Microsoft Research Podcast - Can we AI our way to a more sustainable world?

Episode Date: April 20, 2026

Doug Burger, sustainability expert Amy Luers, and optimization researcher Ishai Menache examine the global emissions implications of datacenter operations, efficiency gains, and AI's potential across electrification, materials, and food systems.Show notes

This is the shape of things to come, a Microsoft Research podcast. I'm your host, Doug Berger. In this series, we're going to venture to the bleeding edge of AI capabilities, dig down into the fundamentals, really try to understand them, and think about how these capabilities are going to change the world for better and worse. In today's podcast, I'm bringing in two experts to have a dialogue about the future of AI and sustainability. One, Amy Lures is an expert on sustainability and the intersection of sustainability, technology, and science. And the other, Ishaim Mnasha is a world-renowned expert in optimization.

And so thinking about how technology can optimize systems, we're going to talk about whether AI has the potential to help with climate change and sustainability and the degree to which there are challenges associated with AI. And we're going to try to get to the root of the issue because that will determine the shape of things to come. I'm really excited about the two distinguished guests I have today. We have Amy Lures, who's Microsoft Senior Global Director for Sustainability Science and Innovation. And we have Ashai Manasha, who is a partner research manager at Microsoft Research. And then the topic, of course, is AI and climate and sustainability, which I think is on a lot of people's minds. You know, we have a climate crisis happening. I've been a climate hawk since the 1990s.

That's something I worry a lot about. I care a lot about. Of course, Amy has devoted her career to it, so I can't really talk. But it's a really important issue. And now we have this AI transition happening. We're doing across the tech industry a large buildout of large, large, large computing systems, mega data centers. And there's a lot of concern in the world about how this might affect the climate,

how what this means. Amy will talk and think about local communities as well. And so I really wanted to dig in to the facts. What does this really mean? What is, do we think the impact's actually going to be, like let's separate the data from the hype, and then also talk about some of the opportunities ahead because I do think there are things we'll be able to do.

And that's why we have a shy here. So maybe I'll first turn it over to Amy. Can you tell us a little bit about your job? job at Microsoft and what got you into this space, maybe a little bit of your story. So, as you said, I laid the sustainability science and innovation in the Microsoft Corp sustainability team, which really means I get to work with really smart people around the company, around the world, at MSR, on shaping and informing sustainability solutions for Microsoft, but also for the world. And part of that is leading our strategy on AI and

sustainability. And how I got into it, I've been working on sustainability and climate my whole life. And I've worked from the tech sector. I was at Google actually previously. Also was in the White House working at the intersection of the CTO's office and environment and resources and energy. I also led a international research institution, UN-based network, rather, focus on sustainability. And in that context, after coming out of Google, where I was really started to think about the power of compute and digital tools for transformation, and which is why I was brought into the White House to work at that intersection, when I started leading the sustainability network, research network globally, future earth, I really brought this need to think about innovation and digital technologies in that space.

And I will say the sustainability science network at that time, you know, it was 10 years ago, eight years ago maybe, was a little resistant to thinking about AI and technologies in this space. And I started a global initiative called Sustainability of the Digital. age, where I really brought together the digital technology and AI community was in Montreal, which there's a lot, a big AI community there, and the sustainability science globally, and really started to think about what are the potentials, what are the risks, and led a big international study to put together a research and innovation agenda in this space. And that sort of really shifted my approach from just big compute can help things, which I,

was really focused on it at Google to this role of AI and machine learning in this space. And Ashai, so you're a world-renowned expert in now ML and optimization. You've published extensively. You're, I think, famous in your research community. You've had, I think, broader, you've had a lot of impact on Microsoft's business. You've also been published in the Harvard Business Review, so you know, you're a little bit polymathy and sometimes a little intimidating to me. Yeah. But I'd like to hear a little bit about your background.

Just, you know, a short version of your story for the listening audience. Yeah, my background is actually in engineering. However, my graduate studies were, as you mentioned, like in ML, reinforcement learning, later on distributed optimization, game theory, a little bit more on the theory side. So my story is that when I was doing my postdoc at MIT, you know, the cloud was kind of on the rise, circa 2009 or so. And I got fascinated by the cloud. My initial interest was actually in the economics of the cloud.

And you know, pricing, how you price the cloud. And I got to know about MSR, because around that time, there was a new kind of lab opening just by MIT, MSR New England. And I got fascinated by the cloud and not only the economic aspects of it, but more fundamentally, how do you utilize resources more efficiently? And that's what got me to Microsoft research in 2011. So I was consulting in MSR New England, but then moved to Redmond at 2011 to join

a lab called Extreme Computing Group that was actually dealing with the Cloud Futures. If I can mention, Doug, you were also part of that. So I've known you for quite some time. And you know, so, you know, so. So sort of my, let's say, my angle into that. So there were like a lot of systems people thinking about the, you know, infrastructure of cloud.

