Transcription
Over the next year, our industry estimates that the vast majority of programmers will be replaced by AI programmers. According to research groups like OpenAI, Anthropic, and others, right now, approximately 10-20% of the code in their research projects is already generated by computers. Technically, this is called recursive self-improvement. And what will happen when this process starts to accelerate exponentially? A great deal. The year 2025 is becoming a turning point for artificial intelligence and its development. But, according to leading experts, the real shock awaits us in 2026. It is then that AI will displace millions of jobs. In an interview with Eric Schmidt, former director of Google, frank words were spoken that you definitely won't want to miss, so let's listen to them. And this is your specialization, but today we are talking about how it is combined with biotechnology. The goal of this report is to raise awareness, expand knowledge, and even inspire people, perhaps to encourage them to work in this field. If you could give one example of such a merger of AI and biotechnology that you believe is capable of inspiring people, what would it be? First of all, I want to thank Senator Yang, other members of Congress, Caitlin, and the entire team that prepared this report. I am proud of our work. The commission has invested a lot of time in working out the details. When you participate in such projects, you not only learn a lot but also make a real contribution. One of the things I realized and want to draw your particular attention to is that biotechnology specialists have already found a solution for scaling their technologies. They know how to deploy developments to huge, broad markets. I myself live in the world of technology, where scaling is key, but there is actually a huge valley of death. And I would urge all of you who are somehow connected with this field to think seriously about it. Our recommendations in the report mainly concern the science of scaling. Technically, this means, for example, how many vessels to use, how to grow cells and cultures in them. These are all very complex biotechnological issues. In addition, we provide recommendations for creating infrastructure for scaling. And this is important because it shows how technologies can be deployed to broad markets. And this experience is also useful for AI. But what is happening today? There are amazing, talented startups. They attract money, do truly interesting things, but they find themselves in a difficult situation. Their products are not yet ready for large-scale market entry, and attracting new investments requires too much proof of success or resources. As a result, they fall into the so-called valley of death, where there are ideas and initial funding, but the transition to mass growth is difficult. This is a well-known problem. And if there is one point from the entire report that you should take away, it is this. Of course, there are many other aspects that I am sure everyone here understands. Basic science, the ability to accelerate drug development, as well as the risks of its misuse. I have personally worked extensively on issues of potential AI misuse. And there is compelling evidence. New AI models, if not restricted, can create dangerous pathogens, particularly viruses. Let's just take an existing virus and modify it slightly, and the consequences can be very serious. The reason we brought AI and biotechnology together is that this is an extremely important topic for America. We are talking about a multi-trillion dollar industry. There is already a lot of evidence that China is investing colossal amounts of money to dominate this field. They have a lot of technology, but some of it they have stolen, and we have documented all of this in detail in the report. In other words, artificial intelligence can accelerate drug development, but at the same time, it creates a threat that pathogenic viruses can emerge. In the next part of the interview, Dr. Eric Schmidt calls AI a ticking time bomb and explains the rather curious logic of this statement. Listen very carefully. I am the main investor in a group of people who have created an AI model initially trained in chemistry. This is a fundamental model for this field. It has been connected to a robotic laboratory. What does it do? It generates variants for drug development. It simply generates. Whether they are correct or not, God knows. And then at night, the robotic laboratory tests these variants and by morning provides a report. And the cycle begins anew. I mention this because it is an excellent example of a future model of AI and biotechnology fusion. AI generates many variants for testing, reducing the so-called search space. From an algorithmic perspective, the task is incredibly complex. There are too many possible variants, so you need to somehow reduce this search space. This group uses AI to narrow the search, test variants, and so on. Their goal is to identify all potential biological targets for human drug development within 2 years. If they succeed, pharmaceutical companies will be able to immediately use the obtained data to develop new drugs. This is a completely different, deeper approach to work. And it immediately gives researchers targets against which drugs can be created. I think this is very interesting. The combination of AI and a robotic laboratory that actually conducts experiments, not just models them on a computer. As a result, we will have fully robotic laboratories. There will be no human operators, but mechanical manipulators that will continuously perform operations like pipetting and do it around the clock. This seriously changes how the biotechnology industry is structured. If you take a scale from zero to