Harnessing AI for Good: Kira Radinsky's Vision for a Better Society

Hi everybody, thank you for the opportunity. Quick background about myself. I was born in Ukraine, I came to Israel when I was about four, and since I can remember myself, I always wanted to be a scientist. When I was 15, I started my bachelor studies mostly in computer science, and a lot of things I was excited about is how it can take everything humanity wrote for the last 140 years, identify patterns, and predict future events. This was the very early days of AI, so I started building natural language processing systems that take all this information, identify patterns, and predict events. At that time, I worked at Microsoft Research, I was part of the small team who was building some of the algorithms for the search engine, and in parallel I was applying predictions with the Gates Foundation. One of our most famous predictions: I was able to predict the first cholera outbreak in Cuba, the first one in 130 years. The way the system worked, it identified that if you have no water in a certain location for a very long period of time, if you have a drought, the probability of two years later having significant rains that will cause cholera in high probability is very high. It learned from hundreds of different other cases, and this is what happened in Cuba. It identified that there was a significant drought, the toughest one in the last 15 years. Two years later, there were huge storms coming in next to Haiti, and the system predicted in advance the probability of cholera. What was really interesting about that is, unlike many other predictions, there's a lot you can do if you know it's going to happen. You can send clean water in time, and if the water gets in time, mortality rates drop from 50% to less than 1%. We've been applying this work also in collaboration with the UN. The UN was asking us to apply those algorithms to predict additional genocides. Since World War II, there were around 35 different genocides, again this was a definition of a killing of 10,000 people and more. What the algorithm I developed identified is if you have somebody important in a country and they call a minority one of two words, that's all you need: cockroaches or mice. The probability of genocide almost quadruples in the next three years. We've been doing similar work in riots prediction as well. I was able to predict the Sudan riots that happened in the end of 2013, beginning of 2014. The pattern the system identified was the following: if you have a subsidized product and you stop subsidizing it, people are going to be very angry and go into demonstrations, specifically students, and if the police kills one of those students, you're going to have even bigger demonstrations. This was based on things like the Egypt demonstrations. Bread prices were subsidized, they stopped subsidizing them, people went to the streets, major riots, and they eventually changed the government. The one that did predict the Sudan riots was all about gas prices that were subsidized, they stopped subsidizing them, people went to the street, one of the students got killed by the police, and then major riots for two to three weeks. At that point, I understood that well, this is really cool, maybe AI works, who knows, and I started a business doing economic predictions. So I started working with B2B companies. Me and my co-founder Yon Zakai started a company called SalesPredict, and we worked with B2B companies helping them predict who's going to upsell, who's going to churn. But at some point, we understood that we have an AI system that gets all of the macro information, macro economy, we also started getting information from internal systems of B2B companies, private and public, and we know whether they're going to hit their targets or not, who's going to actually sell to them or not, because we have all of this information. I was hoping that maybe we can start building a theory connecting micro and macro economy. We sold our company to eBay. I was eBay's Chief Scientist from Israel for a few years. We were mostly building, structuring all of the data from eBay and enabling additional B2B interactions. Then a few years back, I said, well, I want to go back to doing global health, and I started a company with another professor from my university. He led a couple of companies in healthcare. We're both from the Technion, and we started Diagnostic Robotics, which is trying to predict which patients are deteriorating and what's the best clinical next step. This company at its maximum treated almost 28 million patients monthly. We partner with health plans mostly in the United States, health systems. At some point, one of my PhD students had this idea of, well, hey, maybe this AI thing can even create the drug, who knows. Me and another professor in chemistry were playing around with fun stuff, and we identified that we can create AI to build delivery mechanisms for mRNA. We can talk about this a little bit more. We started another company called Manab Bio. This company raised from NFX and from recent Horvitz, and have phenomenal initial results for where AI is building recipes that chemists can actually produce, and having a lot of interesting results that we already have in monkeys as well. And yeah, my latest one is, well, with all of the demonstrations happening in the academic front and people starting to see more and more anti-Semitism, how can we create an effect of consequences for those actions? One of the systems we started developing with my PhD students is how can we monitor the entire social media, identify anti-Semitic events, report them to the social media, but also building reporting for future employers, lawyers who want to be compliant with everything that they're reporting, and making sure there's no workforce harassment.

