Loading summary
Transcript86 lines
- [00:03]
Chris Hoffman
Welcome to Humanitarian Frontiers in AI, the podcast series where innovation meets impact. In each episode, we dive deep into how artificial intelligence is reshaping the future of humanitarian work. From enhancing crisis response to making a delivery smarter and more effective, AI is opening new doors in the way we support communities in need. In this series, hosts Chris Hoffman and Nassim Motelaby bring you thought. Leaders from academia and the tech industry discuss not only the vast opportunities AI offers, but also the ethical considerations and risks we all must navigate. Join them on this journey as they explore AI's potential to transform lives and address humanity's most pressing challenges.
- [00:48]
Host (possibly Chris Hoffman or a guest host)
Hi everyone, and welcome back to a new episode of Humanitarian Frontiers in AI. And I cannot be more excited. For the panel that we have on here today, we've got David Master and Scott Turnbull. I'll talk a little bit about their experiences, et cetera, but this is going to be a fun one. We're going to really dive in today more deeply into demystifying LLMs, trying to understand what it actually means, what it's going to take to actually use them, and what goes on in the background. Right, Because I think what we see is just the kind of the prompt center where we type in something and we get these beautiful answers, or we go into a WhatsApp chat and we type in something and we get this great conversation. And the reality is there's a lot that goes on in the background and there's a lot of considerations that NGOs are going to have to have as they move forward, whether they host them themselves, whether they use an API, call to another organization's AI LLM, and really trying to understand then what does machine learning mean and what is the data that we need to have? And what if we don't have the data? What should we do? What shouldn't we do? We're really going to try and dive into a lot of these questions today. And so as I said, I've got Scott Turnbull here, the founder of techtavern and was the CTO at Data Friendly Space, a really important nonprofit that was breaking ground across the sector with utilizing artificial intelligence. And David Master here, an AI specialist at aws, Amazon Web Services, has been working a lot with nonprofits, is really deep into the agentic stuff. So hopefully we're going to have a great conversation today also about agents and what those do and what they mean and potentially what's the difference between an assistant and an agent and all those things, and start to really, again, demystify some of these things. So this podcast today is Brought to you by the Innovation Norway, who's funding this podcast. We're really thankful to them for being able to provide us these funds to bring these great people together. So, Scott and David, really happy to have you both here. It's going to be exciting.
- [02:39]
David Master
Yeah, thanks for having us.
- [02:41]
Scott Turnbull
Thanks to you. Yeah, thank you.
- [02:43]
Host (possibly Chris Hoffman or a guest host)
Absolutely. Nassim is traveling, unfortunately, my illustrious co host and she might be able to pop in a little bit later in the podcast. She's in an area with kind of limited connectivity, so we'll give her a pass on that. But first question, David, to you is can you just walk us through as a five year old, what is an LLM?
- [03:03]
David Master
Basically, an LLM is just a model that understands language patterns. So it's trained on enormous volumes of data and it allows it to understand how words and phrases interact and work together. So essentially what it does is it just predicts the next word in a sequence. And so common misconceptions about large language models is that they know things and it actually doesn't know anything, it just understands those patterns in language. So one of the issues you run into is if you're not careful with choosing which models and knowing how to interact with those models is if you ask too specific a question to too large a model, you get back coherent nonsense. So it's what's called hallucinations and it's applying those principles and rules and patterns that it understands from analyzing large bodies of text and it just generates a coherent response. But part of what using it responsibly is understanding how they work and knowing both what to put in to these models to generate useful outputs and how to interpret and be critical of those outputs and continue to work with the model to get something that's usable. But essentially what it is, it just understands patterns and languages. And then what really becomes valuable is when you apply that understanding of language and those models to authoritative knowledge basis and to material and information that you can trust and pull out actionable insights that can inform decision making.
- [04:26]
Host (possibly Chris Hoffman or a guest host)
Absolutely perfect segue to Scott, because I wanted to ask about machine learning. I know you guys at Data Friendly Space, we're looking at a lot of that and figuring out how to train good data sets into actionable information. Can you tell us a little bit about what machine learning is again for the 5 year old out there?
- [04:41]
Scott Turnbull
Yeah. So machine learning is basically just setting up a list of tasks to teach computers how to accomplish an end goal. So you do things like adjust a reward value for the system and you let it begin to work through Problems and often discover on its own, this is a better technique than that. And it's the infinite monkeys on infinite typewriters problem. But we start selecting the monkeys that are doing the best work. And so through that reward function, it builds up its own understanding of how to achieve pathways to solving a problem, and over time builds the most efficient way to do that.
- [05:15]
Host (possibly Chris Hoffman or a guest host)
Oh, that's great. That's super, super helpful. And, you know, it really is that thing of trying to understand. And so let's say that we want to look at a large data set, a humanitarian data set, and we're looking at predicting a crisis, okay. Or a type of crisis. For example, let's call it civil unrest. So we look at civil unrest, we have a lot of data from years, it's very unstructured, it's all in different languages, it's all in different types of formats, et cetera. But organizations really talk about this all the time. We want to be able to predict crisis. Predict crisis, right. So what's step one for them when they're looking at this as an organization? I mean, the problem is that they don't know when conflict is going to happen. Right. So we know we're solving for when could a potential conflict happen and then work from there. So Scott, maybe over to you first on this one. But where do they start? Where does an NGO start? If they say, look, we got a million dollar grant to do predicting crisis around the world, what do we do next?
- [06:13]
Scott Turnbull
I mean, I think really you want to engage a few machine learning academics would be a good way to go, but the process is actually fairly simple. What you want to do is take data sets from previous similar events, establish that as your training set, build from that a mathematical, the computer machine learning. You might have heard TensorFlow. You'll build a mathematical model from that previous set to start to recognize what kinds of patterns it might be in language or movement of people or whatever data you have, and then apply that to the new set to give you a percentage likelihood of how much did this new data set match? Some of the significant patterns from your old data set, and that gives you a percentage of this is very, very similar. The language you're using is very, very similar. The migration patterns are very, very similar. These are the political situation, whatever you're basing it on, it's really just a way to develop a similarity between those two data sets. There's actually even the branch of math is called Bayesian mathematics. So it's just around the similarity of those two patterns in the data and.
