Transcript
0:00 To turn a medical breakthrough into a worldwide cure, you need an access multiplier. Tech leaders predict AI will eliminate major diseases in 10 years just by discovering new drugs faster. But take the smallox vaccine. It was proven effective in a lab in 1796. But it didn't actually end the disease until almost two centuries later. A brilliant discovery isolated in a lab makes a disease treatable for a privileged few while it continues circulating naturally in the broader population. Meaning the disease remains endemic. Even if an AI discovers the perfect molecule instantly, the physical cure gets blocked by manufacturing limits, specialized facilities, and insurance approvals.
0:44 To drive naturally occurring cases to absolute zero worldwide, a milestone called eradication, you need a coordinated institutional network to physically push the science through those bottlenecks. This is the access multiplier operating at scale. It takes that single discovery and multiplies it through local health workers, surveillance, and containment protocols. So, an AI can definitely invent a better drug in a fraction of the time. But just like smallox in 1796, that science won't eliminate a disease until global institutions physically deliver it.
1:22 Welcome to the debate. So if you looked at a map of human suffering in 1977, there was this very specific singular pin placed in Somalia, right? >> And it marked the very last naturally occurring case of small pox. But by May of 1980, you could look at that exact same global map and there were zero pins. I mean, the disease was entirely gone, >> which is just a staggering historical achievement, you know. And today, we are exploring the promise of AI in medicine, specifically whether faster scientific discovery through artificial intelligence actually justifies predictions that diseases like cancer, Alzheimer's, and heart disease will soon become relics of the past.
2:02 >> Exactly. And to ground this discussion, we are looking specifically at Anthropic CEO Dario Day's October 2024 essay, Machines of Loving Grays. >> Yeah. Alongside his his November 13, 2026 conversation with JP Morgan Chase CEO Jamie Diamond, which was just ahead of Anthropic's planned IPO, >> right? But we are also looking at the historical evidence here. So tracing the WHO smallox campaign, analyzing AlphaFold's protein predictions, and even examining the recent November 12 resignation of Nvidia's Jensen Huang, >> which is a fascinating case study in how we navigate technological authority.
2:38 >> It really is. So the central question before us today is this. Can AI actually cure disease? And well, what exactly counts as proof of that progress? I will argue that AI represents a fundamental paradigm shift capable of dismantling the scientific bottlenecks of biology, turning these medical aspirations into highly probable forecasts. >> And I will argue that technological brilliance does not eliminate downstream clinical and institutional bottlenecks. Basically, I believe we are dangerously conflating scientific capability with population level cures.
3:11 >> Okay. Well, let's jump right into the core of Daario Amade's thesis because honestly, I think it is structurally sound and historically unprecedented. In his essay, he predicts that powerful AI could compress 50 to 100 years of biological progress into just 5 to 10 years. >> Wait, but we have to be really clear on the mechanics there, >> right? Yes. He doesn't mean a large language model has the cure for cancer hidden somewhere in its weights. You know, those billions of mathematical parameters. He is arguing that if we achieve sufficiently powerful AI, it will fundamentally accelerate the speed of data analysis and physical experiments.
3:45 >> Right? And I don't disagree that the tools are incredibly powerful. But we have to define what happening actually means in a biological context. >> Well, let's look at AlphaFold. I mean, before this AI tool, mapping a single protein's three-dimensional structure could take a PhD student five years. using these incredibly expensive, tedious techniques like X-ray crystalallography. It was a massive physical science bottleneck. >> Oh, absolutely. The crystalallography was agonizingly slow. >> Yeah. And then DeSabis and John Jumper presented Alphafold 2. By the time they shared the 2024 Nobel Prize in chemistry, Alphafold had mapped the structures for virtually all 200 million identified proteins known as science.
4:27 >> 200 million, >> right? And it did this in a matter of months. That isn't just a minor productivity boost. That is biology transitioning from a slow physical science into a rapid information science. >> I mean, sure, 200 million proteins is an astonishing number. I won't debate that. But folding a protein in a simulation is miles away from a safely metabolized drug. >> How so, though? >> It's the foundational step. >> Because knowing the 3D shape of a protein is it's like knowing the exact shape of a lock. It absolutely helps you design a key, but it doesn't establish whether that key reaches the right door in a living human body.