And at the other extreme, there are theoreticians. They're thinking about like, you know, the next kind of wave or like, you know, innovating in the area of algorithms. But I think what was sort of missing is a little bit of bridging between, you know, algorithms and then cloud infrastructure. And that's where I sort of found a very interesting niche for myself and later on for the group which I founded in 2019. So you recently announced the system called Optimine and I, you know, I did a LinkedIn post

about it because I was really excited about it. And just tell us what the system does. Like why, it got a lot of attention. So what does the system do? And maybe we'll dig a little bit into optimization for the audience and then, but then we have to get back to AI. For sure. So, you know, know, stepping back a little bit. So what is actually optimization or mathematical optimization? So optimization or mathematical optimization is a way of using mathematics to make the best decisions when there are many choices and some limitations. Okay. And, you know, just a little bit more, you know, concretely. So in every optimization problem, you have, you have, first of all, a description of the problem that you're you need to solve.

Then you have a bunch of decisions that are actually in mathematical terms. These are the variables. You have an objective. What is your goal? What are you trying to optimize? It could be something that you're maximizing revenue, but it could be that you're minimizing costs.

So the different versions or different kinds of goals or objectives. And then there are constraints, which is like, you cannot do whatever you want. There are some sort of limitations, such as capacity constraints in the cloud setting, or other factors that you have to account in order to come up with the best possible decisions. So maybe a symbol, just to be a silly for a sec, so I have a complex drive to work,

and one day I find that the way I usually take is blocked, and my brakes are really worn and I can only stop twice. So your framework might be able to figure out, like, what path gets me there saving the most gas? Right, so that's one example. And maybe you don't want to pay for tolls. so for some reason. So that limits, you know, the roads that you can take.

You know, there's speed limits and such things. These are all constraints that you have to account for. There might be some speed traps, but I'm willing to go by a speed trap if the route is much shorter. So stuff like that. So it gets pretty complicated, doesn't it? It gets pretty complicated because especially when the, you know, maybe you're a single driver,

but in optimization settings, think of like some of the problems that we worked on. actually with Dynamics 365 was also in the context of field service, which is about managing technicians at scale. So think of like, not just you, but thousands of technicians that have to fulfill or have to take care of certain work orders. So it would be thousands or 10 of thousands of work orders. And then you need to assign the technicians to these work orders. And there's a bunch of constraints. Maybe not every technician can do every work order. There's a you have to account for the the traveling of the technicians, right?

So it's like you're not going to send a technician that is in Spokane to do something in, let's say, in Seattle because, you know, all day will be wasted on traveling. And it's not sustainable. It's not sustainable. And also, you know, the gas, you know, obviously. So all these kind of considerations, you kind of map it formally into mathematical optimization. And then there are techniques of solving. this problem to optimality. So essentially, there is some machinery and there are experts that can

take these problems and come up with the algorithms, but not everyone can do it. So it requires some expertise, in fact, graduate level expertise in operation research or in, you know, algorithms, computer science type of algorithms. And when Gen AI was emerging, we saw an opportunity to democratize optimization with Gen AI in the following sense that a person that is not an expert can define what they want to do. So you gave your example about getting to work. Could be like a simple example of packing, which is like I have a suitcase that I have a limit like 20 pounds and I have a bunch of things that I have to that I want to fit in. Like you know, I have with certain importance, some are more critical. like, you know, I don't know, like my laptop and all that,

but then there are books that are quite heavy and maybe I still want to read books. Or I'm running an airline and I have to schedule the flights and I want to minimize fuel. Yeah, that too. And essentially, so you want to be able to describe what you need to solve in plain English, specify the problem, say what the decisions are,

like I mentioned, like what your goal is and then what constraints need to be accounted for. And you want to use AI that will help you take all these considerations and essentially formulate the algorithm itself. So write down the recipe, the mathematical recipe, that would produce an optimal solution. So that's what the optimine is about, is a small language model that was trained, especially for this kind of scenarios, of, you know, taking natural language and mapping it into an optimization algorithm. So this is really great. And I think we're going to come back to this. I want to now go back to Amy.

When we think about AI and these data centers that the industry is building, and they use water, they use electricity. You know, there's contention in some communities about them being placed there. If I, you know, I love to be really data driven and just kind of very factual. So if I look at the overall picture,

like what is the real impact we think of this transition on, you know, climate, sustainability, It's complicated, right? Because there are many sources of emissions. Electricity gen is one, but you have renewable energy, but it takes materials to build these things. So can you kind of give us some framing to help us understand it? Yeah, so first of all, you know, I think when we think about AI and climate,

a lot of people think about just the infrastructure side. And I think it's really important to think about this holistically. I actually personally believe that AI will be one of the most influential factors determining our climate future for better or worse. But I also believe that we actually need AI to solve the climate crisis. So with that as context, let's talk about the infrastructure.

Remembering we have to really think about the full context. Let me put this in the context. So from a climate perspective, what matters is the emissions to the world, the emissions of greenhouse gases to the world, heat-trapping gases to... climate. Not specifically energy, right? Because energy can be in different forms. Right. It's what are you putting in the air. What are you putting in the atmosphere? So, you know, if you think

about it from a global perspective, the world uses about, energy itself accounts for about 75% of all of the emissions that go into the atmosphere. Wow. That's a lot. So that's a lot. But a lot of people think it's the whole thing. So there's other things that are not energy. Three quarters. Three quarters. But in the context of, so from a climate perspectives, data centers account for about 0.5, less than 0.5% of all emissions as of 2024. Okay. Okay. But they're growing. But they're growing.

