AI for Good: A discussion on addressing real-world problems with AI

Sam. Welcome to this panel discussion on AI for Good. I'm Ryan Keenan from Deep Learning AI and I'm really glad that you can join us for this event. We have people joining us from all over the world today, at least from the signups, I think we have people representing more than 140 countries. So good morning if you're in the Western US like me, or good afternoon, good evening, or maybe even good middle of the night to you if you're somewhere else altogether, wherever you are. We're glad you joined us and we're glad to have you with us here. Much of the conversation about AI these days has focused on cutting edge technologies like generative AI in particular, as well as the potential risks that these technologies could pose to society. Risks like the potential for AI to replace humans in certain jobs, or the concern that AI algorithms could generate content like videos or text that would be hard to distinguish from human generated content. These are real risks and of course, need to be taken seriously. At the same time, AI can be a powerful tool when it comes to addressing the biggest challenges we face in the areas of climate change, public health, and disaster management. For example, AI has already been deployed as a key component of many real world applications that are having a positive impact on things like making renewable energy more viable as an alternative to fossil fuels, or providing access to reliable information in more of the world's languages and allowing for more effective planning, logistics and resource allocation in the event of a disaster. In today's discussion, you'll be hearing from a panel of experts who have each led a wide range of real world AI projects with positive outcomes for society and the environment. To get us started, I'll give a brief introduction to each of our panelists and then we'll jump right into the discussion. Toward the end of the hour, the panel will take some time to answer your questions. And so to ask a question to the panel, please check out the live chat link and type your question in. And then please also take a moment to upvote the other questions that you find most interesting. So when we turn to Q and A at the end of the event, then we'll start with the most upvoted questions and get through as many as we can. We at DeepLearning AI are excited to bring you today's event in connection with the set of three courses, the AI for Good Specialization, that we recently launched on Coursera. And at the end of the event today, I'll be showing a quick demo of some of the course content and saying a few words about how the content of today's discussion connects with what's presented in the courses. So with that, let's get started. We have a group of fantastic speakers joining us today. First, I'll introduce Robert Monarch. Robert is an expert in combining human and machine intelligence and is the author of Human in the Loop Machine Learning. He's founded his own AI startups and developed AI products in places like Google, Amazon, Microsoft and Apple, where he's working today. Robert has also worked for more than 20 years in applying AI to address critical problems in the areas of disaster response and public health around the globe. Glad to have you with us today, Robert. We also have with us Juan La Vista Ferris. Juan is the Vice President and Chief Data Scientist of the AI for Good lab at Microsoft. Juan leads a team of dedicated data scientists and researchers in the domains of AI, machine learning and statistical modeling. And the team partners with domain experts, researchers and organizations around the world to create collaborative ecosystems that drives progress toward addressing some of the world's most pressing challenges. We also have Eva Gumin. Eva is the founder of Humans in the Loop, a social enterprise which provides ethical AI data annotation services. To date, the company has provided online jobs and trainings to more than a thousand refugees and conflict affected people across Europe, the Middle east and Africa in the field of data labeling and Human in the Loop pipelines for AI. She's also a board member for the Group for Humanity, which is a nonprofit and nonprofit focused on the audit and certifications of AI systems. Glad to have you with us, Eva. And finally we have Andrew Ng, our founder at DeepLearning AI, managing general partner at AI Fund, and CEO of Landing AI. Landing AI, among other things. With that, we're going to jump right into the questions we have. As I said, people joining us from all over the world. Our speakers are also distributed around the world. We have Eva joining from the Czech Republic right now, I believe, and Andrew happens to be in Hong Kong. So glad to have a global audience and a global panel with us today. First off, I just wanted to pose a question or a prompt maybe to you, Robert, to get started. And this is kind of just get one thing out of the way right away here. I've talked with you a bit about this and also folks actually at the Microsoft AI for Good lab about this notion that AI for Good is really just sort of the most recent version of a centuries old pattern of humans using the technology that they have at their disposal to try to do some good for the communities that they care about. So could I Ask you, Robert, to comment on that first. And I'd also like to get thoughts from Juan after that from the AI for Good lab at Microsoft.