ten, where zero is minimal interaction of AI and biotechnology, and ten is maximum, at what level are we now? It depends on, let me give a politically incorrect answer. It depends on your age. If you are a graduate student in this field, then now every PhD project uses AI in some way, in the format I described. From my perspective, as an AI specialist, we have won. We have full adoption of this technology. By the way, this is happening not only in chemistry but also in physics and even in materials science. But, as I said, in my opinion, AI is still underestimated, not overestimated. What do you mean? I'll explain now. But the main thing to understand is that the core of research work in the country is created by graduate students and postdocs. And all of them are already using these technologies. I personally am not able to fully understand most of the applications, but I see that the tools really work. But you say underestimated, even though I see news about AI every day. If you believe the media, they write about it literally every second. Now every company calls itself an AI company, even if it has nothing to do with AI. But they understood that such a label increases their market capitalization, and in the current conditions, this is very profitable. So, in this audience, everyone understands what the ChatGPT moment is. Almost everyone has used ChatGPT, but the fifth version is already on the way. Gini has released version 2.5, which surpasses competitors. As someone closely associated with Google, I am happy about this. Cloud 3 is also the best for programming, and all these models are roughly in the same class of capabilities. The Chinese model DeepSeek is also in this class. Grok, which has just emerged from the Memphis data center and is now part of Elon Musk's Twitter and XAI alliance, is also in the same class. Think of them as text-to-text systems. You ask a question, you get an answer. I was told that someone uses them for relationship advice, and another person used them for psychological counseling. I said, "Do you understand that these models were not trained by relationship specialists or psychiatrists? Moreover, calling them psychiatrists is likely illegal. They have not passed exams or undergone checks, but it seems nothing stops anyone." This shows that the power of these models is colossal. When I have a complex question, I just ask one of these services. They have differences, but they are roughly in the same category. But this is already a past story that everyone still considers relevant. The next big story is planning ability. Look, for example, at OpenAI R3 or DeepSeek Car 3. There is an amazing demonstration. You ask the model to show the workflow or conduct in-depth research. It shows how it moves up the decision tree. It tries one option, it doesn't work, it rolls back. It tries another, again failure. And so on, until it finds a working solution. Why is this important? Because this is how we humans think. Let's record this. Previously, AI simply answered questions. It was a natural language dialogue. Now, more advanced models can plan. AI has learned not just to speak, but to think with the same decision-making logic as humans. In fundamental models for biology, these methods have been used for a long time. Models predict DNA sequences, proteins, or chemical structures to predict how biological elements are structured and how they interact. I want to add my two cents here. Eric says that AI in biotechnology opens up amazing opportunities. It helps accelerate the discovery of new drugs, reduces experiment time, and finds the most promising targets for therapy. But there are also serious risks. Such technology can be used to create dangerous viruses and toxins without control and can help malicious actors in bioterrorism. That is why all of this must be strictly regulated and checked by law. Otherwise, experiments can be dangerous for people. Now, a major breakthrough has occurred thanks to reinforcement learning technology. This is when a model learns through trial and error. For correct actions, it receives a reward, for incorrect ones, it does not. Over time, AI learns to choose actions that yield the best results. Simply put, thanks to reinforcement learning technology, AI trains on its own experience to do everything more accurately and efficiently. And this is impressive. Over the next year, our industry estimates that the vast majority of programmers will be replaced by AI programmers. We also believe that within a year, AI capable of working at the level of the best graduate students in mathematics at leading universities will emerge. There are many reasons to believe this will happen, and this is our collective conclusion. I cannot solve such mathematics. Very few can. So how can a computer solve mathematics better than all people? Mathematics, unlike human speech, has a simpler and more formal language, so it is easier for a computer to understand it and solve problems. Algorithms are designed in such a way that they essentially learn to guess words. They take a sentence, remove one word, and try to put it back. They made a mistake, the model gets a penalty. They guessed correctly? A reward. This is the loss function that the model tries to minimize. Moreover, the model works with such a huge volume of text that it is simply impossible for a human. Thus, the model gradually learns to understand language and generate responses very accurately. In mathematics, models work a little differently. They take a hypothesis and a statement that needs to be proven, and try to build a proof step by step, proving why the hypothesis is true. And in programming, it's even simpler. You write code until it passes the test. Therefore, when I am asked what language to program