Transcription
Greetings everyone. Today, I would say, we have a truly deep dive. We will be dissecting and thoroughly examining the ideas of Dr. Eric Topol. He is, well, one of the key figures at the intersection of artificial intelligence, genomics, and, of course, personalized medicine. Yes, hello. And our topic today is, frankly, captivating. How this very AI has transformed healthcare, our understanding of longevity, and what awaits medicine, well, let's say, in the foreseeable future. At the core of our conversation today are excerpts from Dr. Topol's recent speeches, his articles, and, of course, his new book "Deep Medicine" (or "Superagers" as mentioned later). So, our source of information is, well, very rich. Absolutely. These are not just fragmented thoughts. This is the result of many years of experience and research. Topol is a cardiologist, a geneticist, a researcher. He sees the picture from different angles. And his book "Superagers" is precisely an attempt to understand how the science of aging and technologies like AI can work together. And our task today is not just, you know, to retell his words. We want to study each assertion in detail. every example, every idea that he expresses. We will delve into the nuances, try to trace the connections, connections between technologies, education, clinical practice, ethics. All of this is very intertwined. Absolutely. Not just to state and this is important, but to understand why it is important right now, what specific opportunities it opens up, and, no less importantly, what problems and challenges it creates. We will, so to speak, unpack all of this layer by layer. Excellent. Well then, let's begin. And I want to start with a rather interesting, almost personal moment. How does Dr. Topol himself manage to do so much? He is known for his productivity. He himself, it seems, jokes about it, calling himself an infovore, almost a sleepwalker. Yes, he is indeed very productive, publishes, speaks, heads an institute. And behind this joke about the infovore, of course, lies the question: how does he cope with this flow of information? And here, I think, lies the answer. He actively uses modern tools, including AI-based tools. So, he is not just a theorist, he is also a practitioner in using these technologies for his own work. Absolutely. And he emphasizes that most people, well, perhaps even among our listeners, do not fully realize how extraordinary the opportunities that AI already provides are. He gives examples that sound almost like science fiction, but they are reality. What examples, for instance? Well, he says: "Imagine you need to generate a complex scientific report, summarize a mass of data. It used to take weeks, months. Now AI can create a draft, a very high-quality draft, in minutes. Or, let's say, analyzing huge arrays of individual data. Genome, lifestyle, dynamic analyses, to develop a personal strategy for healthy aging. It's simply physically impossible for a person to grasp." He mentioned specific tools that are used. Yes, he spoke about Chat GPT, about Claude from Anthropic, about Google's NotebookLM, which, for example, can take several PDF files, say, transcripts of his speeches, and create connected text based on them, for example, a podcast script. He describes these capabilities as mind-blowing, as something that is beyond comprehension for many to this day. And this mind-blowing aspect, it leads us to a very big and, frankly, a little frightening topic of artificial superintelligence, AGI. How does Topol relate to this? He refers to the famous economist Tyler Cowen, who boldly stated: "What if this is it already? What if we are already living in the era of AGI, we just don't call it that." Topol himself seems more cautious with terms. He says: "It's not so important what we call it, AGI, advanced AI, or something else." What's important is to realize the scale of the phenomenon. It's already a lot. It's already changing the rules of the game. It changes the rules of the game. And where should these rules change first, if we are talking about medicine? Probably at the very beginning, in the training of doctors. Absolutely. This is one of the key and, I would say, most painful points for him. How should AI change medical education? He recently spoke at Mount Sinai Medical School and reflected a lot on this topic. And what conclusions did he reach? Very radical ones, actually. He directly states: "Medical education in its current form, based on memorizing vast amounts of facts that are constantly changing and that AI knows better than any human, is obsolete. It simply does not prepare doctors for the future that has already arrived." So, cramming anatomy and biochemistry is no longer the main thing. Not that it's not important at all, but the emphasis should shift. The most striking thing, in his opinion, and he uses a rather strong word, "embarrassing," is the almost complete absence of AI courses, data work, and digital medicine in the curricula of most US medical schools. Yes, and not just the US, I think. But not all doctors should be programmers or data specialists. No, of course not. It's not about that. It's about doctors understanding how these systems work, what their capabilities are, and where their limitations are, what an AI "black box" is, why it can produce biased results, how to critically evaluate its conclusions, how to integrate AI into their practice so that it