- [07:13]
Host (possibly Chris Hoffman or a guest host)
Then from an infrastructure point of view, David, So we've got the academics, we start to build out this mathematical model. Now what do we need to make that mathematical model function?
- [07:24]
David Master
Sure, yeah. A place to put it, a place to store the data. And then maybe most importantly is just good data practices. A lot of what we talk about with customers is building a data culture and how you're thinking about data, how you're training staff and personnel on using data and understanding data stewardship. And a lot of what we help customers with is building a data strategy. When they start talking about AI, we help set the vision and talk about what's possible, what they want to do, and then say, okay, let's talk about your data. What do you have to really inform this project and what you're looking to get out of it? And so we have these pillars of a data strategy. So mindset people, process and technology. So you don't even get into the technology until after you talk about mindset, how you're thinking about data. Are there organizational standards? Is there someone who's thinking about as an entire organization, how do we treat and handle data people? Do you have the right roles staffed? Is that personnel equipped with the right training and knowledge for how best to select a model and utilize that model? And given the tools the model set up in the right environment to engage with it processes, how well can the organization understand their data needs and then build data solutions and respond to those data needs? And then the last part is technology. You know, once you have that culture in place and those practices and those feedback loops, it just becomes a matter of picking whatever technology best services those needs. And so, you know, a place to store the data tools to access models and compare models and be able to pick and choose which models are best for a specific use case is really kind of the needs organizations have in order to meaningfully apply AI to whatever work it is that they're trying to do.
- [09:06]
Scott Turnbull
Absolutely.
- [09:06]
Host (possibly Chris Hoffman or a guest host)
That's really helpful. And then the next piece is then what's even a budget? Like, where do you even start with thinking about? So using my same use case. Right. So are we talking 10 million? Are we talking a million? Are we talking 500,000? You know, what kind of human resources does it take and then what kind of financial resources does it take?
- [09:26]
David Master
Sure, yeah, yeah. So with aws, at least it's pay for what you use. So there costs around data storage, so how much data you're storing, and then the compute as well. So if you're doing anything around training or using the models, there are compute costs associated with that. And then leveraging third party models, that's on a token basis. So the text is broken up into tokens. It's words or phrases or little chunks. However, the model chunks the language that's being processed and then you're charged per token. So it's really a matter of how you're using it and there are ways to use it more efficiently, but those are the kind of parameters that dictate cost and is data storage costs and then leveraging the model, you know, the compute costs associated with actually querying the model.
- [10:10]
Host (possibly Chris Hoffman or a guest host)
And Scott, you know, being in a nonprofit at Data Friendly Space, how open were donors to wanting to support those types of initiatives?
- [10:20]
Scott Turnbull
Well, AI traditional data initiatives, I think they're very open to. That's something they sort of understand and they recognize the need for over a decade now AI is emerging. So they're not quite sure. I think it sort of depends on the donor. I think that their interest is great interest. Sometimes we can get lost in the safety conversation, which is extremely important. But we need to do more than just say safety. We need to actually have an idea what that means. But I think there's a lot that nonprofits can actually do to help shape that conversation first. Sometimes it can feel like opening a vein when it comes to a budget. Right. So we need to sort of a CTO at a nonprofit really needs to shape for that potential donor. How do you maintain? How do you control costs? How are you keeping this achievable? How will you know you're successful? And really the short of it, by the way, shout out to David. He's actually a partner of us at Data Friendly Space. He helped us make our humanitarian decision AI Gannet. So he was incredibly helpful and his entire team. So big shout out to Octavia. I'm a subscriber and Casey, yeah, it's really. We love it, but you can actually get started for surprisingly small footprint. The critical piece is to not reinvent the wheel and to really, especially for humanitarians, we are are focused on a specific domain case and we need to focus on our specific value add prospect. Right. Not get lost in just recreating a spreadsheet or recreating tableau. We really need to focus on that value add piece. And then when you can do that, when you can hone how you can achieve it, that you can show you can use minimal resources to do it, that you have a strong business partner like aws, then donors are a little bit more friendly to releasing money because you have to give them the art of the possible and they have to, in their mind, see both how this can be achieved and how that aligns with their giving priorities.
- [12:03]
Host (possibly Chris Hoffman or a guest host)
Yeah, absolutely. You're so right. I mean, I guess switching the use case a little bit. And I don't know if we have to go down the agent or the. I think there's more of the assistant kind of conversation, but a lot of conversation is around beneficiary facing engagement. Right. And so now we're talking about not just processing data, but being able to have a conversation with somebody and then be able to give them information. So have that kind of flow of engagement. And there's a lot of reticence. You know, in the last different conversations that we've had on the podcast, and even David, when we were in a conversation with some NGO CTOs at one point, you know, they, they were saying, we are not going to let this AI get anywhere close to the people we're serving. And is that reticence, is it founded or actually are there really strong use cases that can be developed and be trained and be accurate to be able to engage with people in need to answer the questions that they have and be able to not necessarily provide assistance, but provide that knowledge assistance that they might require, for example, legal information or where to find the nearest doctor's office for X sickness and those types of things?
- [13:14]
David Master
I'd say yes. Yeah, there are definitely real concerns and it's important to be thoughtful about that, but not necessarily as a deterrent, just to be mindful of. So, like EU has a framework for. I can't remember the terminology they use specifically, but around responsible use of AI, and they have different levels of risk. And so going up to. I can't remember the exact. But it's basically untolerable risk, I think, is what it's called. Yeah, we'll go with that. Unless someone has something better.
- [13:43]
Host (possibly Chris Hoffman or a guest host)
I think it's close enough.
- [13:44]
David Master
Yeah, yeah. So it goes to high risk to intolerable and intolerable or something that you just don't want AI involved with at all. And then high risk is something like impact someone's livelihood. So we had a customer that was looking at doing intelligent document processing on applications and transcripts. So that's the type of thing that a decision could impact someone's livelihood in the course of someone's life. So that's a high risk area. It doesn't mean don't use it, it just means understand the implications and how it's going to be used. And how can you understand bias in the data, bias in the model, and mitigate for those concerns. So at Amazon, we have a responsible AI framework as well that covers these different topics around fairness and transparency. There are eight different pillars, and the first step to responsible AI is assess the risk and understand what are the possible implications of applying this tool and this technology and then what are risk mitigation techniques that you can apply throughout the process. So just because something's high risk, risk I don't think should be a deterrent. It's just something to be mindful of.