Yeah, happy to.

And hi to everyone joining. I'm looking at all the comments of people joining online. Hi. Ready? Who's in San Francisco. I'm around the corner from you today apparently, and I saw someone else said they were joining from Earth. If there's anyone joining from the International Space Station, I would love to. I'd love to hear that too. Yeah. So I think obviously AI is really in the news lately, but both as someone who's worked in industry for a long time, like, like today I run machine learning teams at, at Apple and, and someone who's worked in disaster response and, and public health for a couple of decades now, AI is not that much different from the other kind of technical tools that you might use. So, you know, for as long as we've been inventing new technologies here on Earth and maybe in the International Space Station, a lot of good people have been thinking about the best ways that we can use new technologies to have a positive impact on the world and at the same time be really cautious about the ways that introducing the wrong tool or introducing the right tool in the wrong way could amplify imbalances in inequality or amplify biases in the world. So it's really exciting to be on a panel like this today and think about the ways that we can play AI to have a positive impact on the world. But I don't think we need to get too, too far ahead of ourselves in our excitement about the amount of impact that AI can have. Even if, like me, you're very excited about the technology itself.

Thanks. Yes. And Juan, I actually remember talking about this very thing with some of the folks at the Microsoft AI for Good Lab. Do you want to, do you want to also comment on that thought?

Clearly there's nothing new in trying to help the world. I think that's something that I think comes in majority of the people on Earth and this is much more common than we think it is. And I think that in many ways AI is helping allow people to have that impact. I think that in many ways I consider myself really lucky to be working on this space. I'm not the only one working on this space. We are not the only one working in this space. But clearly there's a lot of people around the world that would love to work in this space. And I think AI is allowing them now through AI and through other type of technologies, allowing People that are not necessarily near those problems or near those issues to make an impact. And I think that's one of the things that has been changing throughout the years and just through technologies is our ability to influence and have an impact in world problems with, without necessarily needing to be there, basically. So I do think that's a big change.

That's a great point. And just that makes me think of a sort of adjacent point that when we think of AI projects that are doing some good in the world, I think it's important to remember that that's not just on the AI impacting communities, but also on people impacting how AI systems work. So, so Eva, this, this makes me think of the work that you do. Could you comment on that aspect of how AI could have a positive impact? And I want to ask you more questions in a moment about, about some other things related to your work, but just, just that as a thought that, that the impact of AI, a positive impact of AI in the world is not necessarily just about how the AI systems impact.

Yeah, absolutely. I think it's very important to think about the positive intent AI systems and their purpose of creating positive impact. But it's also important to think about all of the external, let's say, stakeholders that participate in the AI industry and how it affects them and how they affect the AI supply chain. And we're active exactly in this space of preparing data for AI for a lot of projects for AI for good, but also for some kind of like neutral projects that are not necessarily explicitly for good. We are definitely against working for AI for bad purposes. We do not approve of any such projects. But it is important to think about the entire supply chain that you're using whenever you're designing and developing an AI project, whether it's for good or for neutral purposes, because it also has a tremendous impact on the environment, on the workforce that you're using, and so on. So it's really important to think about these externalities that your project has, even if it's not explicitly for an impactful purpose.

Absolutely. And so you mentioned that there are those projects out there that are explicitly for good or for some sort of positive intent for a particular community or for a particular cause. There's also industry projects maybe that are, that are for specific goals of a company. And so actually, Andrew, I wanted to ask you about that specifically. You've worked on AI applications across many different domains, including healthcare and renewable energy. Some things that, that might be in what people are thinking about when they think of AI for good also More industry related projects, manufacturing as well as many other things and pushing the boundaries of what's possible in AI research and what I've heard you talk about the phases of an AI project or the considerations that you need to keep in mind from scoping something out to designing and implementing a system. I don't hear you saying a lot about what the domain specific considerations are or it's like whether you're doing something for industry or whether you're doing something for some other explicit purpose. Could you say more about maybe the commonalities or differences across building AI solutions for different areas of.