5:04 >> Well, obviously you have to test it, >> right? But does it accidentally open other doors in the hallway? Does it burn the building down in the process? We are falling into a trap here. Honestly, a conflation of aspiration and forecast. >> What do you mean by that? >> An aspiration is saying, you know, we have this incredible new keymaking machine and I am going to try to cure disease. A forecast is saying cancer could be a relic within 10 years. That requires us to rigorously examine the assumptions behind that timeline.
5:32 >> I think the timeline is based on the compounding speed of the tools, not just blind hope. >> But right now, we are seeing a massive amount of what we can call authority transfer. This is where success in one specific domain like building incredibly powerful AI models is incorrectly treated as permission to stop asking rigorous questions about medical timelines and crucially the institutions required to validate them. Wait, let me stop you there because framing it as just an aspiration or authority transfer entirely misses the scale of what AI has already pushed into the clinic. We are moving far beyond just mapping the locks.
6:04 >> Okay, let's talk about the clinic then. >> Let's talk about actual clinical validation. In June 2025, Nature Medicine published a randomized phase 2A study for a drug called Rentura. >> Right. The pulmonary fibrosis drug. >> Yes. An AI discovered an AI designed drug targeting idiopathic pulmonary fibrosis, which is this devastating disease where the lungs progressively scar and stiffen. The researchers measured lung function using a clinical endpoint called forced vital capacity. >> Right. So essentially how much air someone can forcefully exhale, the volume of a deep breath.
6:36 >> Exactly. And at the highest dose, patients saw a mean increase of about 98 milliliters in their force vital capacity compared to a decrease of about 20 milliliters in the placebo group. >> I've seen the numbers. Yes, >> we are talking about a tangible concrete clinical endpoint where an AI designed molecule improved human lung function. The AI didn't just play in a simulation, you know, it designed a physical molecule that helped real people breathe better. The pace of drug development is accelerating. I hear you on the clinical endpoint, but let's actually look at the math and the biology of that trial for a second. You have to look at the boundaries dictated by human biology, not just the AI speed.
7:12 >> The biology responded to the drug. >> Yes, but that rent a certip trial enrolled exactly 71 patients. We are talking about 17 or 18 people per treatment group observed over a period of just 12 weeks. >> But Encilico Medicine announced in July 2026 that Rent assertive entered phase 3. That is a massive milestone. It is a corporate development milestone, yes, but it is not a completed demonstration of long-term population benefit. And this is exactly my point about the biological bottleneck, >> which is what phase 3 is for.
7:42 >> Exactly. Because human bodies are not simulations. The FDA explicitly notes that preclinical research, no matter how supercharged by AI, cannot substitute for learning how a drug interacts with the human body over time. >> Sure, I agree with that. >> Why? Because of idiosyncratic toxicity. A drug might work perfectly in 18 people over 12 weeks, but what happens when person number 400 takes it for a year. Does it cause liver failure? Does it trigger an autoimmune response? But that's always a risk in any drug development, >> right? And NEI does not remove the biological requirement of running those phase 3 trials, which demand thousands of patients and multiple years to statistically prove a drug won't harm the population. You simply cannot disrupt the physical reality of the human body like it's just some inefficient bureaucracy. Okay, drug discovery isn't just a bureaucracy. I completely agree. But think of it this way. Historically, finding a viable drug candidate has been like, well, like navigating a pitch black maze.
8:36 >> A very expensive maze. Yeah. >> Right. Researchers literally bump into walls for decades trying to find a molecule that works. >> A very expensive maze. Yeah. >> Right. Researchers literally bump into walls for decades trying to find a molecule that works. AI doesn't just make the researcher run faster in the dark. It acts as a GPS, showing us the direct path to the exit. >> But you still have to walk the path. >> Yes, exactly. To your point, the FDA still requires us to physically walk every single step of that path to prove the floor won't cave in. But if you give researchers that GPS, the overall time to get a drug to market drops massively.
9:12 >> I'm not convinced it drops by the margins Ami is claiming, though. >> Well, his argument doesn't require AI to magically bypass the FDA. It requires removing the most restrictive bottleneck at the very beginning of the pipeline, the discovery of the molecule itself. Compressing 50 years of progress in a 10 doesn't mean skipping the safety trials. Then what does it mean? >> It means arriving at the FDA's doorstep with perfectly optimized, highly probable candidates 10 times faster.
