The Next Wave of AI in Agriculture

Ranveer Chandra, Vice President at Microsoft, shares his perspective on how artificial intelligence is reshaping agriculture. From the importance of high-quality data and digital infrastructure to the role of collaboration and specialized AI models, he explores what it will take to unlock AI's full potential and how these technologies can help solve some of agriculture's biggest challenges.

Transcript

Copy

 [00:00:00] Artificial intelligence is creating new opportunities across agriculture, from improving decision-making to helping address some of the industry's biggest challenges.

In this episode of AgTech 360, we're joined by Ranveer Chandra, VP of Microsoft, to discuss how AI is shaping the future of agriculture and where he sees the greatest opportunities for innovation in the years ahead.

Thanks for joining me, Ranveer. Yeah, great to be here, Adrian.

So let's start with your story. What, what led you from computer science and technology development into the world of agriculture?

Yeah, this, uh-- as you know, Adrian, my background is in computer science. I did my PhD in computer science, joined Microsoft, was developing a lot of tech. But I did grow up in, in India, spent a lot of time in my grandparents' farm, having seen extreme forms of-- very primitive forms of agriculture, extreme poverty, things like, you know, bullock-driven tractors, hand-based seeding, hand-based weeding, and so on.

And, uh, ever since, uh, like for some time, since [00:01:00] 2010, I was-- it was on my mind that there's no point developing all this tech if it doesn't get to agriculture.

And in 2014, I was able to realize it with a memo saying, "This is what Microsoft should do." I was given a few interns and said, "Try it out." And since then, I have been able to realize a lot of my passion in computer science with its applications in agriculture by working with experts like yourself and other people in the industry, in academia, who understand agriculture much deeper than I do, but by bridging and bringing the best of computer science to this industry.

So it's very fascinating to hear how things work in, within Microsoft, but maybe we'll get into that a little bit more. But first, um, we've been very fortunate at the Plant Sciences Initiative because you've actually been working with us for, for many years, and you've been on our external advisory board, and your mentorship and, um, advice has been really helpful and important for us.

You were, actually at our, our inaugural PSI symposium, which was five [00:02:00] years ago, which I cannot believe I'm saying that. But if you think of that five-year period, I mean, what have you seen in terms of changes in agriculture, particularly linked to, you know, the use of technology and, and AI?

Yeah. I remember, Adrian, when I had visited, uh, you five years back, uh, just around COVID, right around that time, and the fact that you have PSI at NC State, and I'm very fortunate to be part of the, uh, the advisory board. NC State is one of the leaders in, in agriculture. They have very strong engineering team.

It's like bringing together all the experts and to shape such an important agenda around, uh, around plant science, it's just, uh, amazing to see. In fact, I was very thrilled to go there in person, uh, for, uh, for the inaugural event. And since then, things have changed a lot. In fact,, earlier at that time, I remember things had changed a lot since the time I started working on this in twenty fourteen.

And the last five years, we've seen everything [00:03:00] from, you know, generative AI with, uh, the post-ChatGPT moment, which has transformed every industry, and similarly it is transforming agriculture as well. So yeah, this industry is just going through, like the technology sector is going through a lot of change.

Back then, we talked a lot about satellite data, about using weather predictions and how you could combine them to make forecasts. Now we've taken it to another level where, you know, earlier you would still write an app for it or you would, uh, go and, you know, have a separate tool to make those predictions. Now all of that is just a chat away. You just ask a question and you could get an answer with all these GPUs, with cloud compute and everything else , in the background delivering you that, that insight. So the entire space of technology has seen a massive, massive, uh, transformation with AI, and technology and agriculture is going through the same thing.

Of course, there are a lot of challenges remain, that's what we're here to talk about. But I think, uh, the [00:04:00] technology sector is seeing a lot of transformation for agriculture.

So can you help us a little bit with some terminology and some definitions? So, you know, we've heard through numerous guests we've had in the series that AI has tremendous potential in agriculture, but the term AI in itself means different things to different people.

So how do you yourself define AI in the context of agriculture? Yeah, great question, uh, Adrian. So AI is, uh... , this term was coined back in the '50s, uh, and it's since then evolved. It's gone through its ups and downs, and now it's an up again. And it's evolved. It's essentially about how do you get computers to think like a human.

Uh, and this, in the context of, uh, agriculture itself, it's evolved, over, over time too. Just to say a little bit about what is AI, right? AI is ... people started with the broader definition of AI, then they translated it to something called machine learning. Essentially, it's about [00:05:00] if I have huge amounts of data, can I, from that data, build a model to predict what the next data points are going to be?

