compare the average number of participants per phase three trial broken down by the same treatment areas, we see the same stark institutional bias embedded in clinical trial design. Cancer prevention requires the largest trials by far, 30,000 participants on average, while early stage cancer treatment requires only 3,000. Late stage cancer treatment needs just 600 participants. Trials for late stage disease are smaller and faster because advanced stage patients experience clinical outcomes, in this case progression or death, more quickly, allowing companies to reach their endpoints and complete trials with less time and expense. Prevention trials, by contrast, must follow healthy people for years or decades waiting for disease to occur, making them prohibitively expensive and slow. The data we've just discussed analyzing phase three trials expose why pharmaceutical development gravitates toward treating sick patients in their final months rather than preventing disease in the first place. Fundamentally, the more individualized a disease and the earlier you seek to intervene, the more difficult diagnostic and therapeutic development becomes. Myth of the Eureka moment Discovery is not impact Silicon Valley often misses the bigger picture in therapeutic development by placing an erroneous equal sign between novel insight and clinical impact. The Eureka myth refers to the narrative that invention derives from a moment of brilliant insight. As Sir Harold Evans wrote in the Harvard business review in 2005, we romanticize the epiphany of of total clarity, creating neat origin stories that obscure the messy reality of innovation Eureka, Greek for I have found it, remains the motto of Silicon Valley's home state of California, which fittingly appropriates a moment of scientific discovery to commemorate the Gold Rush. Like the great founder narrative, the Eureka myth spotlights a lone genius and packages innovation into a single divine moment. This romanticism resonates particularly strongly in tech, where founding a software company can genuinely occur through such moments, a brilliant mind, a laptop, and venture capital ready to invest in scalable solutions. But therapeutic development bears little resemblance to software engineering. Discovery is not the end point, but the beginning of what most closely resembles Homer's odyssey a decade long journey tossed at sea, facing insurmountable challenges, gatekeepers and beating improbable odds to arrive at just the the right moment. Drug development requires precise intervention in the human body, a highly complex, chaotic cooperative ecosystem, all while avoiding unintended consequences. Success requires coordinated advances in clinical data generation, rigorous measurement, enabling technologies, relevant experimental models, substantial capital, evolution of scientific paradigms, regulatory policy alignment, social dynamics, and invariably, quite a bit of luck. Evans noted that while we obsess over the invention moment, the act of inventing and improving is far more often a long hard slog, and the act of capitalizing on invention and of managing the transition from a brainwave to the bustle of the marketplace is the really hard part. Derek Thompson captures this sentiment in his essay on American Progress, observing that if there were a Nobel Prize for deploying and adopting technology at scale, our legacy wouldn't be so sterling. In biology, where discovery launches rather than concludes a story, this deployment challenge is exponentially more severe. Even if a potential cure were discovered in a laboratory today, there is no guarantee we would recognize it as such, nor that it would successfully navigate translational research, regulatory approval, manufacturing scale up, reimbursement negotiations, and clinical adoption to reach patients who need it. Like we saw earlier in the case of tenespamycin, even drugs who successfully navigated most of the Odyssean voyage can be shelved steps from Ithaca. We likely already possess undiscovered life extending therapies, but they're stuck behind roadblocks and knowledge gaps in our broken system, staking hope on new technologies as rosetta stones. Overhyped promises of transformative new technology unlocking therapeutic revolutions predate big tech's arrival and the current fanfare around super intelligence. When President Clinton announced the completion of the Human genome project in 2000, he proclaimed that humankind is on the verge of gaining immense new power to heal and that the technology will revolutionize the diagnosis, prevention and treatment of most, if not all human diseases. The statement could have been written by an AI CEO today. Yet it wasn't about AI, but genomics. The project's lead, Francis Collins, speculated that Perhaps in another 15 or 20 years you will see a complete transformation in therapeutic medicine. The assumption was straightforward. Mapping the human genome would provide a Rosetta stone between disease cause and cure. If we knew each person's fundamental code, we could simply identify and debug the errors causing disease. It was software thinking applied to biology. Find the bug, fix the code, ship the cure. 25 years later, we confront the reality that this assumption was profoundly naive. Nearly all diseases proved vastly more complex than single bugs in source code. As a Scientific American review of the Human Genome Project's overstated promises, noted biologist Kenneth Weiss presciently paraphrased Tolstoy in 1993. All healthy families resemble each other. Each unhealthy family is unhealthy in its own way. The diversity, an individualization of disease mechanisms has only become more apparent as new methods of measuring biology have come online. The hope that new technologies will simplify our understanding of disease, only to find that new measurement techniques make things more complex has been the recurring story of modern medicine. At face value, this is obvious. As one zooms into sand on the beach or human tissues. Things get more complex at each increasing level of resolution. Each technological advance that promised to unlock biology's secrets instead revealed additional layers of complexity. Junk DNA turned out to be functional. Epigenetics showed that genes are regulated by more than sequence. Proteomics revealed that RNA translation to protein is far from deterministic. Metabolomics highlighted the importance of small molecules. The microbiome demonstrated that human biology cannot be understood in isolation from our bacterial passengers. The virome added another layer. And single cell sequencing showed that even the same tissue contains remarkable cellular diversity. It turns out the whole deal of biology is more complicated than the knee bone being connected to the thigh bone. For tech, where keep it simple, stupid remains a guiding principle of software development, such complexity is quite often the inconvenient truth. With each advance, disease has become more complex and individualized, not less. We may discover that we are all healthy in unique ways as well. Medicine, humbled by this ever growing complexity through lived experience of failed predictions, tends to approach new technologies with cautious optimism. Silicon Valley, by contrast, operates with its geographic and ideological bubble of software, where problems do have clean solutions and intelligence can be directly translated into value. Lacking medicines, hard earned experience with trials, failures and entrenched inefficiencies, tech often holds idealized notions of health care's challenges and opportunities. Chapter 4 AI solutions derived from first principles or data no first principles for biology on the surface, AI is ideal for complex problems. The giant asterisk lurking out of sight is that AI is ideal provided you have first principles and or robust, high fidelity data to work with. Equating increasing AI capabilities with scientific breakthroughs is correct. In domains with well understood rules of the road, like math and physics, human biology has neither ingredient. The history of medicine tends to validate H.L. mencken's axiom. For every complex problem, there is an answer that is clear, simple and wrong. From fundamental biology to clinical care, every stage of scientific inquiry involves substantial information loss and compression, limited by what is measurable, captured and communicated. Human biology is neither an engineered system nor fully deterministic. It is the product of billions of years of evolution. Human biology was not designed, but learned through cycles of death and survival. Silicon Valley often sees biology as a complicated engineering problem rather than an incredibly complex evolved system. In engineering, if you've an incredibly challenging problem, like getting a rocket to Mars, you turn to physics. Once you've got the basic rules and building blocks, you have everything you need to build at any scale of ambition. Hence, it is understandable that AI's potential to model biology's basic building blocks, molecules and cells, may be mistaken as providing tools to effectively solve biology at any scale. How wrong this is. Human bodies are composed of approximately 30 trillion cells, each following local rules that give rise to emergent properties at higher levels of organization. Further, as one travels across layers, from physics to chemistry to biology, and then from cells to tissues to organs to organ systems to entire bodies, emergent properties arise that are not fully predictable