Jeffrey Hammerbacher was one of the first data scientists in Facebook and later in Cloudera. Very intelligent individual. Once said, the best minds of my generation are thinking how to get people to click on ads. And this is sad. And true at the same time, considering the world faces great challenges.
Millions of children die before they get their fifth birthday. Hundreds of millions are affected by climate change. And 1.6 billion people live with a severe disability. The world needs all the help it can get.
And technology and AI can make a difference. But the problem is the organizations and non-profits around the world that are working on these problems do not have the tools. the capacity or structure to hire the AI talent. Majority of AI talent today works in the tech sector or the financial industry.
But what Hammerbacher didn't consider was that even though the problem of making sure understanding which child has higher chances of infant mortality or understanding which people will click on your ads, even though these two problems cannot be further apart from a societal point of view, from an AI and data science perspective, these problems are basically the same. problem which means that if in your organization you have people that are really good making sure people click on your ads you can use the same talent to help with some of the challenges and that's what my team does we partner with organizations that have subject matter expertise and we bring AI to the table today I'm gonna be covering some of the lessons learned from these four years working on the air for good space and the first lesson is for some world problems Relying on AI is the only option we have. Diabetic renal body is the leading cause of blindness around the world. 450 million people suffer from diabetes. A third of them will develop DR. And if they are not diagnosed and treated, they will become blind.
But the problem is the world only has 200,000 ophthalmologists. So it's not physically possible for these 200,000 ophthalmologists to diagnose this disease. But today, our AI models are as good, are on par with a very good ophthalmologist, and these models are in production already in some organizations and some countries.
This is an example where AI is not just a solution, but it's the only solution we have. But AI expertise alone cannot solve these problems. We need to partner with subject matter experts. Low birth weight is one of the leading causes of death in the world. And if you have a data set that has low birth weight babies, and you train an AI model on top of that data set, what the model will tell you is that if the mother of a child smokes, the chances of survival for that baby increase.
But of course, this is not because smoking is helping. Actually, smoking is a cause of low birth weight, and a lot of children die because of that. But among the causes of low birth weight, smoking is less severe than other causes.
AI models are great for understanding predictive power, but cannot tell us anything about causality. This is why it's key for us to partner with subject matter experts, and that's what we do. For every project we have, we partner with subject matter experts, and we work in teams together.
As humans, we are addicted to complexity. We like complex projects, we like complex things. This is the reason we put a man on the moon before we added wheels to your luggage. This is also the wrong addiction.
If you want to impress people and look intelligent, your solution can be complex. But if you want to have an impact in the world, your solution needs to be simple. And building simple solutions is hard, but it's certainly worth it. This is something that we learned closely working with SEEDS. SEEDS is an amazing organization in India, and they work with families to better prepare them for climate change.
In India alone, 89,000 people die every year of high... temperature and in the scenario of 4 degrees Celsius increase this number could increase to 2 million. Their solution is very simple they need to understand which of these what is the material of the house of each family and there's a lot of papers written about this and how to get the material of a house using satellite data But when we talked to them, it became really clear that none of those solutions worked.
There's a big difference of solving a problem on paper versus solving a problem in production. The distance between theory and practice is larger in practice than in theory. So what we did is using high resolution satellite data, we built an AI model that could detect the material of each of the houses, and with this information we provided the maps.
Seeds and with this information is helping seeds save lives One thing on AI models we learn that all of us need to pay a lot of attention is we can get fooled by bias Let me show you what I call the left-handed dilemma if the Probability of being left-handed today in the room is independent of being in the room. The distribution of left-handed people in the room will be similar to the distribution of left-handed people in the population. Meaning 10% of you will be left-handed and you will see the same amount of people on the right side or on the left side. Please raise your hand if you're left-handed. You see, it works.
I have really bad news for the left-handed people. There's a paper published by Halpern and Koren from California State University and the University of British Columbia where they took a random sample of people and asked the family members whether they were left-handed or right-handed. What they found was really disturbing.
Left-handed people were dying nine years younger than right-handed people. This was published in the New England Journal of Medicine, that is the most prestigious medical journal in the world. If this was true, being left-handed would be as bad as smoking 120 cigarettes per day.
Really bad. Well, the good news for the left-handed people is that what the researchers didn't consider was for a long period of time, there was a discrimination against left-handed people, and parents would force their kids to be right-handed. And I know that because my grandfather was one of them. He was forced to be right-handed. be right-handed.
Eventually they stopped doing this and this generated this artificial increase in the left-handed population and this artificial increase is the one that gives us the illusion that left-handed people are younger when in reality that's not the case. What's the problem from an AI perspective? is that if you're a life insurance company and you train a model using this data, what the AI model will tell you is that you need to charge more to the left-handed people because those people will die younger. Majority of the data we collect has some biases.
We need to understand the biases and we need to understand the risk of those biases. AI can make a difference. And I'm going to share some of the examples. We have witnessed this through these examples. The first example is what happened after the earthquake in Turkey and Syria.
We partnered with Planet. Planet Labs is an amazing company. I think Andrew Zoll is in the room today. And we have this very strong partnership with Planet where they provide the satellite data and we bring the AI expertise.
We partnered with them. And we're using AI on top of satellite data. We created maps of every single city and every single building that was affected by the earthquake.
the earthquake. These information and these maps were used by people on the ground to help them prioritize resources. These type of models can save lives.