in, I answer it doesn't matter. The main thing is that the result works. What code the machine generates is completely irrelevant. This is a completely new world, and it will be in just a year. And what will happen in 2 years? I have already told you about logical thinking and about programming and mathematics. And programming and mathematics are the foundation of the entire digital world. According to research groups like OpenAI, Anthropic, and others, right now, approximately 10-20% of the code in their research projects is already generated by computers. Technically, this is called recursive self-improvement. And what will happen when this process starts to accelerate exponentially? A great deal. In the next part of the interview, Eric Schmidt makes a prediction about when Artificial General Intelligence (AGI) will appear. He expects it to happen within 3-5 years. And, frankly, this is not very good news for people, because it is assumed that AGI will replace more than 90% of the workforce, but at the same time, it will create even more new jobs. So your survival depends on your adaptability. Will you be able to adapt and learn to live by the new rules? By the way, I am writing a book on this topic. I hope to finish it by the end of the year. So, subscribe to the channel, like it, and hit the bell to not miss its release. In 3-5 years, we may have Artificial General Intelligence (AGI). Smart like the most outstanding mathematician, physicist, artist, writer, thinker, politician, and so on. All this is impossible to combine in one person. But AI will combine all abilities, and all this will fit into one computer. Interesting, isn't it? I call this the San Francisco consensus because everyone who believes in it is in San Francisco. What will happen when each of us has the equivalent of the smartest person on any issue right in our pocket? This means that you will have, for example, the best architect in the world when you need to build a house. Another trend is the development of AI agents. Agents are systems that receive data, provide answers, remember, and learn. And here is an example. I want to buy a new house. I like Virginia. I grew up there. And I say, "Find me a house in the McLean area." One AI agent will do this. Look at all local rules and laws. Determine what size house I can build. A second agent will do this. Make a deal to buy land. A third agent: design a house with a human architect, but let him interfere as little as possible, and just sign the papers. And in the end, I approve the project. Find a contractor, hire him, pay the bills, and sue the contractor for poor work. Now I wrote this very simply and almost foolishly, but in fact, I have just described the essence of all business processes, all government processes, and all academic processes in our country. So not only programmers may be out of work, but all of us in general. No, not necessarily. I will touch on this issue later. But the reason I want to emphasize this today is that the foundation is being laid in the next year or two, and it is unstoppable. And then everything will become even more interesting, because computers will now begin to self-improve. They will learn to plan, and they will no longer need to obey us. We call this superintelligence ASI, or artificial superintelligence. This is the theory that computers will emerge that will be smarter than all of humanity combined. The San Francisco consensus states that this will happen within 6 years based on current scaling rates. But to make this a reality, we will need a huge amount of energy. I already talked about this yesterday. We need, well, I could talk for a long time about how many gigawatts of energy are needed, how many nuclear power plants, and so on. We can discuss this separately. But today, our society simply does not realize what will happen with the advent of such superintelligence. There are not even words to describe it normally. But I wrote a book about it. It's called Genesis. We wrote it with Henry Kissinger. I recommend reading it, after all, I wrote it. But the most important thing is that everything is happening much faster than our society, democracy, and laws can adapt. And this entails a lot of serious consequences. That's why many still underestimate the scale of what's happening. People just don't understand what the emergence of an intelligence of this level means, which will also be practically free and independent. That's the point. But how can we prepare for this? Well, at least we should start with a simple conversation. And, by the way, about work. Everyone thinks that automation will destroy jobs. But if you look at the history of automation, starting with the power loom 300 years ago, you can see that jobs have changed, but in the end, more jobs have appeared than disappeared. Therefore, you will have to convince me that it will be different now. If you look at Asia, where for various reasons people are having fewer children, the birth rate there is around one or even lower. Because of this, the population is rapidly declining. Therefore, Asian countries are very actively and rapidly transitioning to automation. The tools I am talking about will allow the few who will still be working in 30-40 years to do much more, and the rest will depend on these workers. At the same time, the productivity of those working will increase many times over. But we are not the only ones working on this. Tell us about the competition in this field. Firstly, the American model is large companies that you all know. For example, Meta recently released version Lama 4, which is also top-class. They operate a little differently and have done an excellent job. They released the model in an open weights format, meaning they actually showed publicly how