from lower layers. Biologist Michael Levin argues for reframing biology as collective intelligence, given that functionality of higher layers, I.e. organs, is contingent on the ability of lower levels, I.e. cells, tissues to cooperate with each layer evolved to solve different problems that are inaccessible to layers above or below it. Physicist Stephen Wolfram introduced the concept of computational irreducibility to highlight that in some complex systems, the only way to reach fundamental answers is through simulating each step. For some problems, like biology, there are no shortcuts and no compression algorithms to capture the essence of the system. Trying to solve human diseases like cancer through pure computation is akin to trying to solve climate change by simulating every molecule of air. Even if biology were governed by physics alone, simulating human biology from first principles is physically impossible. Simulating just one week of a single full human, 10 to the 28th atoms using classical physics with GPUs covering the entire Earth, running at their theoretical thermodynamic limit would take the age of the universe. Even simulating a single second would be impossible on a timescale relevant to humans. Worse, classical physics is unlikely to be up to the task, and simulating the quantum physics correctly is exponentially harder, requiring either an impossible amount of classical computation or a quantum computer orders of magnitude larger than we can presently build. While advances in AI simulation capabilities may unlock exciting insights into molecular and cellular biology, there remains an enormous gap between molecular simulation and human simulation. We cannot compute the fundamental truth of biology we can only measure it. However, this does not mean that we abandon the role AI can play in improving our ability to simulate biology, improve our measurement abilities, and discover potential rules of the game, a task scientists are working on. But it does require us to acknowledge that biology is noisy and stochastic. Even in a perfect simulation, one random molecular collision or mutation could be the difference between health and and disease, given the interconnectedness of biological systems, posing a fundamental limit on our ability to predict clinically relevant individual biology from physics data desert and biomedical epistemics. If you don't have first principles to work from, then you need data where intelligence limited bottlenecks do exist in biomedical research, they are likely solvable with current AI capabilities provided sufficient high quality relevant curated data, which is itself often contingent on innovation in measurement techniques. These critical measurement and data caveats are often glossed over in discussions of AI's potential. Google DeepMind's successes with AlphaFold and Gencast illustrate this point perfectly. Both solved long standing problems that had stumped researchers for decades, validating AI's capability to accelerate scientific progress in intelligence limited domains. The secret ingredient in both successes? Decades of curated high quality data sets spanning global scientific collaboration. AlphaFold was trained on the Protein Data Bank, a database of protein structures maintained by a global consortium of scientists since it was first announced in 1971. Representing over five decades of careful data curation, Gencast relied on 40 years of weather data from the ERA5 data set. These weren't data sets scraped from the Internet or assembled ad hoc. They were the product of sustained institutional commitment to data generation, standardization and sharing in a particular domain in science. The potential of AI tools is likewise fundamentally contingent on having relevant, representative, high quality data from both the bench and the bedside. It is reasonable to infer that DeepMind chose to tackle protein folding and weather prediction precisely because such data existed. Fields where foundational data is broadly available and well curated, like chemistry and materials science are primed for AI acceleration. By contrast, biology and medicine, where data is siloed, outdated, incomplete and often locked behind corporate walls, face far greater challenges to realize the benefits of AI intervention. Further, medicine is full of tacit knowledge that instructs clinicians when to look beyond the data, with being able to walk into a room and determine if a patient is sick, not sick, being a key tenant of medical training. Patient Data in applying AI to medical data, patient records would seem the most logical place to start. Most medical data represents highly compressed, coarse representations of underlying biology, retrospective symptom descriptions, a few hundred standardized blood tests, and medical imaging all functioning as inaccurate proxies for the biological processes being studied. Electronic health records, the primary source for much medical AI, were not even designed to collect health data, but rather to extract and structure information from providers to increase billing revenue, optimizing for sickness, not biological reality. As ambient AI scribe systems marketed to restore the human side of medicine by automatically taking notes are in reality being used to upcode visits and maximize revenue extraction. The clinical accuracy of electronic health records data is only likely to worsen. Treating messy, incomplete data that systematically misrepresents clinical reality as ground truth will not yield reliable insights and may even teach systems that chronic disease maximization is the goal. Like AlphaFold's ability to leverage the Protein Data bank, we recently had a successful example in medicine of how a large, high quality clinical data set can change the game. The UK biobank tracked genetic, lifestyle and environmental factors in over half a million volunteers with the goal of collecting 30 years of longitudinal data. Since the release of the first batch of data, applying AI has led to breakthroughs in early diagnosis, clinical care, drug discovery and precision medicine. Major medical insights historically have come from large scale, structured longitudinal studies of healthy patients, such as the Framingham Heart study started in 1948, which revealed that reducing high blood pressure and lipids was key to reducing cardiovascular mortality. Similarly, the Nurses Health study, started in 1976, taught us about the dangers of trans fats and the role diet plays in driving type 2 diabetes. Beyond scale, time and accuracy, the UK Biobank is arguably the first longitudinal population scale data set incorporating modern measurement techniques. Data collection went far beyond the few course measurements routinely collected in clinical care to include state of the art biology measurement techniques ranging from whole body imaging to proteomics and metabolomics. The project leaned into the arsenal of modern measurement technologies, adding in new tools as they came online to generate torrents of data about human biology at resolutions unimaginable even a decade ago. Mapping precise measurement of biological states to clinical states will be key to unlocking the potential of AI in an otherwise computationally irreducible system. Such measurement is vital to move beyond unrefined descriptions of symptoms and precisely understand what is health, what is disease, and the variance between individuals. Yet outside of the structured research setting, we've barely begun to integrate genomics, proteomics or metabolomics into routine practice. The standard annual clinical panel includes roughly 135 routine blood tests, a laughably limited snapshot. Even worse, while some tests are reliably stable with little variants, many have wide variation, making it difficult to interpret a single result in a medical record or compare between tests taken at different time points. In one case study, a healthy young man had blood work done after intense exercise and weightlifting. Initial results suggested he had fatty liver disease, yet when repeating those same tests at rest, they were normal. Even novel modalities like organ age estimation, longevity clocks vary significantly across the day and almost 30 years of variation depending on what tissue type is studied. AI medical impact is constrained by our failure to measure biology better and refine methods to distinguish signal from noise. These are ultimately prerequisites for defining and diagnosing disease with precision, while routine implementation of next generation measurement into clinical care is not feasible given the current state of the United States health care system. Spinning up more data moonshots like the UK Biobank is within reach. Such data moonshots are expensive but are a drop in the bucket relative to expenditures on the race for superintelligence. The UK Biobank is transforming medicine with the 303 million pounds 413 million US dollars invested so far. Yet the projected spend on the super intelligence race for 2026 alone is estimated at $540 billion. In more tangible terms, if the total AI spend was $1,000, the entire UK biobank would amount to a mere $0.77. If the AI buildout is to cure disease, shouldn't a fraction of that be invested in the data generation required to solve the problem? As medicine becomes more personalized and precise, the data requirements scale faster than our ability to collect them. No amount of AI capabilities can extract signal from data that was never captured. Intelligence cannot