In some sense, you can think of it similar to, you know, if I give you a bunch of points and I say, "What's the next point?" You could try to put them in a line and say, the next point will follow that line. It's the same thing that's happening for LLMs as well. It's about what if you had all the words i- on the internet, if someone just learnt everything and then said, "If I say a few words, what are the next set of words that are going to come?"

That's at a very, very high level what AI is doing. It's building that model here based on all the data on the internet. You're talking of trillions of data points that are used to actually construct, uh, that, that model to make these predictions. And in the context of agriculture, you know, earlier we looked at various data points to make these predictions.

We looked at satellite data. From satellite data, for example, if I give you a satellite image, what is the crop [00:06:00] stress? Where are the pests? Where is my nitrogen level down? Now, how do you input it from an image? From an image, you need to translate it to nitrogen use efficiency and nitrogen in soil.

How do you make those transformations? You'll get a few samples, you'll use it to build a model and make those predictions. So if I'm able to, say, take an image from North Carolina, label it, mark it, and then use it to train the model and then make predictions, if I give an image in Iowa, what's the state of the crop there?

This is what could be done with-- can be done with AI, right? And there has been a lot of work with remote sensing and all of that. Now, the new opportunity is to go use those, all the work that has happened. But, one of the challenges had been a few things. One is, how do you make that satellite image accessible to any grower anywhere?

What kind of questions am I even supposed to ask? Now, this would hinder a lot of, uh, you know, technology adoption in agriculture, and especially if you look at smallholder farmers. They are [00:07:00] not, you know... They, they're just talking. They don't have anything to see. What kind of questions do they ask? With large language models, you can just ask a question and it'll come back with a response.

Behind the scenes, it will take whatever question you asked- Translate it to how should I be creating the satellite image and come back with a response. This is one transformation that's happening in agriculture with generative AI. Every stakeholder in agriculture, from farmers to input companies, to data scientists, to people working in the supply chain, now information is so easily accessible because all you need to do is just ask a question.

It's no longer about I have to install a hundred apps. I remember back then when we had come, one of the questions was, for everything, for planning, for planting, I have a different app from someone else, and how do I even go across all of them? Now, a lot of that can get aggregated by just a central place.

You ask a question, and you'll get a response. That's one place where AI is really taking off. The other place is, just extending it from language to images. Like, you know, how do you [00:08:00] take images and add insights on top of that? Uh, to the, things like if I take an image of a, of a, of a disease to if I take an image of, of a plant to what's happening.

To even taking it a step further to what will come. This is not where things are right now. Existing models, these large language models that you're looking at, they are book smart. They are not world smart. The next phase of AI is going to be world smart. What that means is these robots, these drones, they'll know what to do next.

Uh, and that is going to drive a big, big innovation in agriculture. And I think when all these are combined with image, text, world models, all of them, when they come together, you can imagine a world where every, every stakeholder, I'm purposefully not just saying farmers, of course farmers would benefit, but every stakeholder, be it a scientist, be it an entrepreneur, be it, uh someone in the supply chain in food processing, everyone is going to benefit by just the amount of efficiency gains they are going to get with artificial intelligence.

[00:09:00] So before we get too much further can we talk about some of the practicalities around- Yeah ... AI and agriculture? And, one theme again that we've heard throughout the series is that AI is only really good as the data and infrastructure behind it. So what is required to support AI-enabled agriculture when you're, you're generating data and collecting data in very remote parts of the world and, of, of the country, for instance?

That's the thing. You know, when we started this work at Microsoft, that was the key insight that drove the entire work we did, which is that, you know, uh, agriculture is data sparse. So if you remember, Adrian, the number, the number one things we started looking at was how do we get data from remote sites?

And the challenges were that there's no single data stream that can give us that data about what's happening in a farm. Like for example, if someone asks a question, "What is the nitrogen level throughout the farm?" Or, "What is the soil moisture level six inches below the surface?" There is no way to get those numbers.

I can try to put a few [00:10:00] sen-sensors, but sensors are only giving you the, the, uh, the data at at that location, not broadly throughout the farm. You can get satellite imagery, but that's only giving things at the surface and very coarse granularity. So agriculture has that challenge. The question you asked, Adrian, is a very good one.

How do we make AI more ready, more useful to agriculture, given that we don't have as much data? And the fact coupled with that is that this data actually varies quite a lot across location to location, like even soils, and you're the expert in this, Adrian. But some of my learnings is that soils we define as the physical, chemical, and biological properties of soil, and these things, biological most people, like, we don't understand that well yet.