- [14:47]
Host (possibly Chris Hoffman or a guest host)
And Scott, I mean, when we think about this, let's use an example of a chatbot that is going to give people information about Ebola. There's an Ebola outbreak right now in Uganda, and you want people to access information quickly and accurately. And so you whip up a chatbot with an LLM, you train it, you know, you give it all the training documentation on the WHO guidelines on Ebola and what to do is that that's possible?
- [15:13]
Scott Turnbull
It's absolutely, yeah.
- [15:14]
Host (possibly Chris Hoffman or a guest host)
Not, not that expensive to do, impactful. And you can pretty much mitigate some of this risk, right?
- [15:20]
Scott Turnbull
Yeah, we've made huge advance. We, the entire AI industry has made huge advancements in reducing hallucinations and other risks of bad information coming out there. If you've been following AI, you've probably heard something called rag, which is retrieval, augmented generation, which is sort of a graph database that relates similar topics. There's a new technology called, I don't know how to pronounce it. R, small E, a G. And instead of taking similar lines in a piece of text, it's actually taking similar concepts. So it's even taking that an order of magnitude further in terms of more accurate responses. Actually, I would love to expand on that answer. Risk. I absolutely agree with everything David said. However, full disclosure, I'm a techno optimist, but I would say any risk model is incomplete without the benefit analysis as well. And we have to understand there's also a risk to be had in not acting, especially in humanitarian context. If you could have saved 10% more people or served 10% more people, what's the impact on that event if we fail to act out of fear? That's one of the things we have to do. It's very founded to be cautious, but that caution has to inform our mitigation strategies and our deployment strategies. It's not sufficient to simply not do something because you're afraid. We have to take our professional pathway forward. We have to make sure that we're doing our due diligence and then use the things that are promising to deliver higher impact to our charges. And I think that piece, again, I agree it's founded. I'm just saying that we have to take into the equation the impact of failing to act as well.
- [16:57]
David Master
I'm with you that it should be used and applied and experimented with, but just understanding the potential impacts. So I don't think that should slow the pace of innovation and experimentation. It should just be done thoughtfully. But I'm a techno optimist. Right there with you, Scott.
- [17:13]
Host (possibly Chris Hoffman or a guest host)
I think we all are here. I hope that was the whole gist of this podcast when I was pitching it to Innovation Norway. I said we're trying to take a positive view, but a pragmatic view at the utilization of AI in humanitarian settings. I mean, that's the real truth.
- [17:25]
David Master
That's the way to do it.
- [17:26]
Host (possibly Chris Hoffman or a guest host)
Yeah. And the thing is, I think what I keep hearing and I keep seeing is a big call for efficiencies. Obviously, we know what's going on in the United States. All of us have been impacted in different ways through that, terribly and unfortunately. But, but in that idea, this idea of creating efficiency, people tend to err on the side of technology for creating efficiencies. Right. They think that the technology will solve their efficiency issues. We know that that's not true and we know that it takes a lot more than that. But when it comes to then creating these efficiencies, Look, I'm a 20 year humanitarian in the field. I'm not a technologist, I'm an armchair technologist, as my friend Andre likes to say. But the reality is that they're going to need to have the right people to help them understand how to solve these efficiency problems. And something that I ask in every podcast with Naseem is what's the HR of the future for nonprofits? What does it need to look like? Who do they need to have there? We talk about, is it a data scientist versus a data engineer versus all those different things that they don't have today or have one out of an organization of 10,000? So what are some of these key components that are going to help maybe accelerate these potential efficiency gains by utilizing AI through a staffing lens? What does that look like?
- [18:45]
David Master
I think the role of hr, large part of it is going to be data curation and compiling. What are the company policies? Training, I think, is another function that ought to fall under hr, or at least enable. But having all this information resources available to the staff in a way that they, you know, it's somewhat self service, you know, they can ask and query the organizational knowledge themselves. So, you know, internal AI tools I think will be a big piece of HR and certainly training. I think one of the most exciting applications of AI is around education. And I think a big problem we have with education is the complete comprehension of a topic. And one of the things that these tools can do is analyze a gap in comprehension and it can generate questions in increasing degrees of sophistication to help a person really master concept. And it gives you an opportunity to build a solid foundation on these principles as you're learning new topics. And particularly with math, you know, each new topic builds on prior principles. So having a solid foundation and understanding and comprehension of a topic really allows a greater mastery of it. So I think that will be a major role that HR of the future plays is what is that organizational knowledge and how is it best delivered to the people who need to apply it to the mission of the organization?
- [20:14]
Scott Turnbull
I would agree with that. I think traditional job titles are going to stay in place for a while, but really we're just going to become much more broad. Everybody needs to become a domain expert in several domains now. You can't just sit in your data table and know how to do SQL queries. You have to have, have a very broad understanding. You know, intelligence is going to be free, but wisdom is going to be at a premium. So that ability to sort of really focus on that domain expertise, understand what your colleagues are doing. Why is it important if you're in humanitarian field right now as a data specialist and you don't understand the sphere standards, you need to go.
- [20:50]
Host (possibly Chris Hoffman or a guest host)
Yeah, exactly, I hear you. Absolutely. Yep. I love that analogy. That's great. Because that was actually the first predictive model that I was trying to build was in Haiti. So this is at the Haiti earthquake and we were collecting data daily from each of the 1500 settlements around the city. And then we were putting that in an Excel sheet and then running different queries and pivot tables on top of it to see where they were on sphere standards for each camp. And that's now what's become the displacement tracking matrix with iom. But that was the second implementation that we'd ever done with that. So I love that analogy. And David, to your point earlier on, the EU rules, you know what the first rule is that just came into effect for the EU responsibility rules is that if any of your staff are using AI, it is your responsibility as an employer to train them.
- [21:35]
David Master
Oh yeah, absolutely.