helps, not hinders, and certainly does not harm the patient. This is basic literacy, which will soon become as necessary as the ability to read an X-ray. And he goes even further. He proposes to revise even the criteria for selecting medical school applicants. Yes, and that is also a very bold thought. He says: "Perhaps it's time to stop focusing exclusively on GPA scores and standardized test results like the MCAT in the US. Perhaps it's time to actively seek people with strong human qualities, empathy, communication skills, humanistic values, because these are precisely the qualities that AI cannot yet replicate, and these will become key in the future doctor-patient interaction, where routine tasks will be handled by machines." So, the doctor of the future is more of a person for a person, not just a carrier of knowledge. In a sense, yes, a carrier of knowledge is also a doctor – it's someone who knows how to apply that knowledge correctly to a specific person, taking into account their unique situation, their values, their fears. And someone who can provide that very human support. But for this to become a reality, the training system needs to change. And what is the current situation with the understanding of AI among, let's say, healthcare leaders, clinic administrators, officials? They see this potential. According to Topol, we are at, well, the very, very early stage. There is interest, but it is mainly limited to, you know, back-office tasks. Schedule optimization, billing, insurance coding, perhaps some operational improvements, that is, what does not directly concern patient treatment. Largely, yes. There is also interest in so-called ambient systems. This is when AI listens to a doctor's conversation with a patient and automatically fills out medical documentation. This is, of course, important for reducing the burden on doctors, burnout issues, but this is again more about optimization. And what about clinical application? Diagnostics, treatment selection. And here, in his opinion, there is still a big gap. There is no broad systemic implementation of AI specifically in clinical practice. This is especially true for multimodal systems that can simultaneously analyze different types of data: medical images, text records, laboratory tests, genomic data. Their potential is colossal, but they reach real practice very slowly. And one more aspect he highlighted is empowering patients with AI. This is also thought about very little, practically not thought about. According to him, healthcare systems are in no hurry to give patients AI-based tools for managing their health, for better understanding their condition. The focus is still on the doctor as the sole carrier and interpreter of information. And this, as we will see further, Topol considers a big omission. Okay, if the potential is so huge, if there are already tools, if even Topol himself actively uses them, then why is the implementation in widespread medical practice so agonizingly slow? He uses rather sharp words, a sclerotic, ossified medical community. What is the main snag? Yes, he really doesn't mince words. And the key problem, as he sees it, is, paradoxically, the lack of convincing clinical data, data that would prove the benefit of AI not in laboratory conditions, not on selected datasets, but in real, chaotic, everyday clinical practice. That is, large randomized studies are needed that would show: here is the group with AI, here is the group without AI, and the outcome is better. Exactly. And there are catastrophically few such studies so far. He gives a very vivid, simply a textbook example of using AI in mammography for breast cancer screening. Tell us more, it's important. A huge study was conducted. Almost 100,000 women participated in Sweden. They compared the standard approach: two radiologists look at the images, and the new approach: one radiologist plus AI as a second opinion. The results were, well, very impressive. How impressive? The combination of AI plus radiologist detected 20-25% more clinically significant cases of cancer. That is, not some microscopic changes, but precisely those that require treatment. At the same time, the workload for radiologists decreased by almost half, because AI filtered out clearly normal images, and the radiologist focused on suspicious ones and worked faster. It would seem, here is the proof, take it and implement it. Logically. And what about reality, for example, in the USA? And in reality, says Topol, in the USA, AI is almost not used in mammography. Moreover, there is a company, Radnet, a large network of diagnostic centers, which offers mammogram analysis, but for an additional fee. That is, a woman has to pay extra from her own pocket for a technology that can potentially save her life or improve diagnosis. Wait, this seems somehow wrong, to put it mildly. Topol believes it's terrible, it's a very bad sign for the future. If we start charging patients extra for using AI, which should be the standard of care, it creates inequality and hinders progress. Instead of making better diagnostics accessible to everyone, it is turned into a premium service. It's some kind of paradox. Technologies are developing at breakneck speed, AI capabilities are growing exponentially, and their path into real clinics turns out to be incredibly long and thorny. Exactly. And while this large, unwieldy healthcare system reflects, weighs, conducts endless