Yeah, I know it feels like I'm probably supposed to say oh they are so different, but honestly I see more commonalities than differences. But I think that AI projects are so diverse. This is useful for so many things. Maybe I'll just highlight one difference that I see among this really very diverse AI project, which is that we'll work on AI projects in sectors such as healthcare, disaster recovery, climate change. It's directionally less likely that the recipe of ship something quickly, see what happens and fix it off to. It's not that that always works even for commercial projects, but that's just even less likely to work. One of the wonder and fun things about machine learning is you almost never know what you're going to get. What hyperparameter works, what's in the data set, how much data do you need, how was the algorithm? A lot of these things we just don't know. So machine learning has developed a methodology that was inspired for me probably by the methodology of Lean startup where you build something, see what you get, test it, get user feedback and then see what happens. And in fact, the recent trend to a generative AI has even accelerated that, because rather than taking a lot of time to collect data, train a model, find a cloud service to deploy it, now you just go to an OM and you prompt it and then in minutes you can ship something. So there's even this trend for some classes of commercial applications where there's no or no observable risk of harm. Let's just build 10 things and ship them all and see what happens and see what sticks. And this actually works in a commercial setting. If you can do that in a responsible way without signaling risk of harm, the pace of iteration is just very fast. But in these sorts of AI for societal good applications, where you're more likely to have complex multi stakeholders working with vulnerable populations or take healthcare, it's clearly not okay to ship something haphazard that then damages someone's health because you can't actually easily go backward on something like that. I find that the approach to addressing multiple stakeholders, this needs and ensuring safety before you then ship it, that that just needs a different set of practices. So that's maybe one average big difference that that I see in these sorts of projects. But I'm sure there are others as well.

Yeah, thanks, thanks for that. I want to kind of stay on the thread of, of like explicitly AI for good projects versus AI for industry, other applications, building, you know, building products at a company. Robert, you've done a lot of work in the sort of explicitly projects for good space as well as a lot of work in the building products and industry space. And so you've done work in areas of disaster response and public health, in many cases combining human expertise with technology. Like your work in Mission 4636 after the earthquake in Haiti in 2010, where in that case text messages from survivors were quickly routed to members of the global Haitian diaspora for translation and geolocation and then sent to responders. In industry you've built products like translation services, for example. And I've heard you say that you feel like perhaps some of those industry products you've built have actually had more of a positive impact in the end than some of the more kind of boots on the ground work that you've done working with refugees or something more direct. Could you say more about your thoughts on that?

Yeah, absolutely. And that's right. And I'll double down what Andrew said as well in that you don't want to be iterating on technology by trial and error in sensitive situations. So I guess a good analogy is that if you're working in refugee camps like I have, you wouldn't iterate on a self driving car. It seems ridiculous that you would do that in a sensitive setting like that. But I think sometimes we get too excited about other forms of AI, whether it's for healthcare or disaster response, and maybe don't realize that the negative impacts we can have by trying to rush out these kind of technologies. But yeah, to your particular point, yeah, I've done a lot of work as a disaster responder directly. But a lot of what I see is people, whether they're victims of sudden onset disaster or suffering from public health emergencies, the most benefit they get is from the existing technologies that they're using. So their ability to go online and find resources to help themselves, to understand where they can get food or water, to know basic things like when schools are opening and all of this information online has to be supported by the building blocks of technologies. So I'm very proud of the work I've done as a disaster responder. But I feel like the work I'm doing now at Apple, which is really focused on privacy preserving, machine learning work I've done previously, like when I was head of product for AWS's first NLP services, when I focused on trying to get them into as many languages as possible. I think these kind of building blocks of making AI just more generally usable across industry is probably where I've had the biggest impact. I'll add one more thing there as well. I think Eva needs to be way more famous than she is. She runs a profitable company employing more than a thousand people who for the most part are refugees or internally displaced people. And I think that is another way that in a very much a for profit environment, that AI can have a very positive impact.