substitute for the measurement of real human beings. The literature the insufficient data problem extends beyond capturing quality clinical data. Much of what we think we know from biomedical research is likely wrong. We just don't know which parts the replication crisis represents more than a quality control problem. It's an epistemic catastrophe that poisons the entire knowledge base on which medical AI would be trained. When 70% of researchers report failing to replicate colleagues findings, and 68% of papers lack adequate raw data for reproduction. We're not dealing with occasional errors, but systemic knowledge corruption. AI systems trained on this corpus will confidently learn the patterns in published research, which include failures like publication bias, career incentives, and institutional politics mixed in among the biological truths in medicine A basic science finding can be true and replicable in cells and mice, yet clinically insignificant in humans. A common myth is that cells in mice reliably predict human response, and that's why they are used as models. The truth is, they're simply the best tools we have, but they're not great. When asked whether he could cure cancer, pioneering cancer scientist Dr. Judah Folkman famously replied, yes, in mice, we can cure cancer readily in cells and mice. Yet 97% of those mouse cures failed to deliver benefits to humans, with the main failure model being lack of efficacy. Given the low predictive value, biological insights about human disease derived from or engineered into these models are often problematic. The classic example is Alzheimer's research, where researchers spent decades developing and testing drugs to clear amyloid from mouse models. Based on the amyloid hypothesis that buildup of this protein is what causes disease, the mouse models worked, yet 99.6% of the drugs failed to meaningfully improve the disease in humans. The emergence of the erroneous amyloid hypothesis was not surprising, but its dominance over decades was. It highlighted how scientific inquiry can be as limited by institutional gatekeepers as governing an overton window as it is by available pre clinical models. The questions allowed to be investigated and which assumptions can be challenged are controlled by a peer review publication and NIH grant process that largely relies on legacy incumbent researchers to award funding and decide which papers get published. The result has been an institutionally low ambition process that funds fewer out of the box ideas in favor of incremental science, making science less disruptive over time. AI has supercharged this misalignment, advancing scientists careers at the expense of exploring novel and breakthrough ideas. Dissent is also quashed. There were scientists who questioned the casual assumptions behind the amyloid hypothesis, but they were largely sidelined by what journalists dubbed the amyloid mafia, institutional incumbents defending established dogma they helped to create in the first place. Admitting when we're wrong is vital to the health of science and medicine. Medicine has repeatedly discarded paradigms once considered definitive. From bloodletting to spinning chairs, medical history is peppered with countless therapies abandoned when evidence proved them ineffective. Yet rather than breeding humility, institutional incentives have pushed medicine toward a culture of authority doctor knows best. The result is heightened barriers to questioning assumptions and pressure to project certainty about what remains nuanced and incomplete. One of the sharpest disconnects between Silicon Valley and the medical establishment is epistemic confidence. As both tech and the public believe medicine knows far more than it does. Medicine's failure to communicate the limitations of existing knowledge informed doubt has eroded trust and created a vacuum filled by anyone with an opinion uninformed doubt. That vacuum fills with questions and theories, much of which is warranted truth seeking, but some of which is opportunistic proliferation of unscientific content. While speculation might not end up in the literature, it inevitably ends up polluting AI models scraping the Internet Complicating matters, scientific literature also suffers from severe systemic biases. Firstly, scientific publications only reward the publication of impactful positive results, leaving the negative results and the many experimental iterations required to obtain a significant signal in biology absent from the literature base. In fact, the value of most of the science being done is is lost. While the NIH has transitioned to mandating the use of electronic lab notebooks internally, externally in academia, scientists are still using paper based data capture. We desperately need massive amounts of comparable raw biological data, yet most experiments are recorded in formats and silos that render them lost to the scientific commons. When AI is trained on a literature base riddled with missing data, irreproducible findings, uncertain clinical relevance and bias toward incumbent thinking, what confidence can we have in its outputs? The problem isn't the algorithm, it's the foundation it's built on. The fix requires mandated data capture, standardized reporting and sharing requirements, tighter links between pre clinical science and clinical outcomes, and sustained investment in strategic data generation rather than research that merely validates what institutions or already believe. ARPA H, the DARPA for Health offers a promising model that isn't beholden to incumbent gatekeepers. The agency aims to fund high risk, high reward science that breakthrough biomedical progress requires hard limits to data generation if data is mission critical. Beyond improving capture and availability of the data already being collected, how can we scale up our capacity to generate more? The current basic biology equivalent of the UK Biobank is the Human Cell Atlas, a project to map all human cells in the body. The NIH is already funding the Brain Initiative, a large scale coordinated effort to map the human brain. This is exactly the kind of work NIH should pursue less incremental grants to legacy academics and more investment in the high quality large scale data sets that build the scientific commons. Following in the tradition of the Protein Data bank, the project begins with Identification of what high quality curated data set is missing to advance science, AI is often put forth as the answer to scale data generation in biology. Robotic automation is already improving standardized processes such as protein engineering, DNA synthesis, car T manufacturing, high throughput screening and quality control. However, discovery oriented experiments are often artisanal because iterating and inventing new methods is part of the breakthrough. This requires tacit knowledge, technical mastery and real time problem solving that resists automation as scientists cannot codify what they're observing and discovering for the first time. In biology, data generation is not brain power limited, it's resource limited. There is already a significant oversupply of highly trained scientists relative to available research funding. In the biomedical sciences, there is a single professor position for each 6.3 PhD graduates, an oversupply that cannot be accommodated by a biotech industry in the doldrums. Unlike software engineers who can build with a laptop, biology is a resource intensive endeavor with experiments needing to be conducted on a physical lab bench with expensive materials, cutting edge capital equipment and a finite amount of hard assets available for experimentation. Improving the scalability and precision of pre clinical work is vital, such as developing virtual cells to better simulate cell biology. But the core challenge remains. Most cell biology has limited predictive value for human clinical outcomes. The best pre clinical models dramatically improve prediction by using human samples, organs on chips with human primary cells, patient derived xenografts that implant tumor samples onto immunocompromised mice, and phase 0 trials that test microdoses in patients to assess safety and target modulation. However, all of these approaches face a hard finite human resources. Patient cells, tumor samples and clinical trial volunteers willing to participate without therapeutic benefit cannot be scaled. They can only be allocated wisely. While AI can generate thousands of potential targets, we need robust mechanisms to identify which ones justify the allocation of scarce patient resources. Realizing the potential of AI to accelerate biomedical research requires radically scaling and improving allocation of our data generation infrastructure. Precisely at a time when both public and private investment are declining, we must identify data gaps and fund coordinated moonshot efforts to fill them. This means resisting the temptation to ignore what's costly and time consuming, but essential for progress rich longitudinal human data sets and novel methods that push the boundaries of biological interrogation. AI won't solve the intensifying competition for dwindling grant funding and venture capital among academic institutions, nonprofits and startups. On the contrary, it's actually pulling money from all other markets like a black hole. Until we generate the experimental and clinical data needed to better model a computationally irreducible system. Even the most powerful AI will have limited impact. Filling these gaps takes time