pilot projects, something interesting is happening on the other side, from the patients' side. What exactly? That patients, ordinary people, are gaining access to the most powerful AI models through their smartphones, through the internet. The same GPT, Claude, other models are becoming publicly available. And stories are starting to appear about how people use these tools to analyze their own medical data. That is, people take their medical records, test results, upload them to a chatbot and ask: "What does this mean?" Well, something like that, yes? Topol gives an example that was described in the Wall Street Journal. A person with a chronic illness, with a disability, who had a lot of data from sensors, doctors couldn't figure out the cause of his condition's deterioration for a long time. He took all this data, uploaded it to Anthropic's Claude model, asked the right questions, and AI helped him find a pattern, identify the problem that specialists couldn't handle. Wow, that's impressive. But this also raises a lot of questions, including legal ones. Absolutely. And Topol mentions the opinion of a lawyer specializing in medical malpractice. The lawyer asks: "Look, if there is an AI tool that can objectively improve diagnosis or help a patient, and a doctor doesn't use it or even doesn't know about it, isn't this, in the long run, grounds for an accusation of negligence?" Before, this wasn't the case. Google searches, of course, helped patients find information, but it couldn't make a complex diagnosis or analyze individual data at this level. Exactly. But now the situation is changing, and AI models can analyze complex relationships, find non-obvious correlations. It's becoming increasingly difficult to ignore this tool. And here Topol returns again to criticism of the medical system. What is it this time? In the unwillingness to give patients full and convenient access to their own medical data. Yes, patient portals exist, but often they are inconvenient, the information there is fragmented, and it's difficult to download it in a format suitable for analysis. He calls this an intolerable imbalance of power. The system, as it were, says: "This is our data, not yours." And, accordingly, the healthcare system does not encourage patients to use AI to analyze this data. On the contrary, it rather hinders it. There are no initiatives to teach patients how to safely and effectively use AI for their health. Although, paradoxically, there are already examples of FDA-approved, that is, regulator-approved, AI tools intended specifically for patients. What examples, for instance? The classic example is smartwatches like the Apple Watch. AI-based algorithms in them can diagnose atrial fibrillation, a dangerous arrhythmia. There are even studies showing that analyzing data from watches can predict the risk of sudden cardiac death. This is technology that millions of people already have on their wrists, and it can save lives by working on the patient's side. It turns out that driving progress in the use of AI in medicine through patients, through their gadgets, and through publicly available models may be easier and faster than through the conservative system of clinics and doctors. Topol leans towards this idea. Downward pressure from informed and technologically savvy patients can become a powerful catalyst for change. And, by the way, he notes that even regulators are starting to understand this. A recent initiative by CMS, the Centers for Medicare and Medicaid Services in the US, to collect information via RFI on how to improve data exchange and expand patient access to it, is mentioned. So, the ice is perhaps starting to break. Okay, let's shift the focus a bit. Speaking of patients, especially those who actively care about their health and longevity, a service like a full-body MRI scan is gaining popularity. Companies like Ezra, Prenuvo, Function Health offer this to healthy, often wealthy people, promising early diagnosis of everything. How does Dr. Topol view this? His position here is very clear and quite tough. He is categorically against such an approach for healthy, asymptomatic people. In his book "Superagers," he directly calls it a very bad idea. Here's why, it would seem, the earlier you find a problem, the better? Because, according to him, it's a practically guaranteed recipe for getting an avalanche of false positive results. The human body is a complex thing, and with such total scanning, some minor deviations, cysts, nodules, anomalies will almost always be found, which have no clinical significance, will never turn into a disease, but which will frighten the patient and trigger a cascade of further investigations. That is, a person gets a diagnosis of something unclear, and it begins. additional tests start, often invasive and unsafe biopsies, for example. Topol gives examples from practice. A patient after such a preventive MRI developed liver bleeding during an unnecessary biopsy. Another patient developed a pneumothorax, a collapsed lung, after a biopsy of a nodule found in the lungs, which turned out to be benign. Sounds unpleasant. He quotes a colleague who called such a practice predatory. Why? Because it plays on people's fear of cancer and aging. It is not properly regulated. And, most importantly, there is no scientific data confirming that such total screening of healthy people brings more benefit than harm. That is, it is not proven that it