A co, jeśli powiem Wam, że celem sztucznej inteligencji nie jest zastąpienie lekarza, ale… oddanie mu czasu na bycie człowiekiem?
Moim gościem jest dziś Vivek Natarajan z Google DeepMind. Nasza rozmowa, zamiast o algorytmach, jest przede wszystkim o empatii, czasie i odzyskiwaniu tego, co w medycynie najważniejsze.
Zapraszam na rozmowę, która daje nadzieję…
Lekarz w kieszeni, naukowiec w chmurze
Punktem wyjścia są dwa rewolucyjne projekty. Pierwszy to „lekarz w kieszeni” (AI Co-physician) – osobisty asystent zdrowia, który dzięki ciągłej analizie danych z naszego ciała i otoczenia, może przewidzieć chorobę na lata przed jej wystąpieniem. Drugi to „naukowiec w chmurze” (AI Co-scientist), który daje badaczom supermoce, błyskawicznie łącząc wiedzę z odległych dziedzin nauki, by przyspieszyć odkrywanie nowych leków.
Byłem pod wrażeniem, gdy Vivek wyjaśnił, co jest prawdziwym celem tych narzędzi. Bo to nie jest technologia dla samej technologii.
Człowiek dla potrzebującego
Zapytałem Viveka o głośne badania, z których wynikało, że AI bywa bardziej empatyczne od lekarzy. Spodziewałem się dyskusji o wyższości algorytmów. Zamiast tego usłyszałem coś, co uderzyło mnie w samo sedno. Problem nie leży w lekarzach. Leży w systemie, który zmusza ich do spędzania 70% czasu na wprowadzaniu danych do komputera.
W tym momencie cała historia nabrała sensu. Największą obietnicą AI w medycynie nie jest zastąpienie lekarza, ale zdjęcie z niego biurokratycznego jarzma.
„My hope, and I think the hope of a lot of people working in this field, is that AI can give time back to doctors. Give time back so that they can actually have these conversations, really learn better about their patients, and really understand their social context.”
To wizja, w której technologia nie staje między ludźmi, ale usuwa bariery, które sami zbudowaliśmy. W świecie, w którym największą pandemią staje się samotność, perspektywa odzyskania czasu na prawdziwą, ludzką rozmowę w gabinecie lekarskim jest rewolucyjna. AI zajmuje się danymi, a człowiek może wreszcie w pełni zająć się drugim człowiekiem.
Więcej niż narzędzie
Rozmawialiśmy o wielu aspektach – o globalnym dostępie do opieki medycznej, o etyce i zaufaniu. Ale przez całą rozmowę powracał jeden motyw: AI jako wzmacniacz, nie zastępca. Vivek użył pięknej metafory, która idealnie podsumowuje to podejście. To nie jest autopilot, który przejmuje stery.
„Use it as a bicycle for the mind. Let it amplify you… But that doesn’t replace the pedaling, the effort that you have to put in.”
Pytania do dyskusji:
- Kto powinien ponosić odpowiedzialność za błąd diagnostyczny AI – programista, lekarz, który z niego skorzystał, czy instytucja, która wdrożyła system?
- Czy bylibyście gotowi udostępnić swoje dane medyczne „lekarzowi w kieszeni”, nawet jeśli miałoby to pomóc wykryć chorobę 10 lat wcześniej? Gdzie leży granica prywatności?
- Jeśli AI da lekarzom więcej czasu, czy system opieki zdrowotnej faktycznie pozwoli im na dłuższe, empatyczne rozmowy, czy po prostu zwiększy „normy” przyjmowanych pacjentów?
—
Masz chwilę? Jeśli podoba Ci się 99 Twarzy AI, zostaw ocenę i recenzję – dzięki temu możemy docierać do jeszcze większej liczby słuchaczy!
Dobrego dnia i niech #AI będzie z Wami wszystkimi!
Transkrypcja rozmowy
Disclaimer. Droga Czytelniczko, Drogi Czytelniku – mała uwaga dotycząca transkrypcji rozmowy. Jak się pewnie domyślasz – transkrypcja została przygotowana z wykorzystaniem magii LLM. Proszę o wyrozumiałość, gdyby pojawiły się w niej niewielkie błędy, literówki, etc. Żeby mieć pewność co do wszystkich wypowiedzi – polecam posłuchać, zamiast czytać. Ukłony.
Karol
Could you briefly say what your area of research is? We are here in Warsaw at the Google research event.