decades for clinical data, years for experimental data, and we haven't yet begun to resource this work in earnest. Chapter 5 Systemic bottlenecks and Misalignments the Cost of Narrow Optimization Silicon Valley's exceptional success has been built on an obsessive commitment to optimization of narrow metrics. Identify a key metric correlated with profit, such as user engagement, search queries or conversion rates, and align all company activities or around maximizing it. The tech bio hubris around discovering a cancer cure is arguably a byproduct of this philosophy. Warning lights should be blinking red in applying this culture to biomedical science, as narrow optimization in complex adaptive systems trades short term gains for long term cascades of unforeseen consequences. The cautionary tale is social media in optimizing for user engagement, narrow AI algorithms successfully drove profits but also left behind increased rates of depression, impaired cognitive development in youth, erosion of social trust, and the spread of misinformation. From big tech's perspective, this approach proved extraordinarily profitable and the negative externalities were borne by users and society, not the companies. Such positive feedback creates a powerful reinforcement function for companies to apply this same process to the next domain of disruption. Medicine and human biology represent far more complex systems and social media platforms. The human body exists in a delicate homeostatic balance. Zero inflammation is as pathological as excessive inflammation. 0 lipids are as dangerous as too many. Driving any biological parameter to an extreme typically causes harm. Yet the tech industry's muscle memory is to pick an optimization function and drive it to maximum. This paradigm, so successful in software, becomes dangerous when applied to evolved biological systems characterized by dynamic balance, redundancy and context dependent regulation. The narrow optimization problem has already triggered a death spiral in the United States healthcare system long before Silicon Valley showed up. Insurers optimize for denying claims to maximize profits. Hospitals and physicians optimize for delivering more care, not better outcomes because they're paid fee for service. Pharmaceutical companies optimize for drugs people take chronically rather than cures since recurring revenue is more valuable than one time payments. Regulators optimize for avoiding high profile failures rather than enabling beneficial innovation since punishment for approving a harmful drug far exceeds any reward for accelerating a life saving therapy. Each stakeholder rationally optimizes their narrow objective function and the cumulative result is a system that fails patients. Survival rates remain stubborn, patients struggle to access care, and those who can risk financial bankruptcy as long as the players benefiting from misalignments are different from those experiencing the negative externalities, incentives for change remain weak. Applying AI, particularly opaque systems, biological or systemic, and optimizing for unclear objective functions risks supercharging existing misalignments rather than fixing them. AI tools have already been shown to increase billing through more aggressive coding rather than improving care, which risks driving an already struggling system off a fiscal cliff. In considering the concept of aligning to desired outcomes, it's worth noting that medical AI faces uniquely difficult challenges. Whose definition of health should AI optimize for a patient or a hospital administrator? When a treatment extends life by three months at the cost of severe side effects, who decides if that trade off is worthwhile? These aren't edge cases, but the everyday reality of medical decision making, where values, preferences and context determine what care actually means. The complexity of biological systems, the stakes involved in health care decisions, and the existing misalignments in the system all argue for extreme caution in deploying a culture of narrow optimization focused AI. We need AI designed with an understanding that biology seeks dynamic balance, not maximum values. We need systems with transparency and interpretability so we can understand what's being optimized and to catch perverse outcomes before they compound. And most critically, we need to fix the underlying incentive structures before deploying powerful optimization technologies that will amplify whatever objectives the system currently rewards, whether those objectives serve patience or not. Regulatory and Institutional Barriers Having spent the duration of their life cycles largely free from liability under Section 230, and as they are currently advocating for federal AI, Amnesty Silicon Valley is uniquely ill prepared to confront the significant regulatory constraints governing the transition of successful science into approved therapy. Even when intelligence could accelerate discovery, the path to clinical impact requires navigating formidable regulatory and institutional structures. These structures exist for important reasons, protecting patient safety and ensuring efficacy, but they also create bottlenecks that computation alone cannot overcome. The FDA is thought of as an agency designed to make sure new drugs are safe and effective prior to being sold in the United States. But as with all things, the devil is in the details. Close examination reveals a 20th century agency ill equipped to manage accelerating scientific understanding. The result is an increasing gap between what we could do to help patients based on the state of the science and what is legally available and accepted by medicine to help patients. The FDA's requirements have evolved through layers of well intentioned administrative expansion. The modern FDA, established in 1938, initially required only that drugs prove their safety. The 1962 Kefauver Harris Amendments added efficacy requirements to prove that drugs actually work. Our understanding of what is a disease, what is effective, and what is safe has evolved dramatically since 1962 and lacks the objectivity desired by the regulations. What constitutes a disease? Cancer classification increasingly incorporates genetic mutations alongside traditional organ based categories. What were once grouped as rheumatic illnesses are now recognized as distinct autoimmune conditions with different mechanisms. Dementia has fractured into multiple distinct conditions, yet regulators often still treat these as monolithic categories. A drug for Alzheimer's disease is tested as though Alzheimer's is one thing when emerging evidence suggests it may be several distinct pathological processes requiring different interventions. Conversely, the FDA refuses to recognize aging as a disease despite growing evidence for shared molecular mechanisms underlying numerous age related conditions. Modern biology reveals that disease categories are both too broad, lumping together distinct pathologies and too narrow, separating conditions that share root causes. Yet the regulatory apparatus cannot adapt to either. The concepts of effective and safe are equally problematic when applied uniformly. Efficacy is currently defined by endpoints selected through negotiation between the FDA and pharmaceutical industry, not by what degree of improvement actually matters to patients lives. A drug might meet its primary endpoint of extending progression free survival by weeks while failing to improve quality of life or vice versa. Beyond this, efficacy is determined by mean effects, ignoring massive heterogeneity in treatment responses. A drug that fails its primary endpoint may work remarkably well for a subset of patients who will never gain access to it. Safety is similarly treated as objective when it's inherently personal. Patients with the same diagnosis face identical regulatory risk benefit determinations despite varying personal circumstances, values and risk tolerances. The 2006 Physician Labeling Rule took things even further, requiring a detailed mechanism of action where companies must explain how their drug helps patients, not just if it helps. Each layer was intended to protect patients, but together they have created an environment where rational drug design has become limited to crafting compelling regulatory narratives. More than the objective biological truth, the current regulatory paradigm demands answers to narrow what is the target? What is the molecule? What is the mechanism? What is the outcome? Given the complexity of human disease and our incomplete understanding of biology, the fact that we've developed any effective treatments at all within these constraints is remarkable. The entire possibility space of multiple targets, synergistic molecules working in concert, multiple potential mechanisms, or acknowledging incomplete mechanistic understanding while still delivering clinical benefit is functionally off limits from a regulatory perspective. Yet many highly effective drugs still in use today were approved before mechanism of action requirements, and we still do not fully understand how they work. Valproic acid for seizures, lithium for bipolar disorder, guaifenesin for cold and flu symptoms. Some of our most successful drugs were discovered through irrational observation or serendipity during development rather than rational design. Sildenafil Viagra was designed for hypertension but proved better at treating erectile dysfunction through mechanisms not originally anticipated. Minoxidil