actually prolongs life or improves its quality in the general population. He gives some other examples. Yes, he tells the story of the famous writer Drew Culhar. He had the opportunity to get a free full-body MRI as part of a promotion for one of these companies. An anomaly was found in his prostate. Not cancer, but something suspicious. And now he is forced to undergo prostate biopsies every six months, living in constant fear and discomfort, although he had no symptoms and still has none. That's the price of such early detection. Yes, the prospect is not great. What, then, is a reasonable alternative according to Topol? How to correctly use modern technologies for early diagnosis, especially of cancer? His approach is deep personalization. No one-size-fits-all. It all starts with a thorough, comprehensive assessment of individual risk. What does this assessment include? All available data about the person. First, their electronic medical record, not only structured data, diagnoses, lab numbers, but also unstructured doctor's notes, medical history. And AI can extract valuable information from there. Second, the dynamics of laboratory indicators, even if they are within the so-called normal range, trends can be more important than absolute values. Third, of course, genomics. not only searching for known pathogenic mutations like BRCA for breast cancer, but also assessing polygenic risk, the cumulative effect of many small genetic variations on predisposition to a particular disease. Fourth, epigenetics, aging clocks, which show the biological age of different organs and systems. For example, through DNA methylation analysis. Okay, all this data is collected, the risk is assessed. And then what? And then, if based on the totality of all these factors, the risk of developing a specific type of cancer, not cancer in general, but, say, lung cancer or pancreatic cancer in this person is indeed high, then it makes sense to use more targeted screening methods. What examples, for instance, the so-called liquid biopsy. This is a blood test for circulating tumor DNA or other cancer biomarkers. This technology is now rapidly developing. If the liquid biopsy is positive, meaning traces of cancer are found, but it's unclear where exactly it is located, only then, perhaps, it makes sense to do a full-body MRI or PET-CT to find the primary focus. That is, a full-body MRI is not the first step, but rather one of the last and only for a very specific high-risk group. Exactly. The current approach of mass screening of healthy people using MRI without considering individual risk, he compares to driving cattle to slaughter, "treating people like cattle," equally thoughtless, wasteful, and potentially harmful. And this idea of personalized prevention based on deep risk analysis is the central idea of his book "Superagers." Absolutely. The main goal, according to Topol, is not some fantastic task of reversing aging or becoming immortal. The goal is much more pragmatic. To prevent or maximally delay the development of the three main groups of age-associated diseases, which cause most deaths and poison old age. These are cancer, neurodegenerative diseases, Alzheimer's, Parkinson's, and cardiovascular diseases. And AI is needed here as a tool to integrate all these diverse data – genetics, epigenetics, analyses, lifestyle – to identify this risk as early as possible and act precisely. Exactly. And this is the very integrator that can see the whole picture, assess the complex risk, which no doctor can do, simply because the volume of information is too large. Okay. So, diagnostics and screening are clear. What about lifestyle? We all hear general recommendations. Eat right, move more, sleep enough. How effective are they, in Topol's opinion? He is quite skeptical of such universal "one-size-fits-all" advice. Not that they are wrong, no. Moving, eating well, sleeping – these are undoubtedly important. But he believes that such general recommendations are not very effective for real motivation of people. Why? Because they are depersonalized. They do not take into account individual characteristics and, most importantly, individual risks. What truly motivates people to change their habits, often quite established ones, is knowing their own personal risk. It's one thing to hear that smoking is harmful in general, and quite another to see your genetic markers that increase your personal risk of lung cancer tenfold. So, personalization is needed not only in diagnostics but also in prevention, in lifestyle changes. Absolutely. It is precisely knowing one's risk profile that gives a person a reason to act. And this, according to Topol, opens the door for truly effective primary prevention, that is, preventing disease before it begins. Before, this was largely a fantasy, a pious wish, because there were no tools for accurate risk assessment and personalization of recommendations. Now they are appearing. What specific elements of lifestyle does he highlight as most important, especially with age? Well, first, it's physical exercise. And not just one thing, but a combination. Aerobic exercise for the heart and blood vessels. Strength training to maintain muscle mass, which is critically important with age, and balance exercises to prevent falls, one of the main causes of injury and loss of independence in the elderly. Second, nutrition. Here the emphasis is on an anti-inflammatory diet. Fewer