Yeah. And so if you're growing and you think about, there are lots of projections, it's hard to project really beyond a couple years, as you both know, because things are changing so quickly, both on the demand, on the efficiency, what we're using. Are we going to be using small language models? Like, we don't know what the future looks like. The IEA projects by 2035 that the electricity use could double. And so from electricity use, actually, data centers use about 1.5% of global electricity. And that could double, could be between three and five, even more than doubled. But that still, from their projections, it would still be less than 1% of global emissions.

So even if that would double in that space. So it's still in terms of a global emissions perspective, which is what the climate cares about. It's a small percentage. Can I just go back for a second, break that down? So energy is three quarters of, generis three quarters of emissions, but that includes burning fuel, transport, and then what fraction of that three quarters, or let's say just total emissions, do we think electricity is?

So electricity is about 20%. The energy that's produced is consumed about 20% of it is consumed as electricity. Got it. Now, in terms of emissions, about 35% of the emissions from energy is from electricity. And part of that is because electricity,

the reason that difference is... You've got coal plants. And you've got coal plants. And it's not as efficient when you do coal plants. You actually get efficiencies when you go right from solar to... in terms of just the energy, because you lose a lot of heat in the thermoelectric plants, right?

So there's an efficiency there. But so about 35% of the energy emissions are from electricity. And electricity production is really the key issue. When you, you know, the key issue of today is like electricity and data centers, right? How are you going to get enough electricity? How are you going to get enough clean electricity? and that is something that is often more of an infrastructure problem than actually the energy problem. I mean, they're both true, but it's getting that electricity in the right location at the right time.

And that's sort of a big... It's a big messy problem that we can unpack a little bit because I do think there's a role for AI, a huge role for AI. Maybe even an optimization problem. We've got this guy in the room. So we should, we should unpack that. But I think before we go off that, just two points that I think are relevant. One thing is that what is often not necessarily realized by people who don't spend their lives, thinking about climate, but to tackle the climate problem, we need massive amounts of more of electricity.

That's just part. I said we had 170,000 terawatts of, energy, most of that to be able to solve the problem has to come in the form of electricity, because that's what we can decarbonize easiest. So one of those challenges is actually more electricity. Got it. So let me again try to break this down to a simple statement. So we've got, you know, about 35% of emissions are due to electricity use. Or 30% of energy emissions, which is three quarters of the pie. So we can do the multiplication. And, and, and, And of course, you know, as we decarbonize electricity, there is a probably too slow but

ongoing transition towards lower carbon emissions, a generation of electricity. You know, you convert a coal plant to a natural gas plant. It gets better. You do put it to solar wind. It gets even better. But then the demand for electricity is going up, in part fueled by, you know, the tech industry and building the data centers and AI, but in part because, like, we have to stop burning fossil fuels.

Right. Electric vehicles, turning all of our heating into electricity, electric, everything, turning everything. The phrase is sort of electrify everything. The head of the IA says we're in the era of electricity. So we have a huge pressure on electricity demand. And we need to have it be low-carbon electricity generation. But the demand is just going to go up and up and up with or without this data center build-out and this AI build-out.

But of course, this is happening, and the hope is we can use it to provide a lot of value. Yeah, so there's one other sort of context, which is really important here, and that is that the other concerns that are being raised around this issue of electricity and data center growth is at that local community level, right? So data centers contribute a very small, perspective, globally in terms of emissions, and even in terms of electricity. but they're really concentrated. They're one of the most concentrated industries in the world.

There's this great figure, if you ever want to look up the recent IEA report on energy and AI, which really shows the concentration of different industries and steals way on one side, which uses 7% of the emissions, which produces 7% of global emissions, and data centers are all the way on the other side, which are real in terms of the level of concentration, how close they are together. And the reason that's important is that there's certain pockets of the world where there's really lots of data centers

and they keep going into those areas. It's changing a bit now because of the dynamics that are happening. But when that happens, then in those areas, yeah, they're a major user of electricity, right? And when the growth happens quickly in those areas, then it can be, it can put a strain in local grid. And so there are concerns that are being raised in certain regions in the world about data center growth.

And, you know, I'm really optimistic that those are mainly infrastructure problems and they can be addressed. And we need to figure out how to do that. And I'm really, that's why I'm so excited about in January, we announced our community first infrastructure initiative where we're really focusing all of our work now on how do we designed to our data center development to ensure that that rapid growth is not a net negative, but actually a net positive for those communities. And that includes, you know, committing to paying all of the prices that requires to meet

our electricity needs so that our data centers do not drive prices increase in communities. So I can see how we can get to a net level. Like, you know, we basically don't drive up local prices, don't exacerbate local water supplies. of bring it in, however you can do that. But how do you make it a net positive? Well, I think that we have been saying we've been net positive at a global scale, and I think we're shifting that to say,

what does it mean for net positive at a local scale? And I think at a local scale, for example, data center water use for cooling can go net positive in the sense that in the data centers themselves, we are actually beginning to design systems that use essentially zero water for cooling. Right.