Absolutely.

Thanks Robert, for this. I really appreciate the support that you always give to our organization. I would say that we were kind of part of the hype of tech projects that appeared with the big crisis that happened in Europe around 2013, 2017. There was an entire organization called Tech Fijis that still exists today that was created in order to kind of coordinate tech efforts around solutions for refugees. And unfortunately what they found was that a lot of these projects just died out because of too much hype and a lot of people creating tons of different apps that you had to encourage users to download and to get onboarded on and so on. I would say, Robert, you're definitely right that it's better to meet users where they are and just design solutions that are useful for them instead of getting obsessed with a novel technology that can solve all of their problems. Suddenly for us, it definitely worked out because we just were trying to meet people where they were in terms of their needs for immediate employment and a source of their livelihoods.

Absolutely. Thanks for that. And I'd like to continue asking a little bit more about the work that you're doing. As you mentioned, you're providing work opportunities for people who, for example, might be subsisting on humanitarian aid or might be have been displaced from the result of armed conflict or something else. And that sounds fantastic. I'm wondering if you also have critics who say something like, well, this isn't meaningful work opportunity or this isn't such an incredible thing. Could you say more about the actual impact on the people that do this work and the folks that are taking these opportunities with humans in the loop?

Yeah, I mean It's a very fair criticism. I would say. There has been a conversation happening recently around companies that are trying to channel such work to vulnerable communities. This is essentially very manual, menial work that does not require very high level of skills. So that's why there are some organizations that see in this an opportunity to generate a positive impact. But there's also organizations that see in this an opportunity to just find the cheapest labor possible and to essentially exploit them in order to maximize the cost savings. In our case, we're definitely trying to subscribe to the first option of using this to empower people and to create opportunities for them to secure their livelihoods. Of course, we could always do better. But I would say the biggest impact that we've had to date is that we're explicitly going into geographies and countries where no one else is operating. Literally no one else is even daring to go in there and to try to develop online businesses or to offer online work. There's no PayPal, no Upwork, no freelance opportunities here. We're talking mostly about Syria, Afghanistan. We've done some work in Yemen as well Lebanon, with people who do not have the legal right to work. So we're really going into and navigating very tough environments to try to generate work opportunities for people. And I would say that this is our biggest impact, along with reinvesting our profits in additional upskilling opportunities for them so that they could progress to further job opportunities in the future.

Thanks. Yeah, it does sound like it's a complicated space, both logistically and just, you know, how to navigate companies needs and the needs of, you know, building AI systems with how people are in need of employment. I wanted to shift gears a little bit because I think for a lot of people on this call, for a lot of people watching today, they'd be thinking about how do I want to steer my own career toward the thing that I want to get involved in or the thing that I want to have an impact on. And so they might be thinking, well, do I want to work with Robert at Apple on privacy, or do I want to work with Eva and employing refugees, or do I want to work with Juan at the research lab or Andrew in the building of companies and, and products and other things. But I wanted to ask specifically about kind of how, how the idea.

I want to work with Ryan. I want to. I want to host panels like this all day. Seems like a great job.

Well, this is pretty fun, I'll give you that.

Yeah.

But I wanted to pose a question to you Juan, you run a research lab. Your team publishes really interesting stuff around applications of AI for, for example, things like assessing damage after a disaster. I've seen some things come out recently from your lab on the assessment of damage after the fires of Maui or assessment of damage after the floods in Libya. And. And then I've also heard you say publishing papers does not constitute impact. You need to actually get out there and turn this into action. That's impactful in communities. And so you work with the Red Cross and other large organizations like undp. But could you say more about what that means to you? That of course you're out there publishing papers, but that's not the end goal, that's not the end result.