9:37 >> But that assumes the regulatory doorstep and the clinical trial phase are the only remaining hurdles. And they just aren't. This is where Amodi's historical comparisons really start to break down for me. You mean the small pox comparison? >> Yeah. In his remarks, he grouped small pox and bubanic plague together as diseases we think of as relics. And you open the show with smallox and it's a beautiful story, but it conceals a massive historical and institutional difference. Small pox was eradicated.
10:04 Plague was not. >> Plague is highly treatable today though. >> It's not a death sentence like it was. >> Yes, >> it's treatable with antibiotics, but treatable is not eradicated. Eradication is a highly specific mathematical term. It means reducing worldwide naturally occurring incidents to absolute zero. >> Sure. Statistically speaking, >> and the World Health Organization's September 2026 fact sheet points out that human plague remains endemic in several countries. Between 2019 and 2025, most cases occurred in the Democratic Republic of the Congo and Madagascar.
10:35 >> But that's a distribution issue, not a scientific one. The science solved the plague. >> Exactly. That is the entire point I'm making. Eradication didn't just require Edward Jenner inventing a vaccine. It required massive grinding institutional machinery >> to deliver it. Yes. >> It took the WHO coordinating hundreds of thousands of local health workers, active surveillance, case investigation, and maintaining intense cold chain supply logistics throughout the 1960s and '7s to reach that pin in Somalia. >> Nobody is denying the logistics are hard.
11:06 >> But AI cannot solve that. AI might rapidly invent a brilliant pill that makes a disease treatable. But an algorithm cannot manufacture the international coordination, the local health workers, and the supply chains required to actually eliminate a disease from the human population. The institutions are the bottleneck. >> Look, I think that draws a distinction that misses the functional reality for the patient. >> How so? >> Well, I absolutely acknowledge the epidemiologic difference between global eradication and an endemic control. But let's be pragmatic here. If you're listening to this and you have a family member with Alzheimer's, hearing a tech CEO say the disease will be a relic in 10 years feels incredibly heavy.
11:43 >> It's massive. It's either the greatest news of your life or cruel overpromise. >> Right. But if AI can make Alzheimer's as treatable as the modern bubanic plague where you take a targeted AI designed therapeutic early in the disease progression and your cognitive decline is entirely arrested. That functionally removes Alzheimer's as a major societal threat >> for the people who can get the drug. But for the patient receiving that diagnosis, the disease does become a relic of the past in their own life. It doesn't matter to them if the disease technically exists in a laboratory or in isolated pockets so long as they have access to a highly effective cure. That reality validates the core of Amday's forecast.
12:21 >> So long as they have access, I mean that is the most dangerous phrase in this entire debate. >> Why dangerous? It's just the next step. >> Because this brings us to what we should call the access multiplier. Even if AI completely revolutionizes the discovery phase and acts as that perfect GPS through the maze, the central political and economic problem remains. Who actually gets the treatment? >> The assumption being that AI discoveries will be prohibitively expensive.
12:45 >> Look at the breakthrough therapies we have right now without AI. I mean CART T- cell therapies for cancer are miraculous, right? But they can cost half a million dollars. >> They are very expensive. Yes. and they require specialized medical facility, continuous biological monitoring, highly complex bespoke manufacturing, and just grueling insurance authorizations. Imagine an AI designs a miraculous therapy for cancer tomorrow. For a wealthy patient living near a major research hospital with platinum insurance, yes, the disease is a relic.
13:14 Okay. But >> but for an hourly worker who cannot afford repeated travel or whose insurance denies the claim because it's considered experimental, that AI breakthrough is practically theoretical. >> I see your point on equity. >> If access is close to zero for a specific community, multiplying the brilliance of the scientific discovery by a factor of 10,000 does not change that community's health outcomes. The equation still zeros out. Amade's technological success doesn't exempt him from the reality of health economics.