- [21:35]
Host (possibly Chris Hoffman or a guest host)
So that's the baseline. That's the baseline where the first rule that came into effect is if you do anything with AI in your organization, your staff must know what it does and what it means and how it works and all of that stuff. So I thought that that was a really good thing that you brought up. I really appreciate that. So as we start to look at more practical applications of what we have done in the past, like I said, Excel, collecting data from SMS and putting it into an Excel, putting a pivot table and KML files and KMZ files before they were API, the first kind of use of APIs and all that. Where are the quick wins with AI in the humanitarian sector? And we ask this to everyone in each one. But I think from both of your perspectives, because you're deep into the tech, but you also have the touch points with everything going on. David, I'm going to ask you to maybe give an example of what was going on in the Amazon. I love that it was a story of the Amazon by people at Amazon. I think it was the Amazon that we were talking about that. But what are some examples, Scott, that you're seeing because you're working in this space with data?
- [22:35]
Scott Turnbull
I think data tagging. I think that we spend a lot of time tagging data. Tagging qualitative data. Dealing with lots of qualitative data is a hallmark of humanitarian work. But there is quantitative data in there. There is valid, valuable referential data, There is valuable logistics data. And AI can be used, in my opinion, AI can be used to extract that automatically generate reports. Consistent reporting. Maybe I'm prejudiced because of my work with OCHA on this, but consistent reporting across multiple organizations in the field, I mean, it's pure chaos, right? Everyone, you know, people who are not in the humanitarian sphere think that a crisis happens and we, we load hundreds of humanitarian relief workers on a plane and send them over. That's not how that works, right? No, we have all these like in country NGOs that are, or, you know, it's just thousands of these little companies. Right. Or community organizations and coordinating that is really alliance effort. Right. And leveraging AI to make sure that coordination is smooth. We're filling in data gaps, we're tagging, reporting, raising it faster, you know, especially if there's a conflict in the area where literally minutes can matter. To me, that's a great use for it.
- [23:40]
Host (possibly Chris Hoffman or a guest host)
Yeah. No, I love it. David, Sorry, I'm gonna, I'm ping you on that. But the example that we talked about a few weeks ago, I really loved on Other applications?
- [23:49]
David Master
Sure. The rainforest one, correct? Yeah, yeah, yeah. So we've got an interesting customer that is listening to the rainforest. And their motivation was to monitor for illegal logging and to listen for chainsaws and trucks in parts of the rainforest where they shouldn't be. But one of the problems was there's this whole soundscape over the rainforest of all these birds and different animals. So what they started doing was identifying the different species noise and then removing that audio. And once they were able to remove the audio, they could hear the chainsaws in the trucks. But it gave them all this data on these different types of birds and these different species of birds. So it created this really incredible data set on what's going on in the rainforest in terms of wildlife and an opportunity to understand different types of calls and how healthy these different populations of species are. And we actually have another customer that's doing something similar with the oceans, and they're studying the impact of water quality on marine life. And so what they had been doing was going out and tagging fish and sharks and different type of marine species in hopes that they would swim by one of their tags. And it gives them a little insight into what that particular fish is doing and how it's behaving in response to water quality. And what they want to start doing is listening acoustically and being able to identify unique fish noises so that they can monitor the health of a reef or a particular area in correlation with water quality acoustically. And rather than going out and tagging it, you get a much clearer picture of what's going on. And far less expensive. And they're starting small. So the fish that they're going after now is a Goliath grouper, which is a good indication of health of the reef, but it has a really distinct boom. So it gives them an opportunity of something to listen for and to create a model that can hear this boom and know, oh, that's a Goliath grouper. And as they do that with more and more fish, they can understand what's the health of that reef and correlate that with water quality and understand what is the impact of runoff water and other water quality issues on marine life. It's an interesting example of as you have something that you're going after, it creates all this other data that you can use and lots of other use cases and studies that can be done with the data that's created and discovered through some of these endeavors.
- [26:11]
Host (possibly Chris Hoffman or a guest host)
And Scott Pre show we were having a small chat about the prediction side of things. Right. Because we're always trying to. What do they say? For every dollar you spend on DRR, on disaster risk reduction, it saves you four in a response. Right. And so this idea of being able to figure this out, what's happening on the ground before the event, whatever that event is, happens. So why are we not? Or are we. And we just don't talk about it or why is it not enough? Where is this trend? Why is this trend not taking root now and accelerating?
- [26:45]
Scott Turnbull
I think people are locked in traditional modeling, data modeling for predictive analytics. So there's a little reluctance to move forward with AI based analytics. So it's part of our culture, is that. But there's other things that are outside the culture, humanitarian sphere. When you make that predictive claim, your credibility is on the line. I don't know. It takes five good experiences to counteract one bad experience. I mean, 80% is a pretty good predictive percentage if you ask me. But if the way the human mind works is one bad prediction counteracts five good predictions, we might worry about our credibility in that case. And I actually think that we are focused. So this is a missed opportunity in the humanitarian sphere. To me, we are focused on predictive analytics for humanitarian responses, which obviously are incredibly important. But we have an opportunity here, as we just heard with this data example that David just gave. We have an opportunity to express to the rest of the world how interconnected these issues are, how important it is for every aspect of life. Yes, the people we want to serve in humanitarian responses are primarily important, but also the communities that touch them, the countries around them for their stability or their resilience, business opportunities, these are all broad, often global impact events. And we are so focused on our humanitarian response, again, which is supremely important. But we are missing an opportunity to help our fellow citizens of the globe understand that this impacts all of us. That these dollars are well placed in helping to mitigate these events. That a collaborative response is often better than just handing it off to the Red Cross or some event like that. That this cross sector communication and this cross sector community building is vitally important for keeping funding to keeping communities of practice together. So the predictive analytics, I think we're a little too inwardly focused. I think we're a little too prickly about our individual models. And maybe we should be focused on outcome more than process. And I think we have a missed opportunity to communicate that impact to our global community. I don't know if I answered your question.