Thank you. A few things. Yeah. From all the academics in the room, like publishing paper is important, really important for academics. It's really important for driving science. But in the case of these, in the for good space that we have, I do think that we need more on the applied research part. We need to make sure that we're working with organizations. There is a huge distance between solving a problem on a paper and solving a problem in production, basically. And that the distance between theory and practice is much larger in practice than in theory. And unfortunately, a lot of times people don't realize that. Like a lot of times when we're working in these problems, people say, well, that problem has been solved. Like you can see that paper, like, yeah, that problem. Yeah, maybe that's a paper that maybe people solve it in theory 10 years ago, but nobody's using. And likely they're not using it for a reason, basically. Correct. So. So I do think is really the most difficult part or these problems is the last mile where you're working with people on the ground and you're working with organizations that maybe do not have the structure or capacity to hire the talent that is needed to solve these problem, but they still need to solve these problems. And we on our side, we want to make it simple enough for this organization to use it. Another thing that is really important, I think you touched on this, is that as humans, we are addicted to complexity. We love complex things. We love complex projects. Correct. So this is the reason we put a man on the moon before we add wheels to our luggage. We want those things we want, basically. And a lot of these complexities, unfortunately, are solutions that are too complex to be working in production. So that one. That's one of the big lessons for us is that in order to have an impact, we need to make sure that our solutions are really simple. And they work. And they can work in production, working with organizations like. Another thing that I think is really important for us to see is like part of the beauty of AI or machine learning is that a lot of these problems, even though, are very different to what we're used to working on. Companies like Microsoft, from a pure AI and data science perspective, they're very similar. What is not similar is how the data gets collected and the trust on the data and understanding the data. And that is really important to make sure that you always work in these problems with subject matter experts that understand the problem, understand the data, and you partner with them collectively with all of these things. It is not easy to have a solution working in production. And that's where I think majority of our efforts should be. How can we make it easy for these organizations to have these solutions working in production?

Yeah, yeah. Making it possible for some of these things to actually be handled and be something that could be maintained by an organization that wasn't the one that developed the technology themselves. I'd like to stick with that for a moment with this idea that research is important. Discovering new possibilities and new ways of doing things is important. Putting things into production is hard. And the connection or the distance between new results and research and an actual deployment for impact in the world is potentially a big gap. I've heard you talk about this as well, Andrew, and of course you have also done original research that was the new thing, that was trying a whole different approach to something, or folks that have been working with you at Stanford or elsewhere have been on that leading edge of developing new things, new algorithms, new ways of doing things, and you've seen some of those things work their way all the way through to a production state or to something where the technology is mature enough and it becomes very important for impact in the world. Could you say more about what that gap really looks like between the kind of discovery of a new idea or the proving of a new technology on through to actually becoming something that's impactful? Like, I'll just say, go ahead, go ahead. I want to comment on something else.