Rogaine had a similar story for hair growth. GLP1 drugs like Ozempic were thought to work on the stomach but primarily act on the brain, a complete inversion of the rationalized mechanism during development. What breakthrough therapies are we missing by demanding premature mechanistic certainty? How many potential treatments or languish because they don't fit into neat regulatory boxes? What is the unrealized potential of current science to save lives and reduce suffering? AI's ability to identify novel multi target approaches, synergistic drug combinations, personalized therapeutics or treatments that work through incompletely understood mechanisms is already constrained by regulatory frameworks designed for simpler paradigms. Again, Big Tech is likely making the erroneous assumption that the lack of biomedical progress is due to scientific shortcomings, naive to the amount of science already bottlenecked by an outdated sclerotic robot regulatory system. Rethinking Paternalism Ultimately, the potential of contemporary science and an AI supercharged future is contingent on a paradigm shift. The FDA embodies a mid 20th century worldview of biomedicine as producing static, objective, uniform truths where knowledge is stable enough to encode in regulations and apply universally. Modern biomedical science reveals the opposite. Our understanding is provisional and rapidly evolving, and the frontier of therapeutics is moving toward personalization rather than universal treatments. In the FDA's defense, it is trying to modernize, but the pace of change is woefully insufficient relative to the pace of science. A regulatory system designed for slow moving population level knowledge cannot accommodate a science that regularly discovers new disease subtypes and increasingly attempts to predict individual treatment responses. The mismatch is not just about who decides, but about whether centralized decision making can possibly keep pace with the speed and granularity of current discovery. Policy already acknowledges this reality in extreme cases. Right to try Laws recognize that terminally ill patients can access experimental treatments without FDA approval after exhausting standard options. If we trust dying patients to evaluate experimental therapies when they have nothing to lose, why not trust patients with adequate information to make treatment decisions earlier when interventions might actually work? The answer lies in mid 20th century paternalism when the presumption was that doctors know everything and always know what's best for you. At the time, such paternalism was defensible when science was stable, information asymmetries were insurmountable and patients had accessible primary care doctors. Benchmarking to an old idealized standard of care most patients never experience is intellectually dishonest. In reality, millions of Americans lack regular physician access, let alone specialists offering personalized recommendations. The relevant comparator isn't a world class specialist. It is more often no care, an overwhelmed doctor with 15 minute appointments or insurance company algorithms. Further, the information gap that justified regulatory paternalism is narrowing with AI's democratization of medical knowledge to patients and democratization of accelerating scientific knowledge to providers. Medical progress requires a combination of regulatory updates and a rethinking of what decisions should be made by centralized regulators and what decisions should be made by providers and empowered patients. Such a rethinking is going to require more than tweaking FDA rules. It will require the development of of high quality, trustworthy AI tools to empower patients. Physicians will require retraining as scientific interpreters rather than protocol followers. Equally critical is liability reform so providers do not bear malpractice risk for individualized decisions that deviate from population averaged standards of care. Limits to Compressing Time There are non bureaucratic factors limiting the speed in which science can be adopted, which is the time needed to see if something works. Clinical trials themselves present timescales that are difficult to compress. On average, it takes 10.5 years for a drug to move from phase one through regulatory approval. For drugs entering phase one, 90% will fail somewhere along the pathway, with lack of clinical efficacy representing 40 to 50% of failures and safety concerns another 30%. This is not a problem of insufficient intelligence in trial design. It is the inherent challenge of safely testing interventions in humans on biological timescales. Some companies report compressing preclinical workflows from three to six years to 18 months. But this creates a misleading impression. While AI's role in accelerating drug discovery sounds like a 90% improvement to the public, the reality is more modestperhaps 10 to 20% time savings because you're only radically condensing the initial preclinical phase of drug development. More importantly, any increase in drug candidates inevitably collides with a fixed bottleneck, the limited pool of patients available for clinical trials already a major challenge with existing pipelines. Clinical trials are harder to compress because they operate on the timescales of human biology and regulatory requirements. AI can eliminate friction in the process and tighten the feedback loop between scientific discovery and clinical application, but compression is ultimately constrained by the pace of human health and disease itself. The AI Paradox let's compare the traditional drug discovery funnel with a Hypothetical AI powered drug discovery funnel. AI can radically compress the discovery phase, but it cannot compress the timescales of human biology or regulatory requirements. Accelerating input into a fixed bottleneck does not accelerate output, it simply creates a backlog. For example, a traditional drug discovery funnel starts by investigating roughly 10,000 compounds. After a three to six year discovery and preclinical phase, the funnel gets dramatically narrowed down to only five to 10 of those compounds surviving to enter clinical trials. Phase 1, 2 and 3 will consume another six to seven years. Finally, one drug is submitted to the FDA, which takes a further one to two years to review the result. One approved drug entering the market. After all that time with an AI powered drug discovery funnel, the initial process is accelerated, the discovery and preclinical phase is dramatically compressed to just one to two years, and the number of candidate compounds under consideration is multiplied by 10 times to 100,000. And yet, the number of drugs which enter clinical trials remains the same as in traditional drug discovery, which is still only 5 to 10 compounds. And every subsequent stage remains the same as it is in the traditional pipeline. The clinical trials phase remains unchanged at six to seven years, and the FDA review still takes one to two years. The critical constraint, the systemic bottleneck, is human biology and regulatory requirements. Unless the clinical trial infrastructure itself expands proportionally, AI driven discovery will produce more candidate drugs than the system can evaluate, leaving the ultimate throughput approved therapies reaching patients largely unchanged. Yet some super intelligence proponents suggest that their genies will cure most diseases in five to 10 years. The evidence cited for this assertion is the speed of the COVID vaccine approval. This was approved in less than a year, and it's suggested that this demonstrates long timelines are due to FDA bureaucracy more than biological limitations. This comparison ignores crucial context. Moderna was founded in 2010 and submitted its first investigational new drug application for an MRNA flu vaccine in 2016. By the time Covid struck, the company had nearly a decade of foundational research, safety data and manufacturing infrastructure already in place. The nine month timeline counted only from COVID emergence, not from when MRNA technology development began. Covid also presented unique trial advantages that chronic diseases fundamentally lack. With a novel virus spreading throughout the general population, where almost everyone is due to be infected, trial participants were easy to find and recruit at scale. Outcomes were clear and rapid. Did people get sick, require hospitalization or die within weeks or months? Some researchers even proposed human challenge trials in which healthy volunteers would be deliberately exposed to the virus. An ethically defensible approach when community transmission meant most people would likely be infected anyway. Such trial designs can produce results within weeks. By contrast, chronic diseases like cancer or Alzheimer's require far longer and more complex trials to know if something is working or not and for whom outcomes are slower to manifest. Trial readouts for cancer treatments often require years of follow up to assess survival benefits. Recruitment is dramatically harder since researchers must identify and and enroll patients with specific disease subtypes rather than drawing from the general population. Disease heterogeneity means that subgroup analysis are essential but require larger sample sizes. The logistics of biological timescales, long term follow up, and patient retention over years present challenges that no AI can simply optimize away. The speed of COVID vaccine approval also reflected unprecedented political will to find a