ultra-processed foods, less red meat, more vegetables, fruits, whole grains, healthy fats. Again, with adjustments for individual characteristics. Some people suit the Mediterranean diet, others another. And third, sleep. Quality sleep, especially deep sleep. He emphasizes the role of sleep in clearing the brain of metabolic waste, including beta-amyloid, the accumulation of which is associated with Alzheimer's disease. But again, all of this should be personalized, based on risks. Yes. And here, biomarkers and AI come onto the stage again. They help not only to assess risk but also to track the effect of lifestyle changes, which provides strong motivation. Can you give an example of how biomarkers motivate? He gives a very clear example with the risk of Alzheimer's disease. A blood test for the biomarker PTA 217, phosphorylated tau protein 217, is now available. Its level can be elevated 10-20 years before the first symptoms of dementia appear. This is a very early risk marker. And what can be done with it? The most interesting thing is that studies show that the level of this marker responds to lifestyle changes, especially regular physical exercise. It can be reduced, and significantly, by 40%, 50%, even 70% in some people. And this reduction can be tracked with repeated analyses. That is, a person sees the concrete result of their efforts in numbers. This is precisely what provides powerful feedback and motivation to continue. Topol compares this to tracking LDL cholesterol levels, bad cholesterol, in people taking statins. When you see the numbers decreasing, you understand that you are doing everything right. This can also work with neurodegeneration risk markers. And for each of the three major groups of diseases – cancer, neurodegenerative, and cardiovascular – there are their own strategies, their own markers that can be tracked. Yes, this is the essence of personalized prevention. For cardiovascular risks, these can be advanced lipid profiles, inflammation markers, assessment of arterial calcification. For cancer, genetic risks, liquid biopsy in high-risk groups. For neurodegeneration, markers like PTA 217. We are living, according to him, in a unique time, a time of convergence, when the science of aging, AI capabilities for data analysis, and the emergence of new biomarkers are coming together. This truly sounds encouraging. Let's talk about medications now. The topic of creating new drugs with the help of AI is now on everyone's lips. How does Topol assess this potential? He sees great potential here, of course, and AI can significantly accelerate the drug development process at the very early stages. Identifying new targets for intervention, predicting the efficacy and toxicity of candidate molecules, even designing completely new molecules with desired properties. These are all areas where AI is already being applied and where breakthroughs are expected. Are there any prominent examples of drugs created from start to finish with AI? Not yet, according to him, there are no such truly prominent examples where one could say: "This drug was completely created by AI, almost." But he is confident that they will appear in the coming years. However, he suggests looking at this topic from a slightly different angle. Using the example of the very well-known class of drugs, GLP-1 receptor agonists. These are Ozempic, Victoza, Mounjaro. Yes, drugs for treating type 2 diabetes that have revolutionized obesity treatment. There is a whole chapter dedicated to this in his book. Absolutely. And the history of these drugs, in his opinion, contains a very important lesson, including for understanding the role of AI. What is the lesson? The lesson is that 20 years ago, when these drugs were just being developed by Novo Nordisk, initially for diabetes treatment, no AI would likely have suggested actively researching their potential for weight loss. Moreover, there were significant doubts within the company about this. Why? Because in clinical trials in diabetic patients, weight loss while taking these drugs was quite modest. Nothing revolutionary. It took the persistence and determination of a specific person, a scientist named Mads Krogsgaard Thomsen, to convince the company's management to conduct separate studies on obese people without diabetes. And what did those studies show? They showed a stunning result. People without diabetes lost much more weight on these drugs, 15%, 20%, and even more of their initial body weight. It was a real breakthrough. But why such a difference in effect between diabetics and non-diabetics is still not fully understood. There are various hypotheses, but there is no complete understanding of the mechanisms. That is, AI, based on the data available then, would likely have missed this potential. Quite likely. It would have seen a modest effect in diabetics and possibly not recommended further costly research on another population. Human intuition, persistence, and perhaps luck played a decisive role here. This shows that biology is still full of mysteries and not everything can be predicted by algorithms. But now, can AI help understand the mechanisms of action of these drugs? Yes, now AI can certainly be useful for data analysis, for finding correlations, for generating hypotheses about mechanisms of action. But the very fact that such a breakthrough class of drugs appeared largely thanks to an illogical human factor is an important