We are also using, also in collaboration with Planet, we are also using satellite data to monitor the Amazon, understand deforestation, understand illegal logging. Also in the Amazon, we are using AI models to measure and... understand biodiversity. We have camera traps and we have AI models working on these camera traps and from these, these help researchers understand and measure the biodiversity. We also use acoustic data for this, where we have recordings in the Amazon and we have AI models that can detect and track these animals and measure the biodiversity of the Amazon.
In the health space, retinopathy or prematurity is the leading cause of blindness in children. ROP is a disease that didn't exist four decades ago, and suddenly, we are seeing an exponential increase in ROP around the world. The reason this disease didn't exist is because this disease affects very small, premature babies that four decades ago wouldn't survive. Now, because of improvements in health and in medicine, a lot of these babies are surviving.
And this is what is creating this exponential increase of these cases. And we only expect these cases to continue going up because we will have more and more improvements in health. A lot of these babies now will survive.
but if these babies are not treated by ROP they will become blind. And we don't have enough ophthalmologists and we don't have the expertise to work on this. AI models running on smartphones today can already do the understanding whether the child has a disease or not.
This will make also a big difference. At Microsoft, we continue investing in Seeing AI. Seeing AI is an app that is used by Empowers.
individuals with visual impairments by providing real-time feedback. They are using these devices to see the world. Saqib Shaikh is a friend of mine, and he leads, he's blind himself, and he leads this team. For him and many others, this technology is a game changer.
But technology has the potential to create huge impact, but it's essential for us to understand and work this in a responsible and safe way. If you ever read a book, a children's book written in the 1800s, if you compare them to those books written now, you will see something that is different. A lot of the people in these books die.
They have brothers or sisters that die. They're very different to what you see today. And when you ask why, the answer for that question is because that's exactly what was happening.
In the 1800s, infant mortality was 30%. Out of 10 children... three would die before the fifth birthday. The average life expectancy was 35, and people have this illusion that people will live through 35 and then die. The answer is no.
If you get to 30, you could get to 60, 70 years old. The only reason why the average is 35 is because 30% of the population was dying between zero and five years old. This is shifting that average. Today, there has been a significant improvement, and 200 years after that, infant mortality around the world is still high, but it's 2.6% that is one order of magnitude lower than it was before. And a good portion of this improvement, directly or indirectly, was tied to electricity.
Through healthcare facilities, vaccination, medicine, storage, clean water, sanitation, refrigeration, electricity played a fundamental role in this. And as one of the things that is important today, some countries would consider electricity a human right, and a lot of human rights depends on electricity. But the fact that we have electricity massively available, was debated at one point in society. There was the whole idea of the War of Currents or ACDC. ACDC is not a rock band, although the rock band name comes from ACDC.
ACDC stands for Alternative Current and Direct Current. Alternative Current is what you have in your house. Direct Current is what you have in your remote control.
These have direct current. In Alternative Current, You can see large-scale power distribution. You can distribute electricity for very long distances.
You can electrify cities. On that side, we have George Westinghouse and Nikola Tesla, proponent of this. On the other side, we have Thomas Edison.
Thomas Edison was about safety. He was a big proponent of DC. And if you look at the arguments that Thomas Edison had, by the way, he was one of the inventors of electricity, his arguments were right. He was so vocal against AC that he would electrocute animals to show the risk of AC.
A bit cruel for me, but that's what he was doing. Well, what Thomas Edison couldn't consider was that society could work together to come a way to make electricity safe. Technology can be used as a tool or a weapon.
Every technology can be used as a tool or weapon. It's important for us to work as a society collectively, making sure technologies are safe and benefit everybody. And that applies to all technologies, but especially applies to AI. To close, large language models bring an incredible potential for AI for good.
Majority of human knowledge is stored in the web. The web is our most important data source. But the bulk of that knowledge is stored in unstructured text that is good for humans to consume.
But until very recently, we couldn't convert that data into information for machines to help us in that process. Thanks to large language models, we can now do that, and that is already a game changer that will produce even more impact in the future. But large language models can do other things. They can bring an amazing power as a language aid. As you likely realize by now, English is not my native language.
This is when you put the face of surprise. But I work in a job that requires me to not only speak in English, but also write in English in a flawless way. Especially if you're going to publish in a scientific paper.
Otherwise, reviewer number two will say, a native person will need to review your paper because you don't know how to write. I have gotten some of that feedback myself, by the way. 95% of scientific publications require you to write in English, but we live in a world where only 4.7% of the global population have the luck in many ways to have English as the native language. For the rest of us, we're kind of at a disadvantage. But thanks to large language models, today we can express ourselves, we can express our ideas, and we can write like a native speaker.
This is already making a huge impact to me, but to many more people. And this shows the power of these large language models. But also with coding. Coding is an amazing superpower that only 0.5% of the population have.
And coding requires you to learn another language that is not even a language. You need to learn something like Python or Java. And that takes a lot of effort.
Thanks to large language models, we can now code in our own language. Whether it's English, Spanish or Chinese, you will be able to code in your language. This will increase the power of having the ability to code without knowing a programming language. A few weeks ago, there was a paper published in JAMA where they asked questions... To doctors, they have questions answered by doctors, and they have questions answered by a large language model.
The large language model's answers were not only more accurate than the doctors, but actually show more empathy. We live in a world where 4 billion people do not have access to medical doctors. The alternative for these 4 billion people is not a human doctor. It's usually no one.
For a long time, all of us... We've been asking ourselves when it's ethical to use AI in some of these scenarios. There will be a time in the not so distant future we will need to ask ourselves when it's unethical not to use it. Thank you very much.