It's a cycling. Certainly. So we can, you know, that isn't in place everywhere, but there's happening that. And then we replenish water. So we can do things. So for example, it turns out in many places, water in some cities is lost with leaking pipes. And AI, it turns out, can help identify those leaking pipes. So even if you get half of that, even if you save half of that, that can be, way, are the

emissions from data set, the loss, the use of water in data centers is only a fraction of what we would save. And so you can amplify and save water by using, oftentimes even by using AI. Another optimization problem. So this is more of a, this is more of a commitment from the company to work with the community to get them to a better place, but of course using our global compute, because we're not going to just run the AI analysis that makes the pipe better in the data center that that community is in. Oh, no, no. And it's not just, but it's not just, But we're not doing it just with AI. Of course not.

Of course not. We're also investing in training. We're investing in NGOs. The real focus is to really understand what that looks like. And, you know, my interest is really to say, you know, can we increasingly co-design? What does community positive mean? This is all new because this rapid growth was the first time that this became such a serious issue of a concern. Right.

Right. Well, I'm really glad you're pushing on this. And I think it's really important for the communities. Because if we can put enough focus and enough innovation to make these things a net positive, I'll certainly feel a lot better about it. Going back to the emissions, so one of the reasons I brought you both here today was because Amy gave a talk at our research showcase, which really inspired me. And you talked about a lot of the emissions being generated through things like transport and inefficient,

management of some of our large, you know, non-digital systems. And I heard that. And I'm like, wow, this might be a place where some of our research could help. And of course, that's my job, right? It is to find places where our researchers can amplify the effect of their work. So, Ishai, maybe just very briefly, can you talk a little bit about when you worked with the supply chain and you applied your optimizations techniques? And this was even before optimized where, you know, you could do scenario planning,

for natural language, what sort of efficiencies were you able to drive? Just working internally with our business teams. Yeah, so I'll give a couple of examples, but starting off with what you mentioned about transport, actually. So as part of this intelligent fulfillment service, we're actually accounting for shipping costs. And one thing that one can do is account also take more explicitly into account, you know, emissions and sustainability considerations when doing this. shipping itself, right? So that's one kind of one option. So we could say like get have the same like

level of risk and time to delivery, but as your objective, just try to drive your emissions down. Correct. As one example. Correct. There are other related examples of when, you know, what kind of hardware you use when you, for to fulfill the requirements. And for example, you want to use hardware that has been sitting for a long time in the warehouse. So that also has some implications. Now, we've worked over the years on various systems that efficiency has been a major goal of us, but sustainability is closely related to efficiency. And I can illustrate it through concrete examples.

So one is a virtual machine allocation, which operates at the timescale. scale, it's the process that essentially maps VM requests to physical servers. That's in our cloud. You know, when a customer wants to use something, just again, we're keeping it like not too geeky here. So one of our goals was to increase pecking density, which means that we want to operate the servers close to 100% of utilization. And it is well known.

Actually, it's a study by Google, I guess you're familiar with that actually, you know, as you increase packing density, actually you reduce the power per unit of useful compute. Yeah. Right? So that's well known. So for example, I don't remember the exact numbers, but if your server is utilized at 50%, you still consume close to 100% of the power. Yeah, you've provisioned it, you know, it's being transmitted,

you know, your chips leak, they have static power dissipation, patient, all the sort of stuff I used to work on. That's actually diagnosed better than me for sure. You know, that's been this area of research. So another example that we've worked on since 2022 is rec placement. So you have these demands and you have to decide how to exactly place them, these wrecks of servers within the data centers. And there you have to account for power and cooling and, you know, space and all that.

One of the things that we were able to achieve with our optimization is reduce power fragmentation by 1 to 2%. So 1% here is huge. And it's not only about the COGS saving, the money savings for the company, but it's also about having to build less data centers or utilize the data centers in a more efficient way. And essentially, another way to view that is that, you know, if we are, we want to have AI, we want to have AI consumed by everyone in the planet, right? So essentially, you're making it more kind of with less, you can do more and have AI, you know, broaden the reach of AI and having it consumed by, you know, eventually all the world. I mean, I think, I think internally in the company and with research and working with our product groups,

been very successful because you've figured out a bunch of ways in supply chain and in one of our largest businesses to make things run more efficiently, which is great, right? And, you know, but to the people in the rest of the world, they care, like, you know, is the plan on a sustainable trajectory, right? You know, are there electricity bills going up? And, you know, we're doing it under competitive pressures to improve our operating margins, which is great for us. But we have to think about the world. So, but I think there's, we have a, something we can do here. So now going back to Amy for a sec, in your talk, you talked about, you know, a lot of these emissions and opportunities, are there a few big buckets you think we can tackle by just being

smarter about how we manage complex systems? And what do you think the magnitudes are? Earlier, I said, I don't think we can actually solve the climate crisis without AI. And I think there are three game-changing capabilities that I see AI really brings to the sustainability process. And one of them is this ability to enable the way I think about it is optimization, but also understanding and predicting complex systems that we had a hard time doing with, or it's really impossible to do with the traditional analytical tools that we've had. Right, right. And the other game-changing capability is the acceleration of discovery, development, and deployment of new climate solutions. and that's also, of course, another area that MSR does a lot of work in. Right. And the third game-changing capability that I see AI brings to sustainability is the ability to enhance and augment institutional, human, and workforce capacity.