reminder of the limitations of AI, at least at the current stage. At the same time, Topol notes that these drugs themselves, the gut hormone agonists, are just the beginning. Dual and triple agonists affecting multiple receptors simultaneously are now being developed. Oral forms are appearing, other hormones with similar effects are being studied. These substances affect not only appetite and glucose metabolism but also the brain, inflammation, and the immune system. And they are generally well-tolerated. That is, this is a harbinger of a new era of pharmacology. He believes that yes, an era where we will use the body's own regulatory molecules to treat complex chronic diseases. And in this new era, AI will undoubtedly play a huge role in accelerating development, in finding new combinations, in personalizing therapy. But we should not forget the role of humans, and that biology is more complex than we sometimes think. It's interesting, how does Dr. Topol himself use AI in his daily life or practice? What does he advise?
He speaks quite simply: "I use everything that works and that helps me and my patients improve their health. He doesn't get stuck on any single technology. His main obsession, as he himself puts it, is prevention, the prevention of diseases. And he sees AI as a powerful tool for achieving this goal. And how does he feel about bold statements, like the prediction by Demis Hassabis from DeepMind that AI will help end most diseases in the next 10 years? Topol agrees with the general direction of thought that AI can fundamentally change the situation with chronic diseases, but he believes that 10 years is, well, too optimistic a timeframe. It will take longer. But the goal is precisely this: to break the trend, to achieve a state where most people in old age are not just alive, but healthy and active. He uses the term "elderly, healthy elderly." That's what we need to strive for. And AI is a key ally here. You know, during the discussion, another topic emerged that I find absolutely amazing, even unexpected. It's the use of AI for empathy and social support. A graph was presented showing that people are increasingly turning to chatbots not only for information but also for therapy, for communication, even in search of the meaning of life. Was this predictable at all? This is perhaps the biggest surprise for Topol himself, and he honestly admits it. When he wrote his previous landmark book, "Deep Medicine," several years ago, he was absolutely convinced of one thing: machines would never be able to help with empathy. Empathy, compassion, understanding – this is purely human territory. It's logical to assume. And now he says: "I couldn't have been so wrong." He refers to a dozen or more existing scientific studies that compared the responses of doctors and the responses of AI, for example, ChatGPT, to patient questions on online forums. The responses were evaluated based on criteria of information quality and empathy. And what were the results? Surely AI cannot be more empathetic than a human. The results are astonishing. In most of these studies, patients, and sometimes doctors themselves, who were shown anonymized responses, rated the AI-generated responses as more empathetic and often of higher quality than the responses of real doctors. How is this possible? After all, AI doesn't have feelings. It cannot truly empathize. That's the point. AI doesn't feel empathy, but it has learned to transmit and imitate it. It was trained on vast amounts of text, including millions of dialogues where people expressed compassion, support, and understanding. And it learned to use the right words, the right phrasing, to create a sense of empathy in the interlocutor. Sometimes even better than a tired, burned-out doctor responding on the go. This is simply amazing. And what conclusion does Topol draw from this? There are several conclusions. Firstly, this once again speaks to the completely unexpected and unpredictable capabilities of AI, including positive ones. We focus on diagnostics, medications, but it turns out it can also help in the areas of mental health, communication, and supporting lonely people. Secondly, and this is very important, it's a huge challenge for people, and especially for doctors. If a machine can imitate empathy so easily and effectively that it's sometimes difficult for us to distinguish, then we, humans, need to double down on our genuine human qualities. We need to consciously develop and demonstrate sincere empathy, attention, and care. Not just be polite, but truly be human for the patient, otherwise, we will be surpassed even in this. That is, paradoxically, the development of AI makes us think about what it means to be human, and especially what it means to be a doctor. Exactly. It's a stimulus for the humanization of medicine, however strange that may sound. Good. And let's move on to the last, but no less important discussion that Topol touched upon. What is ultimately more effective: AI by itself, or a doctor using AI as an assistant? Intuitively, it seems that synergy should win. But what does the data say? The new Health Bench benchmark from OpenAI was mentioned. Yes, this is a very hot topic. Now Topol refers to his recent article, co-authored with Pranav Rajpurkar, a renowned scientist in the field of AI in medicine. They analyzed the results of about six recent studies that directly compared three scenarios: a doctor without