So many years ago, I was about to go on maternity leave, and I was approached by a very visionary officer in the IDF, and he told me the following interesting problem. The intelligence, the way it works, should change completely. If before one army would build intelligence with another army, the way they would do it, you try to monitor the head of that army. If you can do that, you monitor the people who are commanding underneath him, you can do that, you do the ones who are underneath him, but it's all a very small data problem. In other words, a few people can listen to a few people, and only one person is making the decision. In 2016, the world was starting to change. Some people wake up in the morning, they don't know they're going to be terrorists in three hours, and in three hours they're going to go and try to stab a soldier, a civilian, and perform a crime. It just started happening in hundreds in Israel all of a sudden. A lot of hypotheses started happening. Is this the social media impacting those people? What's going on? Why would somebody wake up one day and just decide to go and kill somebody else? Why would you do that? He was asking, can we take all of the data that are publicly and not publicly available and build insights into why would somebody become a lone wolf attacker? Can you explain? I'll explain. A lone wolf attacker is not affiliated to any organization, not affiliated to any terrorist organization, just a single person that says, you know what, I'm going to go and attack somebody. It's similar to shootings in the US. It's not like they're part of some organization. They go, one person goes into a school, starts shooting. Just to give that context, and that's similar to what we had with Trump. I mean, that one 21-year-old, he was supposedly not affiliated with anything, he just decided whatever. And you're exactly right. The problem is we cannot solve it today because you cannot monitor 300 million people and see if they're talking about terrorism or they're depressed, and we didn't even know what to look for. In Israel at that time, there were hundreds of those attacks, and they had all of the information about them. We started building AI systems. Again, my PhD students kind of went back to the Army, and we started building a team inside the Army whose job was to actually try to solve that. We identified two patterns that worked, received the Israeli Security Award, and additionally, it got the lone wolf attacks to almost zero. I wouldn't say this specific work on its own, but there was a lot of consequences to this type of work that together got to this amazing outcome of zero. One of the patterns that was identified and published was that strangely, there was a subset of pregnant women who were becoming lone wolf attackers. You would say this is the most uncommon profile that you could imagine, right? It created an even bigger problem because if a pregnant woman is coming and trying to stab a soldier or policeman, they usually don't have enough time to realize who's attacking, what not attacking, they're shooting, the woman is getting hurt, and then publicly a lot of protest starts because a baby was killed and a pregnant woman was killed. We identified the reason for their attack. It turned out that most of those women had outside-of-marriage relationships, and in the Islamic environment, if a woman has an outside-of-marriage relationship and they find out about this, eventually they're going to kill her. The hypothesis the AI identified was that those women are doing this in order to escape killing of their own family and try to go into Israeli jail in order to be protected, or they understand they have no other choice, they're going to die, better to die martyr than to die nothing. So one of the actions that was done is identifying those women in advance and creating women shelters to prevent them from getting to a point where they are in such despair that they actually go and try to stab somebody to actually find some form of a shelter. This is just a very interesting and surprising example, but there's hundreds of others. It turns out that many times it's depression that's driving you, and it also aligns with the times that we've seen some of those attacks in the economic environment. Depressed people try to find other reasons to be worshiped by others, and there are very distinct patterns of identifying behavioral health problems online that are very easy to identify if we had, let's say, hundreds of millions of psychiatrists watching every social media person, but we don't. This is where the AI scales very well. It will not do it as well as a psychologist or a psychiatrist, although these days I don't know anymore because the developments were really strong lately. But even today, the ability to identify depressed or suicide attempts online are much stronger, so you can use those to save those people but also to prevent what we've been seeing, like those mass shootings. David, what you mentioned about the shooting of Trump, probably those events, and I'm not an expert in what specifically happened with Trump, but probably patterns on the social media were already available.

I'll give you an example from a work we've done that was more on the police side, which I can talk a little bit more about the type of information. This was a small pilot. We received data about historical crimes, mostly property crimes, somebody broke in, somebody stole something. The police was asking, can we predict those property crimes before they happen? The way they were trying to prevent this is they had enough information knowing that if a police car drives around a certain area, it's going to scare the criminals, and they wanted to change the directions in which the police cars are driving. The places where there's a lot of property crime, they already knew. Where there's no property crime, they knew. They were asking about everything in the middle. Can AI help them prioritize their driving path every month? What we did, we took all of the information about those locations, and I'm talking about racial information, average income, social determinants of health of the people who are living there, and all of the Twitter firehoses, like what people are talking about in those locations. It's a statistical game. What we do, we build machine learning models. Think about them as a black box that just goes through all possible patterns in a smart way, I wouldn't say all, but in a smart way goes through a lot of patterns and statistically tests if there's statistical significance in where property crimes happen versus not. One of the patterns that identified was the following: if it is good weather, it turns out criminals don't like doing crime when it's too cold or when it's too hot, right? Probably fair weather criminals. Also, they don't come when it's raining. I'm just kidding. I think people are staying more at home, and nobody wants to do crime when somebody's at home, right? But when it's good weather, and it turns out if there is a report of an abandoned car or an abandoned building in that area, like people are complaining about that on Twitter, there's higher probability of crimes happening in those locations, which by the way makes sense, right? They feel those places nobody's going to visit, nobody's going to check, if everything is abandoned, there's high probability they're going to be able to come in. How did the system identify this? It just went through all of this data, went through all combinations of weather reports, it just looks at those words, and then when it identifies something statistically significant, it just alerts about this, right? So it's as simple as that. Of course, there's a lot of depth because you cannot just go through all possible combinations, there's some guidance around it, but it's just checking statistical significance eventually.