I feel like one of the great things about the AI community is there's so many of us doing so many different things, and I think it's okay whatever you're doing, so long as it's responsible and for good. So if you're an engineer working to publish a paper that makes an algorithm 1% better, that's great. Go make our algorithms 1% better. Or if we can make our GPUs half a percent faster. Go do that. I will celebrate your work. I think what Juan and Ron are alluding to, which I agree with as well, is that in academic machine learning, there are certain things we do to optimize certain metrics, like the test semester. Because in academia we have to benchmark algorithms to drive for scientific progress. And we benchmark algorithms by looking at test set performance on kind of usually the same test set for different researchers. But it turns out that even though there's a lot of attention on improving tested performance on standard benchmarks, and if you do that, that's great, go publish your work, I'll celebrate your work. Even though that's fantastic contributions. There's also a huge gap between what academia has trained up everyone to do, because it's what we know how to do. We know how to benchmark governors and push forward benchmarks, which is totally worth doing. But shipping and production is very different because most of the time, for example, performance on test set is only mildly predictive of performance out in the real world. And the real world data is very different. Concept drift, data drift. And I feel like even the terms concept driven data drift is an academic description. It's like, oh well, sometimes the real world data is different than your test set data. That's a academic problem, but when you actually address it, boy, it's complex. So for a long time I know many people were thinking, surely we outperform radiologists on reading X rays, so why hasn't AI raised radiologists yet? Well, it turns out that we do outperform radiologists when we train and test on exactly the same distribution of data collected a Hyatt X ray machine, and under certain test protocols. But I think for a lot of applications, we're actually not competitive in real world data. And I think while data drift and concept drift, the idea that the real world data is different than the data you collected, that is a accurate description of the problem. When you dive into the complexity, it's so much more complex than those words imply and the solutions to them so complex in which you may need to go to hospital and figure out why is this X ray machine broken? Why are the X ray technicians not doing what I thought they would be? Often with good reason, because the patients don't like the procedure or the machines aren't working. And all of that complexity goes far beyond those simple sounding concepts of data driven and concept drift. Which is why I think it's fantastic that people focus on one slice of the problem. We do need lots of people working on the many, many different slices of the AI problem, including improving an algorithm by 1%. And I also find that to push things to production deployment for either commercial applications, for these sort of four good applications, having a broad view of all the pieces of the puzzle, having a team that collectively has that broad view is often important to then spot whatever the problems turn out to be, which are often surprising ones, and then actually solve them.

Absolutely. And so I'd just like to also pose a question for folks who are kind of thinking about research or work in industry or something else. When you look at something like the Transformer paper that came out in 2017, that was cutting edge and I remember hearing some people saying shortly thereafter, well, you know, this is just the latest thing and it'll probably be replaced soon by some of the new latest thing. And what's been built on top of that architecture now is kind of taking the world by storm. And the LLM technology that's resulted, but it did take quite a while. And so those years of gap, you never know whether a new thing is going to actually lead to impact in production later on. But. But how should people be thinking about that gap between something that's the latest, greatest thing or performing the best on the test set today and the potential for real impact in the future of a new technology?

You know, the Transformer paper is an interesting case study. So the transform paper, it was actually a paper on machine translation, not on, you know, and so it's one of those who would have known that wasn't really. I don't know if it was purely curiosity driven research, but I think the authors were really thinking about scaling. Right. A lot of the designs of the Transformers was designed to scale on GPUs and that ultimately had a huge impact. But this is why I think, I think a lot of research is just fantastic to have curiosity driven research, some of which doesn't turn out to have a huge impact, but some of which turns out to be incredibly impactful. So I think as a community, I like that we can celebrate all sorts of work because sometimes it turns out that algorithms to translate from one language to another. The Transformer turns out to be really good at not just mapping from one language to another, but at mapping from a prompt to generated text. But I want to show you just what I tend to do, which is even on doing relatively theoretical work, I try to do so with a view on at least some of the practical ultimate applications that I think I or someone on the planet would care about. Because I find that if I have some real application in the back of my mind then even when doing really theoretical work, I think the odds of that theoretical work eventually being useful somewhere I think is higher. That's just what I do, though. I don't think everyone should do what I do. And there are many different styles of research and I think that's part of the hope of the AI community too.

I fully agree with Andrew on that. And I want to make sure people understand that we wouldn't be here today if there was not a lot of investment already on the theoretical part on these algorithms, even when the objectives were not aligned at all. Basically, when people designed GPUs, they were not thinking about machine learning. Correct. They were thinking about video games. And that's a great story on how a problem that was solved, very, very great problem that was solved was used for someone else. I do think that from a guidance perspective, like Andrew mentioned, is that trying to see how the project that you're working at least tries to align to a problem, I think is a great practice. Not that everybody will do that, but I think it's a great practice. And I can say for myself, some of these problems that you are solving, if you can have a. If you can have your technology being used for a real world scenario is amazing. I think it's amazing and I think the world needs more people on that space. I think we have a lack of people working on that last mile many times because, well, from an academic perspective that problem has been solved before and maybe let's work on another problem that hasn't been solved. That's the area that I think the world needs more people working on that space. And I can tell you from all my experiences, it's not an easy problem to solve.