Physical and chemical, there is more understanding. But these things vary so much in different parts of the world that even if I get information from one farm, it doesn't easily translate to another farm. This is not well understood because you don't have enough data to start building these kind of data set- to, to build these kind of models.[00:11:00] 

So the number one problem we need to address is that of good data. Right now, and the challenge is right now with, with, with generative AI, you could start asking these, uh, these models questions about agriculture. It'll do its best to answer it. The problem is without good label data, a lot of times it is hallucinating.

And the bigger problem is that because people don't have the real data, they don't even know if it's hallucinating or not. And this is different from a lot of other domains that we work in, where there is data you can actually tell whether it's right or wrong. Here, it, it's just operating in that space without less data.

So I think the number one effort we need to do is to make sure that we get good data and we have good simulations, because it won't always be possible to get good data. So I think this is where working with scientists who have in the past worked a lot on soils, like, you know, all the different models that people have worked on or crop growth models that people have worked on, those need to be taken into account to create these realistic simulations to generate data that is more [00:12:00] believable, which can be used to train these models.

All the way from language models to vision models to world models, we need to have a really good data set. And that data set, it's, it's not available right now. And we talk about available not just in the emerging markets. Even if I look at here in the US, that data is not readily available.

And that's an opportunity for land-grant universities where PSI is already showing leadership here, uh, with what you're doing, Adrian, and that's an area where we need more and more innovation.

And, and if I could ask another question that really is based on a very practical issue, is that, you know, you're integrating all of these, these data points from multiple sources like you've described, but you're doing it in an environment that's perhaps in a very rural area without fiber.

H-how do you translate data that you're collecting into something that's actionable in real time on a farm. I mean, what are the, the practical challenges around that, and what maybe are you working on to kind of overcome that type of thing?[00:13:00] 

Yeah. Bringing this data together

It's such a fundamental issue, right? Yeah, you could, you know, you could use satellite connectivity maybe on one tractor, but that's not giving you everything that you need. You need connectivity throughout the farm, and, uh, you need lots of inexpensive sensors. You don't want this thing to become super expensive.

So cost, connectivity. Battery is also an issue. You need these things to be working all the time. And you need to make it so that, behind the scenes, you're not consuming tokens like the way existing models are because that will make it infeasible. The, the ROI won't be there. You have to think of not just developing the model, but what's the return of investment of building that model.

And so you have to bring down the cost of all of this. And we've been looking at various approaches over the years, all the way from doing edge compute so that I'm doing things on the edge. I don't have to send everything to the cloud. If I have GPUs, like one of these new PCs that we announced last week with, a PC with N- with Nvidia, which you could be doing things at the edge like [00:14:00] in the cloud, being able to do using smaller models as opposed to larger models.

One of the other things we looked at was combining large models with small models, which makes the price go down. The other interesting thing is you could actually, rather than using a huge model, you can actually use a much smaller model, this large language model, smaller language model, which is still big, but it's not as big as the big models.

But at less than 10X the cost of these large models, if you fine-tune it for agriculture, like an agriculture LLM, you could make it perform much, much better. It's like a specialization as opposed to a generalized large model. So there are various techniques. We are working on all of these to, to bring, to make, to democratize AI for agriculture.

Any farmer, anyone anywhere on the planet in the agriculture industry should benefit from AI.

So how do companies like Microsoft, um, interact with agriculture? Are you taking these innovations that you're describing directly to farmers, or are you working [00:15:00] through universities or agribusiness or equipment manufacturers?

How do you, you know- Mm-hmm... connect yourself with the agricultural sector?

Yeah. So, you know, we work with, uh, Microsoft, we work with every organization out there. Nearly every organization on the planet uses our products, and we work mostly in the agriculture space. We work with every company that farmers interact with.

We typically don't directly work with farmers. Of course, you- they buy our Windows PCs, they'll use some of our products, but mostly for agriculture business operations, we work with our, our partners, like Land O'Lakes, which we talked about last week at the Build Conference, Bayer, every... nearly every agriculture company, uh, with USDA, we work with them too.

We work with universities extensively. But, you know, Microsoft is not an agriculture company. We don't sell to, to farmers, uh, a lot of these insights directly, but we work with a lot of partners who then build solutions and take it to different [00:16:00] stakeholders in the agriculture industry.

So can you talk a little bit about perhaps some of the initiatives that you're working on, things that you're getting excited about that, uh, may mean advancements in, in the type of technology you can provide to farmers, albeit through all of these intermediary kind of companies that you work with?

Yeah. You know, we started, if you look at the history, Adrian, we started with working with a cloud system for agriculture, which was the FarmBeats project- Mm-hmm ... which was about, you know, helping companies gather all the data so that they can build insights on top. Then we did this work on Farm Vibes AI, which was basically we open source a lot of AI on top of the data.