the algorithm works. Other companies are all closed. These are serious business decisions, and each company chooses its own path. In China, there was its own ChatGPT moment, the DeepSeek moment. When DeepSeek appeared, our stock market lost a trillion dollars in one day. And then we understood the scale of what was happening. Now, China has a large-scale program for accelerated development of AI technologies. I and some other people in this room have worked very hard on controlling chip supplies. And, in my opinion, these measures have been generally effective. But how did China bypass them? Partly through blatant theft and tariff evasion. But the main thing is that they were smart enough to create new algorithms that use other types of computations. Now they are moving forward. Look, China works in an open-source format, so two things happened. We Americans immediately saw their ideas and implemented them ourselves. Thank you, China. You came up with it, and we immediately used it. But since it's free, the spread of Chinese models has become a very serious problem. This is an extremely complex task. And our government is still trying to figure out how to deal with it, and we need smart people to design all of this. I want to ask you about what is happening today. In the past week, research programs at a number of leading American universities have been suspended. There have been budget cuts and reductions in some key scientific organizations. Some foreign students have decided not to come to the US, while others, on the contrary, are leaving, fearing detention on the street. Some leading American scientists are already looking for jobs abroad and are actively being lured to other countries. Are we at risk of losing the brains needed to remain competitive in AI technologies and other emerging industries? At first, I thought it was just ordinary government political stupidity. But here are the facts. Last week I was in London and spoke with people who are ready to accept those leaving the US because they don't want to work in such an environment. And these are the British, our main allies, mind you. The harm from these 15% is related to the so-called indirect costs, money that universities spend on general expenses, laboratories, administration, utilities, and so on. The current government claims that universities inflate these costs. But this is not true. Such accusations interfere with the normal functioning of science. It turns out that the grant structure, created back in the fifties under Bush, is arranged as follows: employee salaries go into direct costs, and laboratory expenses go into indirect costs. If you look at all expenses as a whole, overhead costs are about 10-15%. This is confirmed, for example, by the Bill and Melinda Gates Foundation. I am a philanthropist myself, so I know these things. They account for full costs, not just a part. The government, however, decided to use this as a tool for a false attack on science. If the government has claims against specific scientists or specific research, please address them. But now it looks like an attack on all science in America. Why is this a problem? For example, the average American earns twice as much as the average European. Well done. Why? Because we are more innovative. And what exactly are we innovative in? In creating business opportunities based on science and technology. If I sound like a Democrat now, let me just remind you, fracking, which has been extremely successful in the US, has made us independent of oil and gas imports and even turned us into the largest exporter. The path to this was the same: innovation in universities, entrepreneurship, government support for 30 years. We have plenty of such examples. Another example of damage from current policies. Universities are so scared that they are freezing hundreds of millions of dollars and temporarily stopping hiring new employees. Imagine a graduate student who wants to stay at the university. Industry offers him $2-3 million a year, but he refuses because he wants to teach and develop the university. He calls, and they tell him, "We can't even invite you for an interview." As a result, he goes into industry. This is good for business, as he still remains in America. But we are losing the seeds of future кадров. Current faculty, not receiving research grants, will not be able to perform the necessary amount of work to obtain a permanent position. This ruins their careers. This stupidity will end sooner or later. It can be easily fixed, but by then it will be too late. The damage is being done today. I want everyone to understand, this is a real problem. We are competing with China, which is investing trillions of dollars in technology development. And meanwhile, we are sabotaging the funding of key people who should be creating our future. I've heard some say, "American science is being purged to the ground." Do you think so too? This phrasing is precisely used by university professors. If we talk about biotechnology, almost all research in this field has overhead costs of 65-85%. That is why budget cuts are so painful. To reduce these costs from 85% to 15%, the budget would have to be cut by about half. You can call it a purge, you can call it a halving, but it's about people's careers and programs that are simply stopping. Are you doing anything about it? There is a group of philanthropists trying to find a way to spend more money, which is always good. I, by the way, can also help. But the problem is that the amounts are too large. Private charity brings in hundreds of millions of dollars a year. But this cannot be compared to the billions that the state provides. Remember the old contract, made even before most of us were born? The essence of it was this: the state