cure. Operation warp speed mobilized resources not only to save lives but also to stabilize the health care system and reopen the economy. Government funding de risked investment, regulatory agencies provided real time guidance and accelerated review and manufacturing scale up began at financial risk before approval. That sense of urgency and resource mobilization does not exist for chronic diseases. Despite their enormous cumulative burden. Cancer kills more than 600,000Americans annually, yet commands nothing close to the focused governmental response Covid received. There are inherent limits to acceleration of FDA review. The FDA Modernization Act 2.0 has already enabled new methods of pre clinical testing, such as microdosing human trials and organs on chips to compress timelines between drug discovery and safety testing in humans. The AI tools are improving clinical trial patient selection and creating synthetic control groups to try to increase the productivity of clinical efficacy trials. AI is also enabling the creation of digital twins, a key enabling feature for conducting individualized trials. N1 clinical trials even if the FDA were to revert to approvals based on the demonstration of safety, the question as to whether a new drug actually moves the needle for an individual patient remains bounded by the pathophysiology of the disease, similar to moonshot investments in clinical data collection. The critical question for new drugs is how to know reliably if an intervention is working before a clinical outcome. In essence, how can we compress clinical time scales? For some conditions, like infectious disease, time cycles are short, but many diseases play out over decades. The ability to know whether something is working is especially salient as medicine transitions to preventing disease in high risk individuals rather than treating sickness. There are promising methods in this area, like complex biomarker development. However, it takes time to see if these surrogate endpoints match clinical outcomes. You need at least one full cycle of disease pathophysiology to get results. Further, in a system that values treatments far more than diagnostics. There is again a resource constraint on both the capital and patience required to develop the tools ultimately needed to compress biological timescales. Follow the money who pays for what? In the absence of understanding the incentives of the United States health care system, one would assume that cures to diseases, especially diseases such as cancer, would be invaluable and certainly worth the cost. But in a system that profits from treating diseases, a healthy patient represents lost revenue. As a Goldman Sachs analyst famously asked, is curing patients a sustainable business model? Gilead Sciences developed a cure for hepatitis C, but analysts watched drug revenues plummet from their peak of 12.5 billion in revenue as the number of patients was exhausted. Consider CAR T therapy for cancer, priced at $475,000 for a one time treatment, or Zolgensma for spinal muscular atrophy at $2.1 million. These aren't arbitrary prices. They're essentially estimates of the equivalent lifetime value of chronic treatment alternatives. The pricing explicitly assumes cures must capture the same total revenue as ongoing disease management. The system has encoded a preference for sickness over health. Similarly, when doctors and hospitals are paid based on how much care they deliver service rather than health outcomes, value based care the health care system thrives economically. When people remain sick, early diagnosis and prevention eliminate the procedures, tests and prescriptions that generate revenue, creating a fundamental misalignment between profit and patient welfare. Financial rewards in pharmaceutical development are equally decoupled from eventual clinical outcome. As the typical exit path for small pharmaceutical companies is acquisition by a large pharmaceutical company. This is often before regulatory approval or even large scale clinical trials. This creates systemic pressure to hype early findings and operate in domains with high acquisition appetite, resulting in herding effects where innovation tends to cluster around popular targets and therapeutic areas. Further, when incumbents purchase promising candidates, they integrate them into existing pipelines subject to existing incentives, or simply shelve them if they threaten profitable drug franchises. The science behind the current blockbuster GLP1 drugs was already known by the late 1980s, and a decision by Pfizer to end exploration with the startup exploring the science in 1990 set the field back decades. This path dependency ensures that even genuinely revolutionary AI discoveries get filtered through the same misaligned incentive structures the technology was supposed to bypass. The pharmaceutical industry's relationship with the FDA is clouded by the appearance of sophisticated gaming as opposed to neutral scientific evaluation. Companies employ former FDA officials who understand agency priorities and informal requirements. They design trials to meet established regulatory endpoints rather than answering meaningful clinical questions. This isn't necessarily malicious. It's a rational adaptation to incentive structures. But adding AI to this system risks enabling more sophisticated gaming rather than genuine innovation. A super intelligence optimist may argue that this the technology can also solve such systemic problems. This reveals the fundamental faith underlying superintelligence promises that sufficient intelligence equals power. Market and regulatory structures don't yield to smart arguments. An AI might propose perfect healthcare policy, but implementation requires navigating Congress, state legislatures, corporate lobbying and entrenched interests. While these incentive misalignments are unlikely to stand in the way of a cancer cure, they do shape the probable paths of capital allocation by companies whose fundamental goal is profit. This prompts reflection as to whether the company's promising AI genies have a history of transcending incentives to advance the betterment of humanity, or whether they are likely to follow the money like anyone else. Chapter 6 Conclusion the seductive promise of superintelligence curing cancer allows us to avoid confronting uncomfortable truths, namely that our scientific institutions have structural problems limiting innovation, that our regulatory frameworks lag behind biological understanding, that incentives are fundamentally misaligned with patient welfare, and that we lack the data infrastructure to fully leverage AI's potential. Waiting for Team Tech Bio's intelligent genies absolves us from the collective responsibility to build, reform and innovate. Today, when patients are suffering and dying each day, bankrupting families to access basic care, this abdication is unconscionable. We have examples of what medical progress actually looks the UK investing in the biobank and making the data freely available to researchers globally the FDA approving the first gene therapies for sickle cell disease after 30 years of foundational research operation warp speed demonstrating that removing financial risk and regulatory uncertainty can compress timelines when urgency is real. The Framingham Heart study, started in 1948, teaches us the fundamentals of cardiovascular disease prevention. The pattern with all of these examples is sustained institutional commitment, coordinated data infrastructure, strategic de risking of valuable research and long time horizons. These successes came from systems building, not intelligence explosions. The prosaic truth is that the path forward is not through intelligence explosion, but through hard work. Edison had it right. We need 1% inspiration and 99% perspiration. Specifically, that means generating better data through sustained institutional investment, reforming regulatory frameworks to enable innovation while maintaining safety realigning commercial incentives with medical objectives and empowering domain experts to build AI tools matched to specific problems they understand deeply. This requires confronting the reality that grand challenges are rarely intelligence limited. Grand challenges are data limited, regulation limited, incentive limited, and coordination limited. Current AI capabilities thoughtfully applied by people who understand biomedical problems deeply can help. They can accelerate therapeutic development, reduce trial timelines, improve preclinical models, and help restructure broken incentive systems. We see this potential being realized in distributed fashion across universities, startups and pharmaceutical companies using current AI as a tool for specific applications. But realizing this potential at scale requires abandoning the intelligence illusion and embracing the complex, unglamorous work of fixing systems that actually stand between discovery and cure. The gap between AI's advertised potential and its limited real world impact on medical progress is not evidence that we need more intelligent AI. It's evidence that intelligence is not the binding constraint. We mistake incentive problems for intelligence problems, data problems for computational problems, and coordination failures for capability limitations. The result is a dangerous complacency. Rather than undertaking difficult institutional reforms, we wait for the techbio bros to save us Superintelligence may eventually arrive, but cancer patients cannot wait