And I think all three of these are important and can be game changes if we lean into them. In the context of specifically the management of optimist or the optimization of complex systems, I think, you know, the biggest challenge that we really have to do is, as I said, transfer, electrify everything, right? And that has many levels of where optimization can fit in and accelerating discovery fits in and the third green paths. Like it has various different levels. But one piece is infrastructure. So the IA projects by 2050 to get net zero by 2050, which is the global climate. goals at a global scale, we need to more than double the electricity capacity of the world.

I mean, now, think about that. That's a big undertaking. And so increasingly, that ability to be able to integrate and manage a system while the climate's changing and the weather's becoming much more predictive, that becomes a challenge that if we get it right, we can integrate more. And if we don't, we're going to be really going very slowly toward that goal. Yeah. I want to actually.

I saw you twitching in your series. So we were talking a lot about optimization, but there's also the AI itself that helps along the way to make systems more efficient. And, you know, what we do a lot in the team is actually combining AI prediction techniques with optimization. And that is very powerful. And I'll give a couple of examples. So for example, in the virtual machine allocation set of problems, so what we did, we actually, you know, we predict the lifetime of the VM.

Like, how long do we think that the VM will stick in the system? So there are kind of AI, complementary AI systems that do this prediction for us. And that produces a more better optimization when you have like intuitively, when you have more knowledge or sort of Better predictions. Better prediction. It doesn't have to be super accurate, but as long as you have like decent predictions,

you could do much better with packing. So that's one example. And you know, in the context of cloud supply chain, predictions, demand predictions are critical. Because if you have very bad predictions, then you need to sort of provision ahead. Like we talked about this, that you have to sort of. You have to buffer capacity.

You have to build extra capacity in the system. you've got redundancy, it's expensive. You don't know what's going to happen. So, you know, optimization helps you automate and find the sort of optimal choice. But if your demand predictions are bad, you're going to have to provision, over-provision these hot buffers. And that also has an environmental effect as well as, you know, cost implications. So I think this kind of blend of AI and optimization, this is something that we are really,

I think really is going to kind of drive things forward. I want to be a little careful here because, you know, we, we, there's a lot of excitement about this technology because it's so disruptive. And, you know, for people listening, you know, I'm an anti-hype guy. Like, I used to like to work in areas that were not hot because, you know, communities would get over-excited about stuff. But the stuff we're seeing these advanced models able to do is just jaw-dropping. And it is the capability.

are actually moving really fast. But those can have bad consequences for society, too. I mean, you know, the Internet had a set of consequences. Social media had a set of consequences. You know, we are on an unsustainable path right now, trying to get on a sustainable path. And that's not all tech. That's just human civilization.

So I really want us to focus, maybe dream a little bit now. Like, what are the places we can try to steer this tech, you know, AI and all of this compute? to solve some problems that really move the needle for the climate. Because my own hope and goal, and I think, I mean, we probably share it. In fact, I think we all do is as technologists, you know, steer the technology in a way that helps people in humanity and overcome some of the bad effects. Because, you know, we're going to get both, as with any new technology. So, like, should we be managing cities with this stuff?

Should we build models that just control the grid? I control not in a sense of, like, you know, just that our agency. able to optimize, like, what are the things we could do five years out, two years out that might be magical and would really move the needle? So, you had the three. Yeah. So maybe we could bucketize it in those. Well, I had three game-changing capabilities.

So I would put those game-changing capabilities on problems, right? So first is this is what, when I look at AI, are the things that you can put on a lot of climate challenges, sustainability challenges, and really make huge differences. And then the question then is, well, what are the three big challenge areas, right? And there are a lot of challenge areas. I guess in terms of we can unpack these, but at the high level, I would say we've already talked about this, but the first one is enabling, electrifying everything. And there are various bullets under that.

The second one, I would say I would highlight industrial materials and chemicals. And that's a discovery, development, deployment type of thing. And it's packaged, like all of these, you can't just sprinkle AI and it solves the problem. But it can make it. It's real work. But I think that there's a different model in the second bucket of industrial materials and chemicals, I think. The thing that I really at least envision that there's a different way of approaching that because AI exists, We can unpack that if you like.

And the third one, I would say, is building, enabling the development of a low-carbon and resilient, secure food system. Food system counts for, you know, about a third of emissions. And it's also one of the biggest impacts of climate change itself, one of the issues of vulnerability. In food insecurity is a huge issue. Is this fertilizers? It's fertilizers. It's also.

like methane in terms of cows. It's in terms of, you know, and also how we grow food and how we distribute it in terms of supply chains. There's food waste is about 8% of emissions. So there's a huge amount of efficiencies. And again, I think the three game changing capabilities can, there are ways it can impact each one of those three buckets. Within MSR, we, just in the last month, you know, we've got, had teams design better a better X, where X was this percent more efficient using, you know, genetic algorithms or neurorevolution. You know, these aren't LLMs, but LLMs, you know, can play a role in this.