Vivek
I am a research scientist at Google DeepMind, and I broadly lead projects at the intersection of AI, science, and medicine.
Karol
Let’s imagine ourselves in 15 or 20 years from now.
Vivek
Mm-hmm.
Karol
What might a future doctor look like? Is it still a human?
Vivek
I think the future is hard to predict, but in my view, we’ll probably not have the concept of doctor’s appointments and things like that anymore.
Karol
So can you imagine that you’ll have the doctor in your pocket?
Vivek
Yes. It’s going to be something that’s always with you, always available to you. And respecting privacy and all the constraints and your requirements, it’s going to be able to process and integrate all sorts of data that we might be able to collect. Imagine your wearables data, things at extremely low resolutions. The big progress that’s happening in the field right now is the fact that we can measure biology and human biology at extremely fine-grained resolutions, so we can understand molecular signatures, how people’s genetic dispositions might vary, proteomics, metabolomics, and all the way up to the phenotypes and the symptoms that you see and the environment around you, which is also very important. So this AI can assess, synthesize, and integrate all that information and give you the best possible advice to take care of your health.
To complete that chain of thought, what I think is going to happen is that if we get this right, we might be able to catch diseases 10 years before they happen. If we are able to do that properly, it means that the cures can be much simpler, the interventions can be much simpler. I imagine that if we get this right, then for a large majority of people, it’s going to be a lot more preventative rather than reactionary as it is today. Is it possible that we’ll live forever? Again, I don’t think there is any physics constraint over here. Even in nature, we know certain animals that live for 400, 500, 600 years; trees probably live for much longer. Trees and then animals deep down in the sea are known to live much longer. There were recent studies looking at what is different, say, between the genes that are expressed in the whale, which enables the whale to live 200 years longer. We are understanding all those things better. We thought it might be just because they slow down metabolism, but it turns out it’s some other mechanisms. But you can take those learnings and apply them, and so it might mean that our lifespans might radically increase.
Karol
When I hear you, I see that the mechanism is still the same. We are analyzing the data, correct me if I’m wrong. But right now, we are analyzing the information about the tissues, the skeleton, muscles, body, joints, et cetera.
Vivek
Even lower level, about our organs. Even lower, starting from the genetic code in your cells.
Karol
But it’s on a kind of macro level right now, compared with the vision you’ve described. So nothing will change, but the amount of information will increase. We will be able to connect the data and analyze it in a different way. We can predict the future not in terms of months, as when you diagnose cancer now, but we’ll be able to diagnose cancer 10 years ahead.
Vivek
Yes, or the earliest signs of disease, basically, or things not happening in the right way, things happening in an abnormal manner. I think that is the hope and the dream for many of us.
Karol
I had an interesting conversation a couple of weeks ago with a great lawyer. What struck me was his admission that, looking at the development of tools that can produce any document, regulation, or prepare you for a trial, lawyers should consider becoming psychologists. You know why? Because the ability to connect people, lead negotiations, and also to sell services will become more important. So if it’s very probable that the systems will be better than human doctors, what abilities or talents should future doctors have? Who should they be?
Vivek
Ultimately, at the end of the day, I feel medicine is a very humane profession. So many times, when people say, „Oh, I need to see a doctor,” it’s not really to get a diagnosis or medication recommendations.
Karol
But you are telling me „I need to see a doctor,” however, research showed that the level of empathy of the machine can be greater right now than the empathy of the doctor.
Vivek
Yes, our studies showed that. But I would say those studies are in some ways flawed, including ours. The reason for this is that the profession is set up in such a way that we don’t actually give time to the doctors. When you’re operating medicine as a business…
Karol
That is what I meant to say. Because when you are a doctor and you’ve got 10 patients an hour, it’s like a factory, and 70% of your time is spent taking notes and putting them on your computer, then you don’t have time for empathy or anything like this.
Vivek
Exactly. It’s transactional, it’s business-like, and it is not what medicine is meant to be, which is supposed to be a very humane endeavor. My hope, and I think the hope of a lot of people working in this field, is that AI can give time back to doctors. Give time back so that they can actually have these conversations, really learn better about their patients, and really understand their social context. If they have that understanding and context, they’ll be a lot more empathetic. But not only that, they’ll also be a lot more effective in guiding people because people are actually just looking for connection, just to have the feeling of being heard, having a human voice on the other side. Because that is the other big pandemic that we have around us, which is loneliness.
Karol
I’ve got a curveball for you.