- [28:58]
Host (possibly Chris Hoffman or a guest host)
You did it. It's a very existential one. Right. I mean, because the reality is, the first thing that comes to mind, because I agree with everything that you're saying, is who. Who's going to coordinate that? There's got to be some sort of command and control. There's got to be some sort of baseline agreement that is facilitated upon all of these groups that you've just mentioned to work together and to do this. We would call it humanity in general. Right. That's who's responsible for this. But they need somebody or something. Right. That sits at that top. And the UN sometimes would say that could be ocha, but OCHA doesn't do pre. It does post, you know, and then. Or it could be, you know, the Red Cross, because they're in every country in the world. Maybe they are the ones that should hold standard bearer for this type of thing. But it's a very difficult thing. The WHO piece, you know, when I did a lot of civil military coordination when I was in the UN and I loved working with the military because it was quite clear who was making a decision and that that decision was going to be followed through versus, you know, we get killed by collaboration, killed by convening, killed by meetings in the humanitarian sector where you're not really sure if the decision was made, and then you don't know who's actually responsible for taking the decision forward. And so it's a tough group. And so I want to go to you, David, now that you've worked with nonprofits, you've started to engage with people that are out there. What are the misconceptions? What are the kind of questions that people ask you that you're like, what, you didn't understand that?
- [30:24]
David Master
Or.
- [30:24]
Host (possibly Chris Hoffman or a guest host)
No, it does not do that or whatever those things are. But are there some misconceptions that you constantly hear that we can dispel right here in this podcast for good?
- [30:34]
David Master
Don't get asked the question again, less so from customers. I think a lot of customers we're working with, Scott, for example, super sophisticated and have a clear vision of how it can be used and how to use it. But I think a common misconception in society in general is that it's all powerful, is that it can do everything and replacing jobs. I think there's no shortage of need. It's not like if you could process the data faster, you'd be done and can go home as a tool for efficiency. It can do that. But that just allows these highly skilled professionals to focus on other areas and to expand their impact. So I think that's a Common misconception that I think is failing to acknowledge that there's so much more need than there is capacity to address that need. And these are going to be tools to help extend the reach of people and help people get information faster to make better decisions. And one use case that came up yesterday as I was talking about this podcast, was disaster surveillance and being able to send a drone out to impacted areas and gather all kinds of incredible data. You can run models on the drone itself and analyze the damage and impact and then use that to inform how you respond to the disaster and where the critical areas that need attention most and can have the largest impact. So as an opportunity to inform how you deploy resources and do that in a more efficient and informed manner, I think is the real potential that this technology has. And you mentioned Chris the other week, geospatial data and how that can be used for predicting disasters.
- [32:17]
Scott Turnbull
Can I rip on that? Yeah, I absolutely agree with everything David's saying. I think there's really two ways it does it. It does it in terms of efficiency to be able to move through capacity quicker. And then there's complexity, which is my favorite part of AI, but I'll get that in a second. But just, just efficiency. One of the things I was talking about in December, landmine removal in Ukraine, you just throw in, you combine drone imagery with AI analysis of that imagery and maybe mine removal drones. How much faster can you clear real estate of these dangerous drones and munitions with that capacity than we could have 30 or 40 years ago? I mean, I think that's the kind of efficiency gain we can get. The really thing for me though is complexity, right? The thing I think people miss about AI, well, LLMs anyway, is that this first time in human history we can scale complexity, right? That humanitarian responses are always focused on bulk delivery of services to vulnerable populations. That's still going to be supremely important. But a great example of this again project I was talking about last month, and the world has changed in a month. But this idea of feeding migrant populations. So if you have a camp of people who are, are suffering displacement, we're always focused on feeding them, right? Or maybe we're sanitation, but let's use feeding as the example. Just getting bulk food into their mouths. But once you can solve that problem with AI and like logistics and get there, you can actually begin to do the complexity of it. How many people have specific dietary needs, how many of them medical needs based on the diet that they need, can you get this individualized service that really no human being could Personalization.
- [33:54]
Host (possibly Chris Hoffman or a guest host)
If we can personalize these services and then still cut the cost. Right. Because the current issue is that they believe that the more complexity, the more expensive it is. And actually, we know that the long tail of this journey is that it's going to actually bring in so many more efficiencies because you don't just have to buy 50,000 metric tons of rice, you can give everybody 50 bucks and then give them 10,000 metric tons of rice or whatever it is. Right? Yeah. Actually, it's so beautiful, Scott, how we're mind melding today in this call, because when David was talking, the first thing I wrote down was complexity. And then you're like, I want to talk about complexity. Yes, yes, that's what I want to talk about too. No, I love it. And so in a previous podcast, Nicholas Thompson of the Atlantic was on and he used a mind clearing example as well. And the conversation started to go around a bit around the idea of the cost of error.
- [34:47]
David Master
Right.
- [34:47]
Host (possibly Chris Hoffman or a guest host)
You talked about the cost of not doing, but this idea of the cost of error. Right. So when you talk about mine clearing, and as you say, if you could clear it 10 times faster than you do today, then you can send out the crews to go and ground truth and make sure that everything's done. But the mindset is still back to what David said, which is everybody thinks that it's just going to do it and be done. No, this is an enabler. Right. It's an enablement. Military calls it force multiplier. And I really hope that by the end of these 10 episodes that if people walk away with one thing, it's to understand that AI is an enabler. It's not replacing David, as you said, not replacing your jobs. It's enabling you to do your job better and more efficiently, potentially.
- [35:24]
David Master
Yeah. I think there's a lot of stuff we do now, just mundane, repetitive tasks that we're going to look back and be like, I can't believe people were doing that. That's a computer's job. So, yeah, I do think it's going to free people up for higher level thinking and being able to look around more and think, where can I have a large impact? And then. And then access the data to inform those decisions. So it's going to be informed by data and then the delivery of the solution will just continue to utilize and use data.
- [35:54]
Scott Turnbull
Yeah. And actually if I can refine, I think utilize the solution is the big deal. Can we get something spun up quickly and go. And actually this is a shout out to AWS and their partnership, like it's never been easier. And we've only been doing this for a few years, right? Never been easier to stand up a quick, quick AI model you can get. Like at dfs, we use for again Gannett's. We love that service. AWS was a great partner for us, but we basically went to the route of doing everything by code because we did it very early. And now AWS all these services where we can put an S3 bucket together, we can drop in our data in that bucket, we can build a rag and a knowledge base from that using the other bedrock services. So literally in a week or two, you can have your AI service again, though it doesn't know it for you, right? It gets all the infrastructure together, it lets you deploy incredibly quickly. But you need to know what are your most valuable sources of data. You need to know your domain problem. And this ability to hyper focus on your domain problem is absolutely the critical skill going forward. So I like this ability to scale and I think the ability to sort of get up. You can have an infrastructure stood up by the time you get your first responders, your strike team on a plane and into the locale, you can have your infrastructure spun up by the time than with cloud services, which I think is a great confluence of events.