And I wonder if we were going to go tackle a set of problems, you know, so as Shai's work could think about taking large, complex systems and figuring out how to run them more effectively, and I think what you're saying is, could we, you know, could we design a better wind turbine that would generate more electricity, you know, for a given wind speed, or just through optimization, or new materials, like, for direct air capture, carbon capture. What would be on your list? Yeah, I think in the context of materials, you know, some examples, I mean, of course,

cement, some of the materials, a lot of the issue is tied to electricity. So it's the optimization of the process. So for steel, for example, you know, a lot of that is what energy you use to do it and how optimise, how you can optimize that system. But there are also lots of different materials that can make processes more effective. You know, materials for desalinization. Would you use a huge amount of electricity? And that also connects to this vulnerability and security piece.

Yeah, well, put up a lot or it would be huge if you could. You could do it with low energy. Yeah. And so the way, you know, one of the things that I have and sort of dreamed about is like it couldn't now, we used to have these challenges, these like grand challenges that, you know, were decades long that we would have moonshots. And I feel like if we focused AI on different challenges and said, let's think about those grand challenges in the areas that are really focused where AI can make a difference and say, Let's think of them as like factory moonshots. In other words, like, let's just say we are going to set up a system to be able to, what are the 10 materials that if we solved and addressed, it would really make a difference in energy and food and all the in society.

And we need just kind of check them off, you know? Like just get and get public-private partnerships to also do that together. And then you have to prepare, though. You have to prepare alongside to. to be able to have that system so that they can move into society, right? That's exactly right. And I'm, you know, I was going to pull up my, you know, pull up a paper here. Kristen Severson, who's an MSR and three of her colleagues at the University of Washington,

I think you know about the work, you know, used, used a form of machine learning, you know, Markovian processes to design cement that used, that had algae in it, but is as strong and, and generally, generates 20% fewer carbon emissions. Now, cement is a huge chunk. I don't remember what it is, seven percent. Yeah, so you take, how you measure it. Yeah, so you chop a fifth off of that and all of a sudden that's one to two percent,

between one and two percent of emissions. But of course, you know, and it looks like it will be as strong and have all the right properties, but getting it from the paper, you know, in cell to wide-scale production across the world is a huge undertaking. So, you know, we can solve these problems locally, but then getting them to scale,

especially when you're working in a company it doesn't produce cement. Like, how do you do that? That's a... Right, that's why you need it... You need to do this as a public-private partnership. Like, this should be thinking about it in terms of, we are going to have a mission to get 10 of these in the next, however many years, not trying to just have a moonshot for one in the next several decades, right?

So my list is like... So number one is everything that Amy said, like, you know, That's a long number one. Yeah. But I should say, you know, one of my mentors said, like, you know, every problem probably has some operation research or OR. So, you know, when standing in a line, McDonald's, that's where he said it to me. You know, to take this queue or that cue, you know, and like, you know, when you stand in line. So that's how he was thinking.

But, you know, to the point, I think the domains that you described, you know, materials, food distribution, electricity. There's a lot there to sort of that we can contribute as optimizers. And in fact, I think we can do a little bit more when it comes to concretely kind of accounting for sustainability metrics. I mean, we talked about that, you know, taking into account more explicitly renewables and such. So I think there is some way to go there. Number two, I think that, you know, optimization and AI can be used for humanity, for well-being.

I'll give one example. There are all sorts of scheduling systems and it's actually a project we started looking into, you know, how do you sort of make the system more efficient? You know, you want to reduce costs but still account, for example, for labor laws and, you know, prioritization or, you know, union kind of considerations. And I can help with that, you know, to actually understand the contracts and understand the fine prints and account within algorithms, scheduling algorithms that are good also for, you know, for the people, right? So it's like, you know, for example, if you take into account how many hours drivers, you know, they have these kind of long shifts.

How can you sort of prioritize that they have enough rest and account more explicitly for their well-being? So that can be done with a combination of optimization and AI. Lastly, and importantly, I would say, you know, part of the premise and like the vision of Opti Guide, which is the project that we've been working on in the intersection of Gen A.I. and optimization is making all these kind of complex, you know, tools that we have for making better decisions. And by the way, it's not only like I gave an example of, you know, optimization algorithms, integer linear programming and on that. But it's not only that. Think of very advanced simulation tools that you have also in the electricity space

when you model grid and stuff like that. Make them more accessible to end users, to planners, business operators, and executives that have to make decisions. So I think that's also something quite important here that, you know, AI can help facilitate. People, humans get, we get very comfortable with rapid change, as long as the change stops. Like we can ingest, you know, if a second moon appeared in the sky,

everyone would look at it, be like, oh my God, there's a second moon. And a week later, it'd be like, yep, there's two moons. And that's just how it is, right? And we'd normalize to it. And so we all think that the society we live in right now is normal, but it is a historical anomaly,

you know, exponential population growth, agriculture, industrial revolution, allowing a massive increase in population. And our whole economic, model is sort of based on exponential population growth fueled by unsustainable fossil fuels and resources. Like that's our standard of living and a big chunk of the world is still really poor. So I think we have to move to something different.