Vivek
Go for it.
Karol
Looking at how people interact with chatbots, isn’t it possible that people will tend to visit the virtual doctor rather than the normal one? If people are spending more and more time with chatbots which, as we see, are empathetic and talk and act in a way we expect…
Vivek
Yeah.
Karol
So why should we go to the doctor?
Vivek
Look, we do see that trend, but I think the way to fix it… I think the diagnosis is wrong over there. The diagnosis is the fact that a lot of people don’t have humans around them to provide the care and satisfy their needs for companionship and things like that. If you’re telling me that an AI can replace human presence, then I’m sorry about the humans in your life. That would be super sad. I think that’s not true. We humans are, we should, and we will definitely be much more kind and empathetic. It’s just a nature of where we are right now as a society. Internet technology has done a lot of amazing things, but at the same time, we can’t ignore the fact that there is a massive loneliness epidemic. When people are turning to these virtual chatbots, it’s just a reflection of wider trends in society, I think.
Karol
Sure. We got used to co-pilots or co-assistants. There’s a new word I came across: a co-scientist. What does it mean and what can co-scientists do?
Vivek
I work on and lead a project called the AI Co-scientist at Google. For me, working at the intersection of AI, science, and medicine, we have this vision which is two-pronged. There are two parallel tracks running. One is, there are a lot of amazing discoveries happening in medicine today. The question is, how do you take all this medical expertise that is available in San Francisco, Boston, in the US, maybe even in parts of the EU and here in Warsaw, and deliver it to people in the rural parts of the world, to the billions of people in remote communities in India and Africa, and so on? That is why we are building the AI co-physician, and that’s the AME project that we have.
But at the same time, we can’t stop progress because there are still so many diseases that we have no understanding about, like ALS, cancer, Parkinson’s disease. There’s still this urgent, unmet need that we need to overcome. That is where we need to fundamentally accelerate the clock speed of scientific discovery, so that we can understand biology better, design more cures and therapies, and deliver them in a personalized manner. That is where the AI co-scientist comes in, and the goal is really to give scientists superpowers, amplify their abilities, and really transformatively accelerate the clock speed of scientific breakthroughs.
Karol
And can you imagine a future where a co-scientist will make a major scientific discovery on its own?
Vivek
Where the technology is right now, and as someone who’s involved in building it, I think there are different kinds of innovations. Let me step back and put it that way. There’s a lot of innovation which is about making these interpolations or these unexpected connections. That is fundamentally bottlenecked by the availability of human time and expertise. There are only so many PhDs or deep experts that we train in a given topic. So there are limitations on that. But I think there’s a lot more to discover which is relatively simple to do, like reading maybe a hundred different papers and making these unexpected connections. With systems such as the co-scientist, we’ll see a lot more of those kinds of discoveries. I don’t want to ignore them because they can be extremely effective, and we’ve already shown that. For example, the co-scientist we have was able to make this connection that a drug originally approved for cancer can also be very effective for liver fibrosis. But if you ask a liver fibrosis doctor, they don’t know much about cancer because that is not their area of expertise. The co-scientist was able to make that connection.
Karol
I was curious about this case you mentioned, and I was thinking that it was about the amount of medicine and the way it is distributed. And when we are talking about this flow between the research…
Vivek
Yeah.
Karol
I see a kind of image, correct me if I’m wrong, and I don’t want to be offensive or negative, but sometimes the research or the academic world is very closed.
Vivek
Yeah.
Karol
And with LLMs, you’re able to connect the worlds very easily.
Vivek
Yeah.
Karol
And to talk between the areas…
Vivek
Yeah.
Karol
Without even asking another researcher about the results of the research he’s done.
Vivek
It is kind of like that.
Karol
It’s a kind of knowledge transfer?
Vivek
It is. In fact, we call it transfer learning.
Karol
Look, the internet was made to connect the academic world. So right now, it’s a kind of connection for the content itself, not only for transferring the information but for being able to mix it up. Wow. It’s the same story, like how I sometimes try to show the abilities of LLMs by changing the perspective.
Vivek
Yes.
Karol
Because I usually show people how, for example, the marketing department can use financial data, use an LLM, and be able to understand the financial data with the dictionary of a marketing team. So isn’t that the same with research and that kind of lab stuff?