Yeah, I agree with that. I'd say I'm less excited about the algorithmic advances. I think we in the computer science side of things included, we get to take a lot of credit for advances in material science which have made compute a lot more scalable and a lot more affordable. And we're still using a lot of the same basic regression methods that existed long before computers even. I'll take it back to the original transformer paper and if I can paraphrase, attention to the details is all you need. It was a machine translation paper. And if you look at all the earliest papers using transformers, really good papers. The earliest papers on Burton and GPT, they'll evaluate on data sets including the Connell 2003 multilingual NER dataset named Entity Recognition. This is a data set that is English and German. German was the the language in the original paper on on Transformers, but the early papers all underperformed on German and the researchers left that out of their papers. It wasn't until a much later paper from some people who I guess were Meta at the time, the XLM paper, where they actually went back, looked at the early bird and OpenAI algorithms and found that even though they talked about doing translation between English and German, they actually underperformed the state of the art in German for it named entity recognition. So I think these adventures are great. I think these are strong and meaningful incremental improvements. I also think that caught up in the hype of trying to present an incremental improvement as a revolution, that people are being less than honest in the scientific sense in overly hyping and burying negative results, which would be very important for people in the real world to build on.

Yeah, I want to add something here because I think even if you're not developing an explicit AI for good project, it's very important to be developing responsible AI and trustworthy AI from the beginning and to preserve that honesty in your results, to be careful about what data is being used to train models and so on. Because all of these trickle down and have a lot of effects afterwards when these technologies are afterwards applied in the real world and they can have exponentially, you know, bad, let's see, effects afterwards. I'm not sure when the Stochastic Spirits paper was published, but it was also, you know, some years before the public launch of ChatGPT. And I think it did warn about a lot of problems that we saw escalate quite quickly when it was released. So I think even if you're not specifically thinking about the impact of your work on specific applications and so on, it's really important to stay within the line of responsible development of AI because once things escalate and start scaling up in the real world, it may be very, very difficult to. To revert all of those effects that they have.

Absolutely. All right, well, thank you, thank you for all of that commentary on those topics right there. What I'm going to do now is shift us to answering some questions because we've got a lot of questions lined up and we have a lot of folks giving their thumbs up to upvote the questions that they've found most interesting. So what I'd like to do is just go right from the top of the list and start asking questions and we can, I may in some cases pose them to one or the other of you. But I'll just put this first one out to the entire panel. See who wants to take this first. Varun Shankar says, I have an idea based on a real world problem. How do I get started working on it using AI? Who would like to take that? And this is upvoted by 137 people. So a lot of people want to.

Know about this so I can try to answer that question. I like what Andrew mentioned is something that is relatively important at the beginning. Thanks to large language models today we have now a lot more power. Correct. So roughly less than 1% of the world have coding skills and that makes it difficult for people to even work with code. Like of course people can go learn code like coding, but sometimes it takes time. It's difficult. But thanks to large language models now that gap is changing because we have our ability to actually write code, literally write code in our native language or at least for a significant portion of the language in the world. You can now do that and you can use large language models to guide you and to help you in some of them. I also like when people show up to my team and say, hey, we want to learn machine learning. I usually always have one answer that is send them to Andrew's original machine learning course. I think that's one of the best courses out there. I think that Coursera and other courses out there that are free and accessible for people are a great way to learn these skills. I assume that that's where it's coming from basically. And then once you have that, I would say there's a lot of problems out there. There's a lot of data sets either from Kaggle or from places where you can practice. But I would say it's not impossible to learn coding and machine learning and it's certainly worth it. So I would love other people to make opinions on these, but that would be my guidance.