And maybe we just get to this enough materials to a sustainable future. But to your point, you know, it's not just about efficiency because these are human beings. The systems we build have to factor in human well-being. and that's incredibly complex. You can't do it with like a bespoke regulation. Regulation's hard, policy's hard. So my dream is that we can use these very complex systems to kind of evolve and learn that balance.

Like can we actually manage society and its complexity in a way that fulfills the human condition? I don't know if it's possible, but we're fighting this fight and it's really tough. So I guess for me, that's my hope. You know, Amy, you have dedicated a big chunk of your professional life to this topic. Like, what is your aspiration? Like, what would you like to solve?

If there's a problem we could solve in research, what would it be? Like, kind of what would you really like to see happen? What's your dream? Well, I think to answer that, I'd like to put a twist on your vision of that in terms of, because the way I see it is right now we have built over the year. a society that's based on a broken system. We have, you know, fossil fuels at the bottom and broken infrastructure

and inequities in the world. Massive inequities in the world. And so right now, any time we, like, put a new thing on that system, it just creates more emissions, more, draws more problems because we have the same model that we're doing. Right.

The unusual thing about AI and the reason I think it's so promising and so scary is that it's the first time we have a tool with such power that can actually change that system that we're based on if we use it to do that. Yes. And so I think we need to focus it on those things. I've written about, you know, certain five, and now it's become six, but I've written about sort of five things that have to happen for us to be able to direct it to change that system. And the first is, yes, we have to use AI for sustainability, right? I know that sounds sort of trivial, but it's not trivial because everybody often says, well, there's a lot of potential out there. And I'm like, yeah, well, the climate crisis, solving the climate crisis is figuring out the potential and then making that a reality. The second is data infrastructure and the data and data infrastructure to be able to solve these problems.

There's a lot of, we have to get that data out there, right? The third is making access to clean energy and reducing our footprint and supporting communities. Like those are those I think are the infrastructure side of it. Okay. The next one is what I call governing AI for Earth alignment. And we've, with colleagues from around the world of paper in nature of sustainability that sort of outlines this vision of what it means to have Earth alignment in AI, a principle, and I think that's another one.

And then the fifth is to upskill the world. Like you said make it more accessible. Some of it's not necessarily learning how to build agents. It's being able to understand how do you integrate these tools into your life in a way that actually can drive this change. And I would add one other, probably is closer to number three than to be in the six, but I do think we need to figure out not just to work with communities, but to align the development of data centers and AI operations with the needs and the trajectory of the electricity grid. And I think that's not just applying AI to solve solutions through optimization, but also thinking about this in an integrated way.

Yeah. Shah, you have a response, and then I have a question for you. Yeah. So oftentimes decisions are made by multiple parties, right? So, and, you know, each party makes a subset of decisions. And there is also, you know, they affect what others have to, can decide. They change essentially the settings for others that makes decisions. So there is kind of dependencies between decisions that I make and decisions that you can make.

And I think that AI can also help sort of can be the glue in this very compound, very complicated systems where there's like distributed decision making. And I think with AI, with all this kind of agentic sort of workflows, we can. can take into account moralistic considerations, obviously in the energy space as well. So I see a lot of potential in kind of solving for real problems without making too many discounts or just saying, like, I'm focusing only in my small world. With these kind of combinations of advanced analytics and AI,

we can actually get much further with moralistic and more global optimizations that are good for society. How do we take all of that and putting all of these decisions and, you know, complex systems and, you know, infrastructure and balancing and put it under control of these massive computational structures that learn things that are now too complex for us to understand and retain human agency? Yeah, I mean, that's a tough one. I wasn't expecting this question. Oh, my next one's worse. you know, I feel that there's not a perfect answer here, but I think that the explainability of these systems, of this sort of complex systems is part of what AI can unlock.

So you're getting these outputs, but as a user of AI, you can also ask questions of like why. And that's actually something we've been really focusing on in cloud supply chain management is like, you know, why are these decisions being made? Can you just explain to me the alternative? So actually with AI, you can explore alternatives quite fast and having this natural language kind of interface.

And then you can get better understanding why certain decisions have been made. Yeah. Esha, you're one of the top experts in optimizing complex systems. You've done phenomenal work. And I'm really proud of the work you've done and the impact it's had.

Do you have a dream, like, if your work could scale to solve some problems or do something intellectually beautiful? Like, would you like, when you look back on your career and you achieved X, you know, what would that be? The problems now are so important, and there's so much that can be done. And you're right in the center, I think, of what's possible. You know, I think the dream is to get into areas which, We thought that were impossible like a few years ago.

So like, no, this is not a place where some of your expertise, that's not for you. Yeah. This is like, you know, there's legacy and there's like, you know, it's just too complicated. The data is not in one place and all that. So we did a little bit with supply chain, which is, you know, a big challenge there is data. So I would say, you know, at least from a technical perspective, you know, using AI, optimization, advanced analytics to essentially create a unified decision intelligence platform where the sky is the limit. When you have maybe multiple decisions making multiple considerations, you don't even know when you start what all these considerations are. So having like a more interactive system that optimizes in a continuous form.