for genies when the real barriers are human made systems. We could reform today if we had the will and the permission to do so. The question is not whether AI can help cure cancer. Current capabilities already can, serving as one tool among many in researchers arsenals. The question is whether we will undertake the systemic changes needed or just continue to mistake the promise of future intelligence as a substitute for present action. The Roadmap Forward how can doctors, patient support groups and AI developers integrate AI for positive change in cancer treatment? And what can be done to accelerate cancer cures? AI vs. Cancer explores why artificial superintelligence won't deliver on the cancer curing hype, and details the many data, economic, systemic and institutional challenges that bottleneck progress. But if superintelligence isn't the answer, where do we go from here to actually have a chance at defeating cancer? And how can AI help? This roadmap outlines the work that is already being done towards this problem, but also highlights the real issues that we need to address if we are going to cure cancer. This roadmap is also an invitation to everyone who hears it or reads it on our website to help us to refine the ideas and to collectively bring momentum to the issue. Cancer has proven an intractable enemy for over 600 million years since multicellular life appeared. If we want to change that, then it will require input from every sector of humanity to bring ideas, skills and the demand for political and societal changes. Why now is the right time? Beyond the urgency to tackle cancer, given the human cost of patients dying every day, we sit at a unique point in modern history that makes going on the offensive against disease more promising than ever. US healthcare has long been trapped in what Scott Alexander calls a Moloch problem, a coordination failure where every stakeholder sees the system collapsing, yet no one can escape the dynamics driving it down. Hospitals are closing, insurers are in death spirals, pharmaceutical Companies face a $236 billion patent cliff by 2030. Science funding is waning, doctors are burnt out, and patients can't access care. All while health care consumes 27% of federal spending and threatens the nation's fiscal and geopolitical footing, it's clear that tinkering with reform at the edges is no longer viable, economically or politically. The same resignation haunts cancer, specifically after Nixon's 1972 moonshot and Obama's in 2016, mortality rates have barely budged, breeding a fatalism that curing disease is simply unsolvable. But there has been no better time to shake off fatalism and go on the offensive. Ironically, a system in late stage collapse also creates the conditions to rebuild Culturally, within US politics there is an appetite for radical institutional disruption, with polling during elections showing 83% wanting substantial change or complete upheaval. Further, existing AI capabilities are already set to radically disrupt the workforce and incumbent businesses. Whether for the better or worse remains unclear. What is clear is the combination of collapsing healthcare, political appetite for disruption, and the availability of AI's tools to reduce friction, align incentives and manage coordination at scale creates a once in a generation window to explore new avenues that may at last enable us to develop a concrete path to better health care and a genuine plan to cure cancer. Step 1 support and scale AI Tools to Accelerate Cancer Cures the AI capabilities that we have right now contain immense promise to accelerate science. There are countless companies, nonprofits and academics hard at work developing AI tools, often coupled with novel methods of measuring human biology to tackle key bottlenecks in oncology. From the Cancer AI Alliance's development of a federated AI framework to empower researchers to learn from de identified patient data to the National Cancer Institute's digital twin development, Nonprofits, large pharmaceutical companies, small startups, government institutions and academics are all hard at work on AI tool development. While Big Tech has tried to claim the mantle of building, these are the true builders we should be celebrating. Resourcing and scaling. Following is a sample of some of the key actors working at various points from bench to bedside to unlock the power of AI to deliver real benefits to patients. Examples of entities developing, using and advocating for AI tools to cure cancer AI drug discovery and target identification Developing AI tools to find disease targets and design drugs to block them faster than Traditional lab methods Developing AI tools to read and analyze genetic data at massive scale to uncover drivers of cancer and disease AI Proteomics and biomarkers Develop AI tools to map the proteins the body produces to uncover early or hidden signals of disease presence, progression and treatment response AI Toxicity prediction Developing AI tools trained on compounds already known to be safe or harmful Predict whether a new compound will be toxic in humans in advance AI in silico Modeling and virtual cells Building AI powered digital replicas of cells and biological systems to simulate experiments and better study basic biology AI Drug repurposing Using AI tools to mine existing approved drugs and natural compounds for hidden potential against diseases. AI Clinical trials and regulatory affairs Using AI to accelerate, improve and redesign how trials are run and approved by replacing slower conventional methods AI early detection Liquid biopsies and imaging Using AI to process blood scans and tissue samples to detect cancer earlier than current methods AI Clinical and surgical Bringing AI into the clinic and operating room to guide decisions, improve precision and personalize care delivery Advocacy for AI tools in oncology Leading cancer organizations pushing for AI driven advances to reach patients through policy, funding and clinical adoption. Again, you can find a list of organizations working on these AI tools on our website. CureCancer AI roadmap Step 2 Double down on the most promising areas for progress in oncology early detection and prevention. We need to scale up proven screening tools like mammograms and colonoscopies, expand cancer preventing vaccines and use policy to tackle risk factors like smoking, obesity and toxic chemical exposures. We should fund large studies to test whether new blood tests that screen for multiple cancers actually save lives. And we need better tools to catch cancer returning or becoming resistant to treatment before it takes hold again. Data and Clinical Trials Cancer research is hobbled by fragmented data. We need a national system linking patient data across genomics, imaging and treatment outcomes. Clinical trials should be redesigned to test multiple drugs at once rather than one at a time. And a non profit drug accelerator could push promising but commercially unattractive treatments through the approval process. Improving Research and treatment we should keep pushing innovation in radiation therapy, surgery and next generation drugs while making these treatments more widely available. Manufacturing breakthroughs could slash cost dramatically, bringing some cutting edge cell therapies from 50 down to under $20,000. We also need large coordinated research efforts to better understand cancer at the molecular level and more attention to the long term health of cancer survivors. Metascience and Institutional Reform Our research institutions need reform. Funding agencies should take bigger bets on unconventional ideas rather than safe incremental work. We need to fix the reproducibility problem in pre clinical research. Drug payments should be tied to whether patients actually get better, not just whether they receive treatment. And we should pursue drug pricing reform so that a cancer diagnosis doesn't also mean financial ruin. Tackle the main blockers to medical progress that are limiting AI tools Despite there being promising areas within oncology and AI tools being built, both will fail to meet their full potential when operating in current structural constraints that quietly determine their ceiling. Success in these domains is contingent on solving deeper data and structural problems. In addition to asking what we should build, we need to ask what is stopping us from building it. An attempt to highlight the top blockers limiting AI's utility in advancing cures for disease and medical Progress more generally. Blocker 1 data creation 1. How do we create data to build a real time map of what healthy actually looks like? Reliable, high resolution definition of disease requires reliable high resolution definition of health. Until we know what deviations in biology are normal and which drive disease, we cannot tease the signal of disease from that of health. We need funding and coordination for more population scale, longitudinal, multiomic baselines for human health across age, sex, ethnicities, geographies and environments. In parallel, we also need to push the frontier of clinical measurement capabilities and integration of established methods into clinical practice. 