- [37:12]
David Master
One of the ways we like to talk about it is we sell Lego blocks. So we sell these little pieces that you need to build these solutions and then builders like Scott can put into a castle or a helicopter or all kinds of cool stuff that he's out there building. But they connect together well. But how you put them together and architect it is still a large piece of it, but we try and make that as easy for you as possible.
- [37:35]
Host (possibly Chris Hoffman or a guest host)
That's super awesome something you said, Scott, on this hyper focus on the domain, right? I think that might not be a terminology that everybody that's listening might completely fully grasp. Can you take a small dive into what you mean by hyper focus on the domain?
- [37:52]
Scott Turnbull
I mean, you really need to understand why they do what they do, why it's important and what their baseline methodologies are. I mean, you need to become a professional. So in humanitarian sphere, you need to become a humanitarian analyst. I mean, in some light way. I mean that's a profession, so you know, it specialists isn't going to do that. But you just can't sit back on your knowledge of databases. You have to deeply understand the interconnectivity of that data, especially for field workers. I remember this became really apparent to me when I first got into emergency response is that you need to understand the context of a relief worker on the ground, what their confusing environment is like, what data they can and cannot pay attention to and how that physically works out for them within the environment. So you need to immerse yourself in that, that profession as deeply as possible. We're overly focused on dashboards. Everybody loves a dashboard. Loves a dashboard, right?
- [38:45]
Host (possibly Chris Hoffman or a guest host)
But an infographic and a dashboard, you're good, right?
- [38:47]
Scott Turnbull
Yeah. But you're a humanitarian in the field and you're just trying to get the most number of vaccine inoculations out in a day. How much time do you have to look on their little widgets and blinky dropdown menus and blah, blah, blah. Sometimes you need to just put that, yes, you've learned a lot, but you need to put that all aside and think from the perspective of the field worker or the planner or the administration specialist and that actually we're focused on field operations. But I think that we need to appreciate we have a role in funding. We have to make our technology and our outcomes relevant to our funders to align with their giving priorities so they see the impact of their investment and their dollar and they go, you know what? I feel great. This is in alignment exactly with what I wanted to do. I'm happy to increase my funding next year by another million dollars because of the high, high impact of this service that I saw in the world. Businesses want that impact. They give lots of money. And we can't rely on governments anymore. So we have to find our partnerships where we can. But we're critical in making that case on why this is important. We can't just tell people, well, you should value this and you're a bad person if you don't. We have to show them the impact. We have to be undeniably good. And I think that humanitarian work, it's more important than ever.
- [40:03]
Host (possibly Chris Hoffman or a guest host)
I mean, David, to that point, then it just came to my mind, how amazing would it be for a corporation that's in this industry to help to develop those domain blocks. So we have the water, sanitation and hygiene domain that we've trained everything on so that it's clear and we understand it that people can then use so that you can apply that into your model or a model for food distribution, a model for gender based violence or whatever it is. So look at those programmatic buckets of humanitarian organizations and what they're dealing with and start to build out that domain knowledge. That would be really cool. I mean, maybe that's something ocha if they're listening, Andre, who's going to be on the next podcast, maybe that's what he can start doing and start to bring in some of that information. But the reality is that I think you're right. I know you mentioned Sphere earlier and the Sphere guidelines standards, sometimes they go back and forth on what they're allowed to be called and what they're supposed to be called, but the sphere indicators, let's call them those and things like that. But there's a lot out there I'm starting to pine for a humanitarian AI that just takes all of. Take everything from Ocha, take everything from the usaid, take everything from Echo, bring it all together and just start to train this model and being able to find out what works and what doesn't work in certain situations in certain countries. I think it's pretty awesome. The big question that I got yesterday, I was talking to a customer and they said, well, what's the sustainability of this? We want to do this. We'd love to have this. We're thinking about potentially having to hand it over to the government in X country once we're done. How do we create sustainable AI? And I don't mean that from the environmental side.
- [41:39]
Scott Turnbull
Right.
- [41:40]
Host (possibly Chris Hoffman or a guest host)
We can riff on that if you want later. But it's more about the sustainability of running it and keeping it going. Scott, I mean, I would love to hear, because it's got to be a big question for you.
- [41:50]
Scott Turnbull
Yeah, this was probably my number one topic of conversation. There's a tendency in it to want to focus on your solution because borrow. We're also very, very smart. Right. Like, I did a lot of work on this and I work with a lot of entrepreneurs too. The first thing in entrepreneurism you have to do in IT entrepreneurism is you have to let go all of this effort and how smart you think you are. And the customer doesn't care. Right. They just focus on the product. So I swear AWS is not paying me here for advertisement, but I think you have to be laser focused and stay as light in code as possible. And I would say this all the time. I don't want to hear about code. I don't want to. I don't want to hear about any of these Python libraries. I want us to leverage N of cloud resources. Absolutely. 100% as much as possible. And we will have a thin layer of where our value add is. And that might be parsing data out of the humanitarian data exchange or wherever you think your value add is. But everything else, you have to Ride along AI, especially right now, you're going to bleed out trying to chase the tail of any AI. It is changing so fast. I've been an IT professional for 30 years. I have never seen things change as fast as they are here. Whether that that's deep seek being released or chain of thought models or now they're going to go to test time compute and now they're going to go to test time training and it's just like every few months the table is flipped and you have to go on. So right along with services like whatever your cloud provider is, they're going to have a great suite of services. AWS has a great suite of services as well. Plan a little bit of head against David and his team was fantastic in giving us heads up of like actually these models are the ones that we're going to stay with our bedrock and is going to work with our agent services. So plan ahead for that and sort of getting ahead of the ball like you're running down. I hate to, sorry to use sports metaphors, but I'm running downfield and I'm the striker. I'm looking to get the ball back pretty quickly from somebody, but I need to be positioned on that field correctly because tech is a chess game. It's not a game of checkers. You have to have the pieces positioned in the right and then the last move might seem easy, but you have to know deeply what the tech does, what the problem space is. So I sorry, I'm over answering but focus on your data, focus on your value add piece and ride along with everything else that you possibly can from turnkey services that you might get out of a partner like Amazon.