AI, a doctor with AI, and AI by itself. The tasks were varied: radiology, reading scans, diagnosis by symptoms, patient management. And what was the verdict? Did synergy not work after all? The results were, to say the least, counterintuitive. In almost all of these studies, AI by itself showed better results than doctors working without AI. But what is most surprising, AI often showed better results than the combination of doctor plus AI. That is, the synergy everyone expected was not observed. Sometimes a doctor using AI showed even worse results than AI itself. How can this be? Why does adding human intelligence to artificial intelligence not improve, and sometimes even worsen, the result? There is no definitive answer yet. There are several hypotheses. First, doctors are simply not yet trained to use AI effectively. They either blindly trust its results (automation bias) or, conversely, ignore its suggestions if they contradict their own opinion. Special training is needed to learn how to interact correctly with an AI assistant. Second hypothesis. Perhaps the studies themselves were not constructed entirely correctly. They were conducted in somewhat artificial conditions, not fully reflecting the complexity of real clinical practice. Maybe in real life, synergy will still manifest. And is there another opinion? Well, there is, of course, the viewpoint that AI is simply objectively better at certain tasks of data analysis and pattern recognition, and human intervention with its cognitive biases and limitations only hinders. Topol's co-author, Pranav Rajpurkar, being a computer scientist, leans towards this idea. Ultimately, in purely diagnostic tasks, AI will most likely win. And Topol himself is a doctor? Topol himself says he remains an irrational optimist. He still hopes for synergy, for doctors and AI to be able to work together effectively. He uses the metaphor of a high school dance, where AI and doctors are currently standing in different corners of the hall and don't know how to start interacting. But he believes that over time they will learn to dance together. But he admits that there is no convincing evidence of this synergy yet. Yes, he honestly says that there is little evidence yet. And he also makes an important point that people outside of medicine often forget. A doctor's work doesn't end with diagnosis. Diagnosis is often just the beginning. Next comes the choice of treatment, patient monitoring, communication with the patient and their family, therapy adjustment. In all these aspects, the human role remains key. And here AI currently acts as an assistant, not a replacement. Well, we have indeed delved deeply into Eric Topol's ideas, going through many aspects from personal productivity and education reform to the subtleties of diagnostics, screening, drug development, longevity, and even such unexpected topics as empathy. The picture is complex. Yes, very complex. On the one hand, incredible breathtaking potential, and on the other, clear tension associated with the difficulties of implementing these technologies into the conservative medical system. Inertia is visible, ethical dilemmas are visible, concerns are visible. What are the key threads that can be drawn from this entire discussion? What is most important? Well, first of all, it's, of course, data. Data is the new oil, especially in medicine. The importance of collecting high-quality, diverse data, providing patients with access to it, and using AI for its integration and analysis is a cross-cutting theme. Secondly, personalization. Moving away from universal approaches to individual risk assessment and individual strategies, both in prevention and treatment. What else? Patient empowerment. Topol consistently advocates for giving patients more control, more information, more tools to manage their health. Further, of course, ethical issues: the danger of predatory screening, problems of unequal access to technologies, issues of responsibility when using AI. And, finally, this surprising counterintuitive ability of AI to imitate and even surpass humans in unexpected areas like empathy. Yes, it makes you think. And in conclusion, perhaps some thoughts for our listeners, for further reflection. If AI is developing so rapidly, if it is beginning to surpass humans where we least expect it – in diagnostics, in data analysis, even in imitating empathy – then how should the very concept of medical care change? Indeed, what should the role of humans, of doctors, become in the future? Should they simply become operators of complex AI systems, or should their role transform into something else, something more, focused on those aspects that are still inaccessible to machines, on strategic thinking, on complex ethical decisions, on genuine human connection? And one more question to ponder. How can we ensure that this pursuit of health and longevity with the help of AI for super-agents does not lead to an even greater societal divide? How can we ensure that these benefits become accessible to everyone, not just the chosen few, and do not exacerbate existing social and economic inequalities? These are absolutely fundamental questions. There are no easy answers to them, but it is precisely these questions that we all – doctors, scientists, patients, politicians, society as a whole – must think about and work on as we enter this new, exciting, and slightly frightening era of AI-driven medicine.