provides basic support for research. Universities conduct it, and venture capitalists create companies based on these ideas. Then the government provided the necessary market support to make everything work. As a result, industrial leaders emerged. This is the American way. Please don't change it. Now we will move on to questions from the audience. In a minute, we will give the floor to the audience, so prepare your questions, and I would like to discuss private investments. We have heard a lot about this today. What do you think about the current economic turbulence? About how it affects the willingness to invest in AI technologies? I think everyone here understands that business needs predictability. Any system where the rules change every week is perceived very painfully, as people make extremely long-term decisions. Most experts believe that we need 10-20 GW of electricity for AI, which means huge data centers. A regular chip consumes about 2 kW, so imagine the scale. Total investments are $50-100 billion. Building such data centers takes a couple of years. Then you have to wait your turn for GB300 chips or others from Nvidia. Economic uncertainty slows down this process. Everything is moving, but slowly and with delays. And all this hinders the picture of the future that I have just talked about. Another example that particularly worries me. Let's imagine you are a good person, and I am a bad one. Good is the US, which is ahead of everyone and doing everything right. And bad, let's say, is China, which is 6 or 12 months behind. The closer you get to superintelligence, the more worried I become if I don't catch up. You are surprised, what's the problem? After all, it only took the Soviet Union 4 years to replicate the atomic bomb. But here we are dealing with the business of network effects. In such fields, the leader gets up to 90% of the market. If you are good and reach the goal first, then you are likely to take 90% or more of all the intellectual power in the world. For me, the bad one, this is terrible. And what will I do? I will try to hinder you. Where will I start then? I will try to steal your intellectual property and your specialists. Exactly. But you, the good ones, have foreseen everything, and I failed. Then I will use my AI, which is almost the same as yours, to attack your system. These are so-called Adversarial Attacks. And you respond: "No, we have better cryptographers. We foresaw all this. I failed again." My next step: I bomb your data centers. Think about it, similar discussions are already taking place in Washington. For example, what to do about Iran's nuclear program? I am not an expert in this, but they talk about it in our country. When China is ahead of us by a few months, will we be ready to bomb their data centers? My favorite example. I spoke with a person who said, "The solution is obvious." What? I asked. We, the good and the bad, sign an agreement that attaches dynamite to each other's power grids. If you get angry, you blow up my electricity. If I get angry, I blow up yours. The idea is clear, right? Some will say that something similar is already happening, yes. But a real attack on a data center is already likely an act of war. Of course, no one will accept such an idea. I just gave it as an illustration. This is an illustration of what we call the needle's eye problem. We need to pass through a very narrow and dangerous stage without destroying ourselves and others, to reach the promised land of Dawn and AI and AI. By the way, regarding international relations, this report and other discussions talk a lot about the need for active cooperation with allies and partners in the field of technology. Let me return to the current moment and ask: will we see such cooperation? We definitely need to cooperate. If you look at the competition with China, this is precisely how it is perceived in Washington. Success is possible only with reliable partners. The best partners are Canada, the European Union, Israel, South Korea, Japan, and others. If you cannot understand this, then you do not understand that we are talking about large-scale business. Japan recently developed a new EUV technology. I don't fully understand it yet, but it's new physics that should compete with ASML chips used in Taiwan. This is great. Such competition historically gives us more choice in chip supplies, which are very
are important for the country and national security. Glory to the Japanese! Who would have thought, the idea is clear. We need to maintain close relations with countries like Japan, because working together is more profitable. But now this is not happening, and this is a mistake. Now we will gladly answer the audience's questions. Don't worry, soon AI will also write a question for you and answer them. Will AI be able to help treat cancer and create personalized medicines in the future? Yes, superintelligence can see what we cannot see. Therefore, it is believed that superintelligence will be able to understand biological and cellular mechanisms at a level unattainable for humans. We used to think that there would always be at least one polymath, a person who could comprehend everything. Perhaps in 10 years, we will find ourselves in a world where we will not fully understand how it works. But scientists will use such AI systems every day. When I was in college, I studied quantum physics, and my graduate student friend grasped it much better than me. I asked him then: "Is this physics even true?" It sounds too strange to be real. He replied: "Yes, we use it every day." Imagine, in 10 years, a young student will approach you and ask: "Is this superintelligence real?" And you will answer: "Honestly, I use it every day, but no one fully understands how it works." This is the