- [44:05]
David Master
And I think another key piece is lighter models. I think if we're using too heavy models, it's overkill for a lot of what we're doing. And there's clever stuff you can do with architecture where you query a smaller model first, then escalate to a larger model as needed. So I think there are little things that can help run more efficiently and have a big impact on the cost and energy consumption required for these tasks. And what you're asking of the compute for sure.
- [44:32]
Host (possibly Chris Hoffman or a guest host)
I mean when we talk about. You mentioned rag earlier and we didn't go too far down there, but I definitely feel like that's an easier win for a lot of NGOs. It's a way to reduce risk. It's a way to train just on your garden. You garden fence it, you put all your stuff inside your garden and now you've trained it just on those pieces. And as we move forward and as NGOs might want to learn more about these and take these types of things forward, they say, I want, like, I'll give you, I'll give you a use case, a quick one. So I want to be able to have somebody enter into a chat with the chat functionality. They're in a crisis. We're in a refugee camp of, let's say, 100,000 people. And they want to be able to find out which organizations offer what services. Right. In the refugee camp. Then they want to be able to register for those services. Right. So now you've got a form that they have to fill. Right. And then you want to be able to allow them to apply, not just fill out the form, but then get feedback from the system on that they were accepted or they weren't, and give them directions on where they need to go or information or where to take their child for the vaccination. So all of these services. So now when I think about this, we already have that, right? And it's called a customer experience.
- [45:45]
David Master
Right.
- [45:45]
Host (possibly Chris Hoffman or a guest host)
And this is what companies have been doing for years. They know your customer profile, they know how many times you went to the website, how many times you called in, how many times you checked their WhatsApp chatbot. Why aren't we applying that and following that model and is it applicable to our way of working as humanitarians? Do you see what I see or am I completely seeing the wrong thing? You know what I mean? And I guess maybe, Scott, to you.
- [46:09]
Scott Turnbull
Yeah, I absolutely see that too. To my colleagues, I lovingly say we're not as special as we think we are. Right. Because everything has to. But it's not a humanitarian interference face. Does that matter? I do think there's some fear of vendor lock in. You know, you get, you get 60% of the way in and you realize this isn't the right solution. And now you're, you know, your spend is like a $500,000 and you, you know, it's wasted. I think this fear, that's legitimate. That's legitimate. But the way you fix that is that you have very experienced, broad set of IT professionals who've used these applications and services, know their capabilities, and can understand your domain as well and tell you how they apply. So you're absolutely right. Like a ticketing system probably works really well for what you just said. I don't know how many humanitarian project planning meetings I've sat in and said, I think you just asked me to recreate Excel. I've actually said those words a lot. Or tableau. Right. Replace either one and sort of this one little special piece.
- [47:11]
Host (possibly Chris Hoffman or a guest host)
So.
- [47:11]
Scott Turnbull
But how do you tell people. But back to my own thing, challenge of understanding the domain, I'm hesitant because my first reaction is, well, I don't understand the problem space enough. And about half the time that's true.
- [47:23]
David Master
Right.
- [47:23]
Scott Turnbull
I get like, oh, you know, they actually have to have, they have data security requirements or they have something about the way their data is collected that this doesn't work with this. Well, but it's only half the time. The other half it could work.
- [47:35]
Host (possibly Chris Hoffman or a guest host)
I mean, David, David, on your side, sitting on the business side of things, when you're engaging with customers, right, and you're talking to them, what are their concerns that they're bringing to you? Is it the lock in? Because I hear that a lot, right. In Scott just now it's cloud lock in, it's now all the services lock in. Or is it what's my long term cost going to be? What are their big concerns that they're bringing up?
- [47:57]
David Master
It seems to largely be a kind of a lack of understanding. It seems to be sort of. And privacy concerns. A big concern is whether or not their data is going to go back and train the underlying model. And cost is of course a concern as well. We talk about how we do our pricing and we have cost estimators. So after we architect a solution, we'll put in what pieces are needed, what's the data volume, what models are going to be leveraged, and we can estimate costs. But the concern seemed largely around privacy. And is there data going to get out? Is it going to be exposed, is it going to be used to train models? Is mostly what I see. So a lot of times what we'll do is we'll put together a workshop on responsible AI and we'll go through Amazon's framework for responsible AI and how we think through our own use of AI and the steps and procedures classifying the risks and mitigating it is usually how we handle that concern.
- [48:59]
Host (possibly Chris Hoffman or a guest host)
Absolutely.
- [49:00]
David Master
But you mentioned filling out forms earlier. One of my favorite use cases from last year was a organization that helped provide caregivers. And a lot of the caregivers were helping people with disabilities. And in order to hire a caregiver, they had to register as an employer. And there was all this paperwork. In some states it required notarization. And a lot of times it's the person with the disability filling out this paperwork themselves. So being able to engage with that person through conversation and have a bot that will just talk to them and ask them they need and then use that to fill out the form and then give guidance on if there any additional steps to get it notarized, whatever the process is, and then help them hire somebody, manage that person. So I think that has a lot of applications in the refugee space with various forms in afl, various legal requirements, language barrier. So helping people who are new to this situation, understanding what information they need and then even helping them fill out the forms and advising on what they need to do with, with that form. And yeah, I think, I think I haven't seen it yet in the humanitarian space, but I do think, you know, the application is certainly there and I'd be interested to see it.