Vivek
It is definitely kind of like that. Taking a step back, as we were talking, in terms of the nature of innovations, what I was getting to is that there are maybe two broad kinds. What we are talking about is the first kind. This is definitely very true. For example, for me today, with LLMs like Gemini or ChatGPT, it becomes much easier if I put in the effort to understand what another scientist is doing, whereas before it would take me much longer. So it lowers the energy.
Karol
So, explain a physicist’s domain to a mathematician.
Vivek
Exactly, or a biologist. And again, a lot of people tend to miss that there’s a connection between all these fields. You can say in some ways that everything emerges from math and physics and quantum chemistry, and then biology. If you take that perspective, then biology is actually fully connected to math and physics. In fact, people like Eric Lander talk about programmable biology, or biology as programs, which means that biology can be connected to computer science. The other big trend I can see sparking right now… I do believe in some aspects of that. Of course, you don’t want to take it too far because these are abstractions that you develop to make it easier for you to wrangle these difficult topics. But the truth is, a lot of the most impressive breakthroughs in the last two or three decades have actually happened at the intersection of fields. So I think AI co-scientists can actually help with that intersection, that intermingling, to happen. Bring together communities of people, help them speak the same language so that they can then more effectively collaborate. The other big trend in science is that we no longer have these individual geniuses making breakthroughs; rather, science has become a collective endeavor.
Karol
If a co-scientist came across a new discovery, who should get the Nobel Prize?
Vivek
Look, the way technology is today, I would say that it is still primarily a tool. It’s an extremely capable tool, an extremely powerful tool.
Karol
So you mean it’s proof of the human in the loop?
Vivek
Exactly. Ultimately, it’s still up to the human to use that tool properly, ask the right questions, guide it the right way, and things like that. But if you’re able to do that, then yes, it can rapidly accelerate your research. I would still say that the credit goes to the person who asked the question and used the tool in the first place.
Karol
Coming back to the AI doctors, should we trust the AI doctors more than people?
Vivek
Again, this is common to both our AI scientist work and our AI co-doctor work that we are doing. We are still trying to do a lot of work to ensure that the AI can accurately communicate uncertainty. Because if you’re always confident and you don’t know when you don’t know something, then it becomes very hard to trust. So I think that’s a very important aspect that we have to nail with these systems. It’s still a research frontier right now to get LLMs to have much better calibrations of when they don’t know about something and being able to accurately convey that information. I would say that until we get that right and until we get it to how good human doctors are, the reliance will be primarily on human doctors. Human doctors will be using AI mostly as a tool for diagnosis and recommendations. But maybe the switch might flip once AI becomes much more accurate at communicating uncertainty.
Karol
Is there anything you are worried about, looking at the development of the technology right now? Do you have any concerns?
Vivek
The problems that keep me awake at night are the fact that maybe there are billions of people who don’t have access to medical expertise, or maybe there are hundreds of millions of families who are facing devastating diseases. And why we as scientists, as researchers, are not able to make faster progress.
Karol
You know, what you’ve just said struck me because I connected it with something I read a couple of months ago. For example, in Africa, there’s very little infrastructure on the continent. However, most of the people have a mobile phone.
Vivek
Yes.
Karol
So there’s a huge chance that without any infrastructure, medical consultancy will be available for the people. Am I right?
Vivek
Yes. I think you would still have to deploy technology responsibly. It is very important that if your AI has been primarily developed in a Western context, you do the right kind of rigorous evaluation, account for the social context over there, and do proper ethical studies and work with patient consent, and only then you deploy. But once you deploy, I can totally imagine rapid progress and accessibility to world-class expertise even in communities in Africa and India.
Karol
I think about this one. Treating a body is one thing. The medicines, all the imaging, the medications. Okay, so the body is one thing, and the soul is another. You’ve got doctors for your body and doctors for your soul. I can imagine that machines will analyze the data that are given from medical devices better than humans.
Vivek
Mm-hmm.
Karol
But how about treating the soul, for example, psychologists? I will give you an example. A couple of months ago, one of the listeners wrote to me that one of the chatbots helped them get out of depression.
Vivek
Mm-hmm. That’s great.
Karol
Great or sad?
Vivek
I mean, the fact that you’ve been able to help someone is good. And you don’t have to be a doctor to help someone with a medical condition. The interventions can sometimes be…
Karol
I had a great conversation with a professor, one of the leading ethicists in Poland. He was chief of the International Neural Network Association. We were talking about the kind of relationships that we have with LLMs. Of course, when you lead the conversation in the proper way, you can get the proper answers. On the other hand, if you see LLMs not like a mirror…
Vivek
Yeah.