Yeah, interesting. And so I guess we might have folks in the audience who are machine learning practitioners already. We might have some folks who are interested in getting involved in one way or another, but don't necessarily have that background. But yeah, maybe. Eva, I would love to hear from you on your thoughts on that.

Yeah, just because as an entrepreneur I also started trying to solve a real world problem with technology. My suggestion, and that's rule number one in startups in general, my suggestion is do not fall in love with the solution, but rather with the problem. So do not fall in love with whatever AI solution you think would be most appropriate for this problem. Try to investigate the problem in depth and to see who is working on this problem already, how are they tackling it? What technologies are they using? Is AI even the best solution for it? Maybe it is, but definitely try to cooperate with people who are already who have been dealing with this problem for a long, long time before you even came to the scene.

Yeah, I agree with that. I think I don't know what reward problem is being talked about in this question, but to build on what Eva said, I think the first thing to ask is can AI be part of the solution? This is something in our course that we come back to repeatedly. I say this. I don't want to be negative. I love AI. I did a PhD in machine learning. I was still excited by the end of it. I'm really proud of all the products that I've built in industry, but when I've worked in public health or a disaster responder, most of the time when I've tried to build technology solutions, AI hasn't been part of that solution, it's increasingly been part of it. So I think to double down what Ivra said, start by understanding exactly what problem you were trying to solve and then think about the different technology or even non technological solutions that you might be able to bring to bear to it.

I'm trying to give some agree with everything that's been said and just let me try to give some quick tips in response to Varun Sankar's question Wasn't sure where you are Varun in terms of the idea you have on a real world problem, but somewhat generic advice. I think if you have an idea I often start off with a quick literature search hopes, you know, gain situational awareness about what else might have been tried, whether it worked or not. And then after a quick literature search if you have the sufficient technical skill to do so, I would often start by looking for open source and see if I can find a solution tonight. Sorry, not sure if my wi fi is bad and then if and then if it's responsible to do so, if it can do so keeping within the boundaries of responsible AI often and if it's feasible, start building a prototype and see what works and iterate from there along the way. If you can find friends and allies and advisors or co collaborators, that helps a lot too. I know the advice I'm giving is somewhat generic, but depending on the details I found this to be literature search open source, build quick prototypes, but only a responsible way. Find friends, allies, partners to push it forward and then I find that often if you have an idea. You understand your idea deeper than anyone else in the world. And the best way to get momentum on it is to do it yourself, at least the first few steps of it. Because often that first prototype, if it's your vision, you understand it much better than so many other people that if you implement a prototype, that's actually a good way to communicate it to others and then get buy in and help the whole project gain momentum.

All right, thanks all for comments on that. We'll move on to the second question that was most upvoted. How do you view the ethical considerations and potential benefits and risks of training AI models on data from the dark web? And so I'm not actually sure what exact data that would be referring to, but does anyone have a comment on that that was upvoted by a lot of people so very curious about impacts of.

My goodness, don't do it. There is enough data which is not on the so called dark web, which is already a huge risk for people's privacy. I mean not just copyright infringement. The, the majority of even published research on things like Twitter or X or whatever it's called, the majority of that is funded by totalitarian regimes who are trying to monitor and identify dissidents. I saw someone in the comments mentioned they're, they're in Belarus and that would be one really good example example. So the data which is not even, not even accepted in academic circles and is on the dark web, there is no way to ensure that you're protecting people's privacy and safety from it. So yeah, I'll let other people on the panel disagree with me, but my, my advice would be to stay away.

Totally.

I saw the question today and it was like trying to find which scenario would be possible that like would be actually positive or would make sense. I agree with, with, I agree entirely with Ron. Like, majority, big majority of the scenarios would be a no, no, like from a practice perspective.