kind of interesting situation that awaits senior researchers like you in 10 years. I have a question about equal opportunities. How can we ensure that small research groups and startups can also develop on par with large companies? The problem is that without GPU resources, Devs, and data infrastructure, it is difficult to verify or reproduce the results of other teams that have more resources, and this slows down scientific progress. We have a national AI research program, proposed by Stanford and supported by other universities. We support it because it will help universities get the necessary equipment. Currently, if you look at the scale of computations that companies use, universities simply cannot have the same capabilities, just as they never had comparable resources in physics or chemistry. But there is also good news. Open pre-trained AI models will be available that are powerful enough to work with. Universities will be able to take such models and fine-tune them for their tasks. And, I think, for them, this is the best possible solution. We often talk about AI and biotechnology, but there are also other exponential technologies, such as fusion and quantum computing. How, in your opinion, can we combine the achievements in these areas so that they reinforce each other? And how can we maintain competitiveness in all three directions simultaneously, developing them at the same pace? It would be great if we could achieve this, but I don't think we can maintain the same pace in all directions. To some extent, artificial intelligence is mathematics, and it obeys three laws of scaling. There is an excellent article by Dario about this. Machines of love and grace. The first law of scaling is what we see in ChatGPT and other models. If you simply add more hardware, data, and time, intelligence becomes smarter and smarter. The second law is reinforcement learning and planning, examples of which I have already given. And the third is training the system directly during operation, when AI learns itself directly in the process of performing a task. The last two laws are just starting to work. But it seems that the very core of AI is now on an exponential growth curve and can continue to move along it for a long time, as long as no one sees clear limits. We thought the limit would come when the training material ran out, but we have already downloaded almost all the knowledge humanity has ever recorded into the models. And yet, there is still enough data to generate new training materials for AI. It seems that AI will develop faster than other technologies. For example, in fusion, AI is already needed for design and control, especially of plasma. But the science itself there is not moving exponentially. With quantum technologies, it's a similar story, we will get there, but AI will most likely lead. One of the reasons why I chose programming and mathematics, rather than natural sciences, is that they do not depend on production scales. If there is enough electricity, you can simply perform more calculations and computations. You don't need additional laboratories, biology, or complex equipment. Therefore, I think the explosive growth will start precisely in AI, and then spread to other areas. I have been writing about science for many years and love science fiction, but I am skeptical about artificial intelligence. Therefore, I have a question: how realistic is the idea of superintelligence? We have been living in the era of chat for several years now. Yes, it's a revolution, but often it shows not superintelligence, but super-stupidity. If you change some parameters and give it only reliable sources with correct links, the result will be maximally simple and banal. Exactly what is already in the database. And this, you must agree, is the complete opposite of creativity, and certainly not like superintelligence. Large language models are most likely a huge distiller of human knowledge. They are already useful in a number of tasks. For example, tons of documents are created daily in Washington that no one has time to read, and AI helps to process and summarize them quickly. But here the question arises: how exactly will all this lead to the emergence of superintelligence, and not just to the development of more advanced media and communication technologies? My favorite example is that I am on the Hospital Board, and insurance companies send computer-generated letters to deny patients treatment. This hospital uses something like ChatGPT to automatically generate appeals to these letters. That is, computers write letters, and other computers write appeals to them. This is how the modern medical system works. Returning to the question of AI's stupidity, the fact that AI performs a task well does not mean that the process itself in which it is used makes sense. But I think that algorithms have become noticeably better in terms of generation errors, so-called hallucinations. The main achievement is reinforcement learning, which allows it to reason step by step and consider previous actions. Where the limit is, we don't know yet. An example of a question that we don't yet know and that will help understand the essence. What is the limit of knowledge? Although I am not a genius myself, I have already studied a lot, but sometimes I just lack ideas. And a true genius sees patterns in one area and skillfully applies them in a completely different one. For example, he notices the structure of prime numbers in one context and uses these tools in another area. This is the essence of genius. Currently, we do not have algorithms that work like this, but active work is being done on it. My answer to your question. I believe we will be able to solve this problem. And if not, we will still have a computer in your