Karol
But, I don’t remember how it’s called in English, the mirror from the legend where you look at it but you don’t see yourself.
Vivek
Hmm. Yeah.
Karol
Do you catch my meaning?
Vivek
It’s a good point.
Karol
It’s the language model. So if you are not using the proper words when describing your problem, situation, or the things that you have, you can get the wrong answers. It’s the same thing, for example, I was analyzing that you can’t treat a chatbot like a friend. Because a human friend, when you are saying stupid things, tells you, „Hey, you’re wrong.”
Vivek
Yeah.
Karol
I asked a chatbot a question: „Hey, I’ve got a brilliant idea. What would you tell me about the business model based on apples on sticks?”
Vivek
Yeah.
Karol
„Brilliant idea. You’ll make lots of dollars on that.” What would a friend say?
Vivek
„Probably it’s not a great idea.”
Karol
Exactly. So what do you think about this?
Vivek
I actually want to make a couple of points here. The first one is, you talked about AI for the body and replacing medicine. I think even there, there’s a lot of misconception around what caregiving is and what doctors do. There are parts of medicine that AI is never going to replace. Imagine taking care of a patient who’s completely bedridden, which can happen if they have surgery, but oftentimes from growing old or terminal diseases. Imagine the kind of work you have to do to change their feeding tubes or their urinary tubes, or cleaning them up. People underestimate and underappreciate the work that caregivers do and the amount of mental effort it takes. When people say that AI is going to come and replace medicine and caregivers and everything is going to go away, it’s nonsense. You have no understanding of medicine or what caregiving is. So I think it’s all about amplification and helping them do a better job and helping them at giving care.
It’s the same with the stuff you talked about in your second point. This is another limitation, or rather, a reflection of how AI is trained. I’ll use the friend analogy you have. When you’re talking to your friend about a problem you have, you are giving them some information, but that is lossy compression. They are not embodying you and having the full experience that you have.
Karol
That’s the power of context.
Vivek
Exactly. So in that sense, they’re taking some compressed representation of your experience, and then they’re mixing it up with their representation and giving you some advice. But it can be imperfect, it can be noisy, which means it may not completely apply to you. That is why I say you should listen to advice from a lot of people, but ultimately, you are the best judge because you have the full context, you know what’s going through your head, you know what you have experienced. It’s the same with AI as well. When you think about how these AI systems are trained, they’re actually trained on all written text and language on the internet, but they don’t have experience. They’re just reading a lot of books and papers and articles. They’re learning from projections of human experience and knowledge, and that’s why they have limitations.
Karol
If context is king, shall we record our lives?
Vivek
I mean, there are all sorts of devices coming, like the glasses and things like that. That’s happening. I can imagine always-on assistants being part of those glasses. But again, whether there’ll be adoption of that technology, I don’t know.
Karol
At the end of our meeting, do you have any favorite prompt technique that you can recommend or any advice for our listeners? Is there any way of working with LLMs that’s super beneficial for you?
Vivek
I think it’s mostly the fact that when you look at social media and the commentary on AI and LLMs, it’s two extremes. One extreme is, „Oh, it’s a super useless system.”
Karol
It’s useless.
Vivek
That’s the first extreme. It’s useless, it can’t do anything, it’s hallucinating all the time.
Karol
If you are asking useless questions, the answers will be useless.
Vivek
Yes. And then the other extreme is, „Oh, AGI is here, it’s going to replace all jobs, it’s going to tank the economy.” But the truth is right in the middle. It can do certain things for you, it can do them effectively and productively, but it also has a lot of limitations. So I think it’s up to individuals who have the context about their professions, their skills, their needs, to be able to experiment with this technology and figure out how it can best help them. But also, it’s not a substitute for your own thinking and your own work. Rather, use it as a bicycle for the mind. Let it amplify you, let it help you do more, let it help you become more productive. But that doesn’t replace the pedaling, the effort that you have to put in. It’s just going to amplify you and let you go longer.
Karol
Great analogy. You know, because when you are going for a walk, when I look around the world, I see it in a completely different way than while riding a bike, because the perspective is changing much faster and we see a lot more.
Vivek
It’s exactly what’s happening over here.
Karol
Vivek, it was a great pleasure to have you on my podcast. Anytime you are in Warsaw, please feel invited.
Vivek
I’d love to come back again in the summer. It’s a great place. And again, thank you so much, Karol.
Karol
Thank you very much.
Vivek
Thank you.