And, you know, one aspect of that is also, which is, I think, a big challenge is all these Black Swan events, rare events that, you know, optimizations, it's hard for them to, you know, account for it. So, you know, how do we kind of incorporate optimization in real time and have it reactive? And maybe, you know, with a bit less human in the loop, but the human is still an important part in, at least for the foreseen future, in verifying these outcomes and being confident about them. And I would add to that also. A part of it is like educating humans that are not like, maybe do not have the same, you know, the level of expertise. in certain areas. That's another thing that we have to think of. How do we explain to humans that, you know, have something to do with the business? How do we explain to them and make them feel comfortable about what the system underneath

with all this AI and all these complexities are doing? You know, I was listening to it. And it's really, that's a beautiful answer, actually, and it's really compelling. I'm going to try to restate it. Number one, take any system that you can define. no matter how complex, and drive it to near optimal for some criteria, for emissions, you know, managing an energy grid, for example, or how much renewable capacity to provision, or which problems should.

So that, like, if you can achieve that, right, then humanity has an amazing tool to just make, squeeze all of this waste. You know, we talk about, like, wasted food. That's a massive target you could go after. So just that capability, any system can be brought to near optimal. We just have to define it and then pick one. And then the other one I loved is how do you plan for resilience if you can speculate on a few of these Black Swan events? We know that these things are going to keep happening and they're going to get worse.

And we've seen the hurricane in Jamaica, right? And we just, you know, at North Carolina. And there's just, so how do we build societal resilience with these rapid changes? That's an optimization problem. Well, but that's an interesting question. I would wonder what both of you think about is that, I mean, I think there is a world that thinks that optimization and resilience are not necessarily always compatible.

And so, you know, there is a world where, you know, that resilience and redundancy in nature is about resilience. You know, that is part of the resilience of these systems. There's the reason we have two kidneys. Yeah, and so, you know, I think a lot about the resilience of ecological systems, and, you know, and there is a lot of concern about that. When you optimize it, are we actually making ourselves less resilient? I'm just curious. I know we're running out of time, but as this tension came up, I was wondering what you guys thought about that in terms of, you know, how we can move forward with these tools in that space.

Well, I'll jump in there and then turn over to shy. I mean, having designed, you know, with my team, large-scale systems that have to be, you know, resilient, you also don't want massive replication. And so there is an art to figuring out where do you over-provision and how do you do it intelligently so that you get, you know, given some model of failures in Byzantine problems like the right outcome. And so, you know, resilience can be just, I'm going to put a ton of food and water and fuel everywhere in case something happens. But you can also be really smart about it. So it really is an optimization problem. But you need resilience in your criteria for optimization. You need to know what you're predicting for and what you're trying to solve.

And then you optimize. And that's what I was trying to say to a shot. You optimize it. So my first answer was more like Ducks is on the system side. And definitely resilience is being accounted for, you know, durability and also things in certain systems, storage systems and such things that that DG has worked on for many years. I should say that more on the supply chain side and optimization side,

It's a big topic, resiliency. Yeah. And optimization can capture in a formal way, measures of risk, right? So you're not like optimizing, necessarily optimizing for the average outcome, average profit, average cost, but you can take into account explicitly risk measures when you're optimizing, like value at risk and other measures that explicitly account for risk. And there are other ways in optimization that you account explicitly for uncertainty, stochastic optimization, where you try to model the uncertainty in a form of, in a distributional form, robust optimization.

Maybe you don't have the distributions, but you still have some sort of polytop of possibilities in the world. And you just kind of, and you do it actually in network routing, for example. You're saying, like, I don't know exactly where I'm going to be, but I'm optimizing in a way that no matter where I am where the state is going to reveal to me, I'm going to do quite well. Right? So there ways to account for that, again, depending. And what I said about AI is that, you know, again, a supply chain planner, a network operator, they might not be familiar with these formal terms.

So how can we make it more accessible to them? Like they have to define the objective now. They know that they have some risks that they want to account for. You know, how do we sort of help them formulating in a way that, is commonsurate with what they actually need. Yeah. Amy, you look like you want to weigh in and then I'm going to bring us to a close. Well, I guess I was thinking that this is the area you had asked about human agency and this idea of human the lute and not having automatic, but there's also another side of it of how you can

have these systems potentially help not just focus on optimization without thinking about resiliency. You know, if you embed that in the system, then it might have this two-way influence that could have a net benefit. 100%. So, you know, that's something I think that's a really interesting area to pursue. I can tell you both that if I were listening to this

podcast after hearing the discussion, and I want to thank you both, this was really fun. And I learned a lot today, which is great. Yeah, I did too. I didn't have somebody saying polytope on my bingo card this morning when I got up. I would be really curious and maybe even a little bit

antsy to see what is that list of those top problems priorities and then from you because you talked about it and so maybe if at any point you get it I'll post it along with this and then which one does a shy pick up and start working on you know it would be great to have some positive outcome so if we do get the list if you do pick up one of these and make progress you know I would love to report out to people and I'm hoping you know the ties we made today in the discussion we have lead to something like that. That makes the planet better. Excellent.

Thanks a lot of us. Thanks so much for having us. It's a lot of fun. This was really great. Thank you both. Yeah. Thank you. You've been listening to the shape of things to come, a Microsoft Research

podcast. Check out more episodes of the podcast at aka.m.s slash research podcast or on YouTube and major podcast platforms.