Search is more like information retrieval, it's like giving existing knowledge, retrieving it. Here, it's creating knowledge. The difference is if you had infinite amount of time, you would go through all combinations of forces and weather and income, right? You just don't have infinite amount of time. The biggest difference is that it automates that. It just goes through some of those combinations and applies significance testing. I'm oversimplifying because if you try to even explain how large language models work today, what do they do? All they do is look at a bunch of words and predict what's the word with the highest probability to appear next based on all of the stuff we wrote online, and we wrote a lot of stuff. You would say, how would something that just predicts the next word based on previous words can be intelligent? It doesn't make sense. There was a movie, I think with Tom Cruise, that they arrested criminals before they committed the crime. It was Minority Report. Yes, yes, we were inspired by that movie, although we were building the stuff for the police and the Army. But here, think about how simple it is, predicting the next word. But it turns out that more lately, I think it was somebody in Microsoft Research that were doing actually IQ testing. In 1989, somebody decided what's the best way of testing if something is intelligent, and they went through intelligence testing through those algorithms. Just predicting the next word, and it turned out it passed a lot of them. There's still differences in things like planning, which is not doing really well, but there's a lot of academic focus on improving that as well. The Turing Test, many years ago, we said we're going to do a test: a human can talk to an AI, a human can talk to a human, and if it cannot identify who's the human and who's the AI, we pass the Turing Test. It was always kind of something we will never achieve in AI. We always talked about passing the Turing Test as something almost impossible, and here you go, we passed it. But if you ask, what does it do? Nothing, right? It just predicts the next word. It just goes through all patterns. And remember, before you want to give over complete control of the planet to the intelligence, just don't switch that button on. Just remember.

I would say AI is already controlling many of the things that we're doing for many, many, many years. I think at this point, we just started seeing more advanced versions of it because algorithmically, I don't see huge changes in the algorithmic design of what we were doing since the 60s. We just have more computing power. I think the biggest jump, if you would ask me, what happened in the last couple of years, we just started pouring more money into training those algorithms, more data, more computing power. That's it. Nothing else. Algorithmically, there was no huge algorithmic jump. Now, will it take over the world? Well, you have automatic pilots in planes. It is considered a form of AI. It's more rule-based, but it doesn't matter, it is a form of artificial intelligence. I got my Tesla upgrade that if I'm somewhere else, I can command my Tesla to come and pick me up. Yeah, me too. Isn't that nice? But I'm too nervous. The interesting thing is people are making more mistakes than autonomous driving. It's just, I think a generation needs to change in order for all of this. I think the biggest problem is how will it drive when there's humans, and I think this is the biggest question. Even in trading, first people started trading, then algorithmic AI started trading, and now you have humans and algorithmic AI trading together. This transition phase is something which is extremely interesting because humans don't always perform logically.

Well, sometimes predicting somebody else's emotions and acting accordingly, that's maybe what some AI algorithms can be doing. If you would have asked me, let's say, five years ago, what would have happened with AI and taking over the world, I would quote something I heard Andrew Ng say. He said, well, it's like talking about overpopulation on Mars. Eventually we'll get to Mars, eventually we will overpopulate it, it's just too early to worry about that. I do think we're still not exactly in those directions. It's like your autopilot in a plane is not designed to kill you, and you never worry about that. Why would you worry about a system that predicts the next word trying to kill you? Nevertheless, I do want to talk about the real risks that I see for tomorrow. It's very hard to predict what's going to happen in 10 years with the pace of developments. The developments are super quick these days. Every day we break something else that we never thought was even possible yesterday. But in the next five years, I start seeing more and more those generative AI systems being used as almost weapons. I think countries cannot allow themselves to not have that weapon anymore. In this race, it's mostly United States versus China right now. You do see a lot of advances coming in from France and other countries should definitely take it seriously. You would say, why would a generative AI system that just creates text even be considered a weapon? But what it does, it creates public opinion. When you start creating wrong information and influencing people, and we go back to emotions, you can influence what they do, what they vote, and how they will act. I think this is a much stronger weapon than many others we experienced before, and it's a complementary weapon to what we've ever seen before. Even you see this right now in the war between Israel and Gaza. One of the strongest weapons being used by one of the sides, from the Palestinian side, is the PR war. Israel has historically never been really good in anything to do with relations. It's not their strongest side. It's a country with a lot of engineers building systems, very low on the marketing and sales sides. We would see this across most companies. This is why Israeli companies are usually split where the go-to-market teams are in the United States because this is where the market is, and the engineering always stays in Israel. We see the same thing in how you handle wars. The same thing you see with Ukraine and Russia, the amount of disinformation coming in from this war to impact public opinion is considered yet another system of war. You can actually get weapons and negotiate weapons from other countries and how the public opinion is supporting them with funds.