- [50:15]
Host (possibly Chris Hoffman or a guest host)
Absolutely, yeah. Well, to be fair, our main technology partner is Twilio and so Twilio's got their new AI Alpha. And so what you're able to do is you create the conversation with the LLM and you train that conversation on engaging with the person. Then when they get to the point of what they need, you can code in the form and the form standard and then it goes back into the LLM. So continue the conversation and then what you're able to do is you can anonymize it as well. So if you're collecting certain information, you can anonymize what goes into the LLM to train it and what stays in your code base. But that goes around then to your whole need of having, you know, your own common platform where you're storing your information. So your data lake. Exactly. And all that. So all those pieces and I look at today and in my optimism I still though with all those cuts that we just saw happen. In the last 24 hours, three major organizations have announced over 10,000 people in cuts in the humanitarian sector. And I start to worry that while we were talking earlier about the efficiencies and moving forward and being able to have a force multiplier that actually I'm really scared of. Also the potential for regret of going back and I really worry about that. And so for the last kind of question of the podcast, I wanted to come to you guys and say, what do you see as the future, the low hanging fruit future and the extended future. And we've talked around it, I think many different pieces of the puzzle and some really great use cases which I really appreciate. But what is the future for AI in the humanitarian sector? We're talking about frontiers for AI in the humanitarian sector on the podcast. So from your perspective, David and Then over to you, Scott. What do you think is, is that low hanging fruit that we can use cases potentially that they can start to think about and then what's that long game that they're looking for?
- [52:01]
David Master
I think there's an expression I like, innovation happens at the edge and scales in the middle. So I do think that's what we're starting to see. And we're seeing the innovation at the edge and that's where it needs to come from. And being informed by the people out there who understand the needs and understand what information would help them do their job better and be able to help more people and then building tools to get them that information. And I do think it needs attention and focus and people. So these cuts in funding and cuts in resources is really tragic. And then technology isn't going to replace that. We still need the people and the attention all the more so because it's iterative process, you know, it requires experimentation and it requires testing ideas and trying new things. And that requires a lot of people and a lot of people at the edge dealing with these problems and coming up with solutions and then the ability to recognize the potential and scale that. And it's, I think cloud providers are in a position to work across, you know, a lot of those organizations doing that interesting work and then share best practices and knowledge learned and help scale those practices. But I do think it requires a lot of people in the field digging into the problems and being vocal about what they need and technologists themselves getting, Scott, getting out there and understanding how can we better inform these humanitarian workers, how can we get them the information that they need and iterate on it and build on how technology can better inform these decisions and increase the impact of, of the dedicated personnel that we have out in the field doing really incredible work. So that's the future I want to see is lots of interesting creative approaches to things and then recognizing where's the potential to scale and where can we take lessons learned and apply it to just make it standard in the industry. And I think that's the future I'd like to see.
- [54:07]
Host (possibly Chris Hoffman or a guest host)
Absolutely, Scott.
- [54:09]
Scott Turnbull
I mean, I would agree with all that. I think in the near term actually stood out about capacity. You said we're losing 10,000 at least temporarily. My hope is that we'll find those resources again and bring those people back into the humanitarian sphere. But we're losing a lot of capacity, we're losing a lot of domain knowledge from the humanitarian sector with the recent development. So education, so UN data 2.0 charter was very. I think accurately stated that we need to focus on capacity building and education. AI can be used to enhance education dramatically. Like I, you know, things are developing faster than I've ever seen it. However, I'm able to learn faster than I've ever been able to learn before. You know, learning doesn't need to be sitting through a really terrible webinar or a really terrible like video series. You can really rapidly learn things quickly and have a check with AI resources. So internally as a way to improve ourselves and bring back that capacity. That's one, two. I'd like to see in the short term a bridge between back office administration and field operations. I actually think field operations are going to lag a little bit because there's too much chaos, it's a little bit too unstable. The infrastructure to run a lot of AI services in the field, although that's coming. But I think we have all this qualitative data that I said earlier we can extract a lot of meaningful insight out of. I think we need to focus on that to make it actionable for people in the field. So bridging back office administrative bureaucracy like reduce the impact and having a quick turnaround and making it impactful for people in the field. I think that's a short term, long term I'm actually going to go off roading and say I think long term robotics is probably the biggest impact. We're going to see humanitarian response, whether that's drones in the field, delivery robots, medical robots. I mean we're talking about 10 years from now or five or 10 years from now. I think field operations are going to be heavily impacted by drones and robotics in ways that I don't think we can truly conceive of right now. Wow.
- [56:01]
Host (possibly Chris Hoffman or a guest host)
But you've given me something to think about for the rest of the day. That's.
- [56:03]
Scott Turnbull
That's for sure.
- [56:04]
Host (possibly Chris Hoffman or a guest host)
I'm excited now for robotics use cases. I mean the drone thing. Patrick Meyer, I think many of you have heard Patrick speak before in the past and his work with drones has really been pretty amazing. And I think you're right. I think I really do agree with you that the idea of drones and robotics. Drones more so Are in my head right now. I didn't go that far off road on robotics but I love the direction you were taking us. So that was perfectly great. David and Scott, I want to thank you both. This has been such a wonderful conversation today. It's been fast paced, fast moving, man. We must have all over caffeinated this morning. So I think we've done a great job.
- [56:40]
Scott Turnbull
A great job.
- [56:41]
David Master
I haven't had any caffeine yet. I'm just jazzed on this conversation.
- [56:44]
Chris Hoffman
Well done.
- [56:45]
Host (possibly Chris Hoffman or a guest host)
That's awesome. That's great to hear and thanks. I know you're in different time zones. We're in three separate time zones, so thanks for making the time. Waking up early this morning there in the US Guys, and couldn't be more appreciative, couldn't be more thankful for your time and efforts and the work that you've done in the past and the work that you're doing, you will be doing in the future around this subject, around humanitarian action and then the frontiers of AI. So thank you both for joining.
- [57:06]
Scott Turnbull
Thank you very much.
- [57:08]
David Master
Appreciate it. Yeah, it was good fun and good to see you, Scott.
- [57:10]
Scott Turnbull
Yeah, great to see you, David.
- [57:13]
Chris Hoffman
Thank you for joining us on humanitarian frontiers in AI. We hope today's conversation gave you new insights into how AI is transforming humanitarian efforts and the steps we need to take to ensure it's done ethically and effectively. If you enjoyed this episode, be sure to subscribe and stay tuned for more discussions with leaders and innovators at the intersection of technology and humanitarian work. Together, we're exploring how AI can bring real change to communities in need. Keep pushing the frontiers of possibility.