pocket, smarter than the smartest person in history. And that is already amazing. Today we talked a lot about data and how important it is for the development of AI and biotechnology. Earlier you mentioned risks, including possible mutual attacks on data centers. Can you explain why it is so important for data to be authentic and of known origin? I am concerned that most of the public data on which models are trained comes from databases where anyone from anywhere in the world can upload information. Often there is no reliable information about where this data comes from. From a security and reliability perspective, this looks very risky. The issue of data provenance is indeed very important. And AI can organize noisy or low-quality data. But if the information has been intentionally altered, it creates a new level of risk for the country. In biology, the main problem is the lack of data. There are many cellular processes, and to study them, laboratories are needed. And now the funding for such laboratories is being cut. These laboratories generate data that should be publicly available, verifiable, and peer-reviewed. Recently, Anropic released the MCP protocol, which has been adopted by all major companies in the last 3 months. It is an open-source solution, and it allows models to structure data in any format. Now you can directly access raw data through this protocol. This is a big improvement. No need to build a bunch of connections to different sources anymore. The model itself navigates raw data and answers complex questions. Can you tell us more about the risks in biology and how they can be minimized? What role do developers and AI play here, and what role does the state play? And most importantly, how can we safely move forward so that technologies bring benefits and not threats? You, of course, understand this much better than I do, but as an amateur, I see two main threats from these models: cyber risks and biological risks. Cyber risks are understandable; the system can create cyberattacks, including so-called zero-day vulnerabilities, which are difficult to detect and do so on a large scale. In biology, the danger is different. An attacker might take an open model and start using it for harmful experiments. Something like a dangerous virus scenario. A new Osama bin Laden, so to speak. Currently, open AI models are tested using special tests, so-called benchmarks, which search for and remove dangerous information. But in practice, these security measures are quite easy to bypass, and then there is a risk that the model will be able to create dangerous pathogenic viruses. The second risk is access to materials for creating dangerous bioproducts. Today, most people believe that the main threat comes not from individuals, but from states. Although the actions of terrorists cannot be completely ruled out. There are examples, for instance, in China, where dangerous bioproducts can theoretically be designed and produced. The good news is that making them simultaneously lethal and rapidly spreading is extremely difficult, but there is evidence that pathogens can be modified so that control and testing systems do not detect them. And this creates an additional threat. We are currently still slightly below the threshold of real danger, but in one or two cycle turns, that is, in about 3 years, given that one turn takes 18 months, these risks will become much more tangible. By then, many of you will be graduates and will be able to participate in solving these problems. Theoretically, can't AI and biotechnology help create countermeasures for these threats? That's exactly what I thought and made this argument. But, working in national security, I encountered the concept of the offense-defense balance. This is a situation where an attack cannot be stopped by the same force that is used for it. That is, the damage has already been done. And most biologists say that although models can be trained for defense, the damage from an attack will significantly exceed the capabilities of this defense. That is precisely why this topic causes us so much concern. I don't want to end on such a grim note, so let's look to the future. Tell us how AI and biotechnology can truly improve people's lives and what positive things await us. But first, I want to thank the financial system, the equipment manufacturers, and everyone who helped build the huge data centers with billions of dollars of equipment, even though it's not yet clear how it will pay off. Everything that is unfolding today will be used by smart people to solve real problems. I'm not talking about politics, I'm talking about fundamental scientific tasks. We will get huge databases with the information we need. For example, it has been impossible until now to create a digital model of a cell, a seemingly basic thing. I spoke with biologists, and they were shocked. What's wrong with you? We have been studying cells for 5,000 years, and we only know about 150 types. The answer is simple. It's really difficult, but we are close to doing it. And this will open up huge opportunities in medicine and pharmacology, allowing us to understand the language with which cells communicate with each other. I was the chairman of the board of the Broad Institute. It's a large project, and we are on the verge of this revolution. Scientists use AI to create these models, control the process, and AI helps them. And this is the correct order. Thank you. And thank you, friends, for watching this video to the end. I worked on it for quite a long time, so I would be grateful if you liked it, subscribed to the channel, clicked the bell so as not to miss new videos, left a comment, and shared this video with someone who needs it. See you in new videos. See you soon. [music] [music]