13:42 But listen, we cannot solve the access problem if we don't first invent the cure. And to invent the cure using artificial intelligence at this scale, we need unprecedented amounts of capital and computing power, >> which is why he's making these forecasts. >> Yes, we can't ignore the socioeconomic drivers behind these promises. You brought up Amod's November 13th conversation with Jamie Diamond earlier. Let's look at the context of that. That was a conversation happening on Wall Street. Ahead of a plan anthropic IPO regarding the massive computing infrastructure required to realize this AI vision, >> right? It's a pitch. It's fundamentally a sales pitch.
14:16 >> It is a pitch, but hope has economic power. Mod's vision serves as an organizing aspiration to gather the necessary resources to attempt work that looks unreasonable to skeptics. >> I don't think it's unreasonable to ask for evidence. Big projects, whether it's the Apollo program or mapping the human genome, need coordination and immense funding. And to his credit, Amod explicitly acknowledges in his essay that his vision is stareyed and that the skeptics might be right about the clinical bottlenecks.
14:43 >> He does say that. Yes, >> that intellectual humility shows he isn't demanding blind faith from Wall Street or the public. He is proposing a rigorous hypothesis and asking for the capital to test it. Well, I have to push back on what intellectual humility actually requires when you are making public medical forecasts. >> What do you mean? He admitted he could be wrong. >> Saying maybe I'm wrong is easy rhetoric. True intellectual humility requires establishing visible conditions for revision. What specific measurable evidence would force Ammoday to revise his 5 to 10ear timeline?
15:15 >> The failure of these drugs in clinical trials presumably. >> But when exactly does the 10-year clock start? Does cure mean prevention, remission, complete elimination? If there is no conceivable experimental disappointment that would cause a timeline to be delayed, then this isn't a scientific forecast at all. It is just an expression of loyalty to a technological vision. >> You think people are just following the hype, like a cult to personality. >> It actually reminds me of what sociologist Max Vber called charismatic authority.
15:42 >> Okay, let's unpack that. >> This happens when followers treat a leader's exceptional qualities as a reason to suspend critical roles. We see someone build an empire of AI capabilities which takes undeniable genius and we accidentally assume that genius applies to everything they touch. >> A halo effect. >> Exactly. Giving them a halo effect that bypasses normal scrutiny. We assume that because someone is a genius at designing neural networks, their medical timelines or geopolitical forecasts must also be genius. I'd argue charismatic vision is an economic necessity for scale though, not a mechanism for evading accountability.
16:15 >> How so? A visionary leader isn't suspending critical rules. They are defining a north star so that the thousands of engineers, biologists, and regulators know what coordinate to row toward. Enthropic didn't just issue press releases in 2026, you know. >> No, they are doing the work, >> right? They published an analysis of 207,000 research sessions examining exactly how people use their AI models for science. They are actively partnering with drug developers. They are generating real data. And yet we just saw a masterclass in the limits of that exact kind of charismatic authority with Nvidia's Jensen Wong.
16:49 >> the resignation. >> Yes, the WHO succeeded because of human institutions. And that brings us directly to how we govern the humans building these AI tools today. On November 12th, Wong resigned immediately as president, chief executive, and director of Nvidia following allegations of an affair, a board investigation, and unverified reports of intelligence connections involving a subordinate, >> which I want to emphasize remain unverified allegations, as the BBC's reporting carefully noted. >> Absolutely, and they should be treated as unproven. But the point here is about governance. Nvidia's chips are the absolute bedrock of the entire AI industry, and Hong's technical and commercial genius in shaping that hardware is undisputed, >> unquestionably, >> but achievement is not a universal credit score. You do not accumulate enough engineering points to settle an unrelated question about workplace hierarchy, conflicts of interest, or corporate governance.
17:41 >> No one is saying you do. >> But his technical brilliance did not erase his need to answer to basic institutional accountability. We must apply that exact same standard to medical AI. The brilliant achievement of predicting 200 million protein structures does not give any tech leader a free pass on the grueling, unglamorous work of clinical trials, manufacturing, logistics, and equitable healthcare access. >> Honestly, I actually agree that an allegation shouldn't erase a scientific contribution just as a scientific contribution doesn't grant immunity from corporate governance. We definitely need that rigorous distinction.