2. How do we create data to compress biological timescales across health and diseases? How do we know something is working in an individual before the disease plays out? This is the highest leverage and most valuable technical problem in medicine. If we can validate surrogate biomarkers and real time biological readouts, we could both collapse decade long trials into years, making drug development for prevention and early treatment financially viable. It would also enable individualized medicine at scale. This will require at least a single cycle of human disease to validate, but will position us well for the future. Further, even without full validation, construction of well found surrogate biomarkers is better than the vacuum of information guiding treatment. Today, confidence in a surrogate can be built incrementally based on biological plausibility and retrospective correlation with hard endpoints even before prospective validation is complete. 3. How do we generate data to build preclinical models that actually predict human outcomes? Cells and mice fail us. The hard limit on experimentation is our inability to study human biology reliably without directly studying humans or their primary cells and tissues. But we are still far short of that limit. Progress starts with optimizing experimental capabilities for for human derived organoids, patient derived xenografts and phase 0 microdosing trials, all of which dramatically improve predictive value over standard cell and mouse models. Ultimately innovating in and scaling our capacity to measure human biology is the deeper solution explored in question 1. 4. How do we build AI models trained on comparable data rather than publication bias? How can we capture and integrate negative results, failed experiments and unpublished data? Essentially, this is solving the garbage in, garbage out problem in distinct ways. The first is capturing what's missingnegative results, failed experiments and unpublished data that currently disappear into filing cabinets. Mandated electronic laboratory notebooks with enforcement for all federally funded research is one concrete mechanism. The second is making existing data comparable across labs and experimental iterations, which require standardization of reagents protocols, batch correction and reporting. The field has pursued solutions for decades with limited success, suggesting both technical and incentive structure difficulties. Tackling both begins with rewarding meaningful data generation, not just positive results. Blocker 2 Economics and Incentives 5. How do we make preventing or curing disease more profitable than treating it chronically? Models such as value based care and outcome based pricing take on new meaning in the age of AI, which can help to democratize data collection and improve both transparency and analysis of patient outcomes. Further political will for payment reform and elimination of administrative waste with AI could generate savings that can be reinvested in experimenting with different reimbursement models. As discussed in the why Now Is the Right Time section, AI enabled outcome tracking the fiscal crisis forcing change insurance death spirals, hospital bankruptcies and an imminent patent cliff create a unique window for experimenting with new models. 6. How do we fund the therapeutic development that markets won't fund but need to succeed? Markets reliably fail to deliver potentially life saving treatments to patients from antibiotics to rare diseases. The antibiotic crisis showed that philanthropy can step in when markets fail, but philanthropy alone isn't sustainable. The more durable answer lies in changing the economics. As AI and next generation measurement techniques like multiomics bring down the cost of clinical trials and potentially enable N1 studies, the financial barrier to researching commercially unviable treatments falls significantly. This opens the door to a broader funding ecosystem. Government grants, ARPA H style high risk bets, single payer systems with different incentive structures and philanthropic endowment models designed to generate returns that sustain the work beyond the initial gift. Decentralized science community governed research funding coordinated through digital infrastructure such as decentralized autonomous organizations is an emerging addition to this ecosystem and improving the economics of clinical research may finally unlock its ability to scale. Blocker 3 Institution Systems and Coordination 8. How do we modernize the FDA to build a regulatory framework that is personalized, not population averaged? The 20th century FDA was designed for a simple world of one disease, one drug, one mechanism, one target, one population averaged result. That world no longer reflects the state of biology as diseases fracture into molecularly distinct subtypes. A drug that works remarkably well for 30% of patients may fail its population level endpoint or pass it while delivering negligible benefit to the majority. Either outcome highlights a regulatory framework mismatched to science. FDA modernization requires at least three shifts. One from binary to conditional approval with mandatory real world evidence collection 2 from population averaged endpoints to adaptive enrichment designs that identify responder subgroups during trials rather than after three from genetic disease categories to molecularly defined indications. None of these are radical and in fact versions of each exist within current FDA authorities but are chronically underused. The harder question addressed next is evidentiary. Personalized approval requires personalized evidence, and generating that without massive trial sizes remains an unsolved problem modernization must confront honestly rather than ignore. 9 How can we accelerate development without compromising safety? The tension between acceleration and safety can be partially resolved by sequencing. The current model demands certainty before approval but then learns relatively little afterward. A better model lowers the bar for initial approval based on early biological signal, but raises the bar for mandatory confirmatory trials with hard endpoints, real world evidence collection, and rapid market withdrawal when confirmatory data fails. Accelerated FDA approval pathways already embody this logic, but weak confirmatory trial enforcement has undermined and eroded public trust. Regulators using the carrot more than the stick gives the appearance that acceleration may simply mean a faster way for drug makers to profit. Public skepticism is reinforced by the revolving door between FDA reviewers and industry. There are real unsolved challenges in the balancing act between development and safety. Shorter trials mean less power to detect rare adverse events, which can only be observed at scale. No approval model fully resolves this further. Ultimately, safety is a deeply personal question. Different people have different risk tolerances yet are constrained by a single risk tolerance. FDA's respecting individual autonomy while protecting against corporate exploitation of patient desperation is the genuine design challenge. 10 how do we prevent AI from supercharging the existing misalignments rather than fixing them? AI accelerates existing incentive structures, as we saw with AI scribes upcoding under the guise of improving the patient experience. Even good intentions are insufficient protection when the underlying system requires rewards the wrong outcomes. AI's turbocharging ability accelerates the speed asymmetry where misalignments compound far faster than institutional reforms. Early governance and prophylactic reforms are vital wherever possible, as explored earlier, rapid technological disruption hitting a system in late stage collapse creates genuine political and economic conditions for reform that have not existed before. Seizing this window requires us to get specific about what gets rebuilt and in whose interest and what objective functions we choose to optimize A new system for these decisions will determine if U.S. health care accelerates off of a cliff or if it begins a redemptive journey of repair. An Invitation to Change as stated in the introduction, this roadmap is an invitation. The conclusions reached here are deliberately incomplete. It is meant as a stimulus for discussion and feedback, not a comprehensive plan. Much of it may be misguided or wrong. What's missing from current discourse is not diagnosis of the problem, which is abundant, but solution oriented generative thinking. Ideas are too often killed by criticism before they're properly examined, when iteration and collaborative troubleshooting would serve us better. Even if AI progress stopped today, the disruption already underway will force us to reimagine how we work, live, and organize our institutions. Meeting the moment will require us to move beyond listing grievances into a mode of generative criticism where the question is not why an idea is wrong, rather what are the positive elements to amplify and identify how it could be different or better. In that spirit, I welcome feedback on what warrants further development in this plan or how plans could be improved, better prioritized, or more complete. A follow up piece will incorporate this input in the months ahead. We need your expertise. If you are a researcher, clinician, technologist, policymaker, patient advocacy organization, philanthropist, or other stakeholder with a perspective to share, we'd love to hear from you. Your input will help us develop a practical roadmap for AI and cancer research. Sign up to access a private feedback survey and receive invitations to curated workshops where we'll bring the roadmap to life. As a reminder, you can find our sign up page and more detailed information on our website at. Curecancer AI Roadmap.