18:15 >> So, you agree the timelines are suspect? No, because applying that logic to medicine doesn't negate a mod's 5 to 10ear forecast. I think it actually supports it. >> Wait, how does an institutional failure support a 10-year timeline for curing disease? >> Because we should rigorously separate the AI discovery engine from the human governance engine. >> Okay, go on. >> Let's say AI discovers a highly effective cure for Alzheimer's in year four of this 10-year window. If the human governance engine, so the FDA, the hospital networks, the insurance companies spends the next eight years slowing down the distribution of that cure due to bureaucracy or pricing disputes. That is a failure of our institutions. It is not a failure of the artificial intelligence. The AI still did exactly what Amade predicted. It compressed a century of biological discovery into a fraction of a decade.
19:02 >> But the patient doesn't experience the AI in a vacuum. The patient experiences the medical system as a whole. You cannot neatly separate the discovery engine from the delivery engine when you are making public promises about curing cancer and Alzheimer's. >> I'm just saying the AI did its job in that scenario. >> But if the FDA requires a 4-year phase 3 trial to ensure a drug doesn't cause liver failure, that is not an institutional failure slowing down the AI. That is the necessary speed of human biological verification. Now, I'm not saying we should bypass safety, >> but the FDA's public guide to clinical research is very clear on why preclinical results cannot substitute for human studies. We cannot treat the physical reality of the human body as just another inefficient bureaucracy to be disrupted.
19:44 >> And I am absolutely not suggesting we disrupt safety or bypass the FDA. What I am suggesting is that AI will give us such an overwhelming volume of highquality, highly targeted therapeutic candidates that the sheer probability of successful clinical trials will skyrocket. >> Volume doesn't guarantee safety, though. >> No, but if you use the GPS to find a 100 perfect paths through the maze instead of bumping into walls for a decade to find one, your odds of reaching the exit safely change entirely. The rent assertive trial wasn't a fluke. It is a preview.
20:13 >> A preview that still requires years of human testing. Yes, it needs larger phase three trials, but the fact that an AI mapped to target designed a novel physical molecule and that molecule successfully improved lung capacity in a randomized double blind trial is just staggering. We are standing on the precipice of an unprecedented era in medicine. It is unprecedented. I will give you that. If we maintain the ambition and yes, the capital and the visionary leadership required to build these systems, AI's proven capacity will unequivocally alter the landscape of medical discovery, we are fundamentally changing the odds of human survival against disease.
20:50 >> We are changing the odds of discovery certainly, but true medical progress requires matching that technological capability with fierce institutional accountability, >> which we should demand, >> right? But if we want to genuinely conquer disease, we have to follow the bottleneck relay to its actual conclusion. We must demand clinical proof in living human populations, not just simulated successes in a neural network. >> I agree. >> We have to mathematically account for the access multiplier, ensuring that breakthroughs don't just become luxury therapeutics for the ultra wealthy. And above all, we must strictly refuse authority transfer.
21:25 >> You mean separating the genius from the halo? >> Exactly. We can applaud the technical brilliance of people like Daario Amade and Chinsen Huang while simultaneously demanding that their forecast be rigorously tested and their institutions be held accountable. Ambition deserves our attention, but it does not make our biological or economic reality disappear. Well, it seems we have arrived at a solid point of convergence on the reality that artificial intelligence will irrevocably alter the landscape of scientific discovery. I mean the transition from physical science to information science driven by data from Alpha Fold and early clinical successes is simply too profound to ignore.
22:02 >> Agreed. The capability is very real and the scientific bottlenecks are definitely shifting. Where we continue to diverge is on whether accelerating that single stage of discovery can conquer the downstream realities of clinical biology and equitable access. And that is a tension you, our listeners, will have to weigh for yourselves as you encounter these technological promises in the news. View them through the dual lenses of scientific capability and historical institutional reality. We really encourage you to explore the source material yourself. Read Amod's essay, look up the FDA's guidelines on clinical trial phases, and you know, review the WHO's archives on how smallox was actually defeated. Ask yourself, what specific measurable clinical milestone would you require to trust a 10-year forecast for the end of disease?
22:50 >> Because ultimately, the true measure of artificial intelligence in medicine won't just be the moment a neural network invents a cure in a simulation, >> right? It will be the day we can look at the map of human suffering and pull the very last pin. Thanks for joining us for this discussion. Hey, oh a Hey, hey, hey.
24:12 Heat. Heat. Heat. Heat. N.