Once you get the data, how do you translate it to insights? It's, uh, in GitHub, it's an open source. We've worked with several partners. Most recently there was work we did in India, which in, uh, in Baramati, in Maharashtra, this is a drought-prone region where we showed working with, Agriculture Development Trust and, uh, Oxford University and a partner there, Click to Cloud, where [00:17:00] they've deployed it for 10,000 plus farmers where they are benefiting from, from AI, where they showed that how they could, uh, double the yield of sugarcane.

The new thing that I'm currently working on is a lot in the AI space. Our team is going deep in AI, building new AI capabilities on post-training models.

Like the thing I talked about of how do you create... What does AI for agriculture mean? If I think of anyone in the agriculture industry, how can we help them incorporate their own knowledge, their own knowhow of what, their workflows into AI so that they get the most of AI at the least cost?

This goes back to this thing of, we need to bring down cost to show the ROI for, for AI. So all the way from, , if I give a question, how does it come up with an accurate answer at the lowest price point? So we actually change the model weights- Like, you know, it becomes your own very personal model that you can then create for your institute, for your [00:18:00] farm, for your, uh, uh, for your organization.

The new thing we are looking at, which is what the post-training... What post-training does is it's going to modify these weights so that the predictions are very personalized to you and your data and your knowledge, and that's the capability that we're building. And the new thing, if just getting a little bit deeper, hopefully people on the podcast will...

Who already know it's great, but people who don't know will learn the new term on reinforcement learning. If you're using AI, you know when you ask it a question, people who saw the transformation from the earlier ChatGPT days to thinking models, they show a reasoning.

You click and say, "What did it think? How did it come up with the answer?" That's done using reinforcement learning. This thinking didn't happen before o1 came out in OpenAI. Before that, you'd ask a question, it'd come up with a response. This thinking is because of reinforcement learning. The way reinforcement learning works is I'll give a question to the large model.

It'll come up with a response. What we do is we grade [00:19:00] the response. We say how good is the response? If the response is great, it's good. But if it isn't, what we do is we do a reverse loop. We compute a loss. If you know the good answer, you know where it was, what is the delta, and we use that to change all the weights.

And we keep doing this over and over and over again. So it becomes yours if you train it on the way you think.

The results are better, but it also reduces cost. You can get a very small model to behave like a large model and actually better on your data. So it becomes... It talks like you, it thinks like you, and it works like you.

That is fascinating. Let's move to crystal ball time. So, um, I'd love you to kind of describe what you see as the big, new opportunities and technological advancements in agriculture.

We've seen obviously so many things, but what are the kind of things that you believe that AI will enable for us to do tomorrow that we can't do today?

This is something that I [00:20:00] think a lot about, about the future of agriculture with AI. What is the future farm with AI? What's the future supply chain with AI?

If you ask me for the holy grail that I would like to achieve, a few things. Uh, you know, one is for every farmer, with AI, they can be more profitable. They should be able to increase productivity using this. They should be able to reduce costs because I'm not putting in more inputs and all of that, and it's also more, better farming, more sustainable for, uh, agriculture as well.

This is the-- what every farmer should be able to do with AI. And everyone in the agriculture and food industry will be more efficient once they start using AI. We, we have multiple use cases. We've actually done this for every stage of farming.

How is this going to transform? How much improvement should people be looking at? How much fewer, how much fewer inputs would they need? How much more, uh, agile they can make the system? I can figure out, uh, where should I [00:21:00] sell, uh, my, my product? When should I sell? How long should I store? A lot of these things, not just during farming, but post-farming as well.

Post, post-harvest, you can take a lot of decisions, uh, using AI. This is one thing which I think AI will enable. And every year, I think every stakeholder in the agriculture industry, I would challenge them to think about it. Like, to see over the last year, how much did they improve on all the metrics, basically around profitability, around productivity, all of that, using AI.

The part that, you know, someone needs to work on for agriculture, which is completely missing, is that in agriculture, the-- it's-- the entire value chain is very siloed. Like, for example, the data at the input companies is not visible to the data, uh, what's happening in the farm, to what's happening post-harvest, to what the consumers want.

This entire thing is very disaggregated. And this is why if someone can close the loop, end-to-end loop, it's [00:22:00] going to lead to significant efficiencies throughout the value chain. Someone will do it.

Because everyone in the industry is going to benefit. A lot of wastage, a lot of delays, uh, the supply chain issues will go away. If someone can bring that data across, because once you bring the data across, , I'm very confident that AI will uncover a lot of, uh, efficiency gains throughout, uh, throughout the value chain

So you've kind of anticipated my last question, which was around the barriers that exist that would need to be overcome for AI to be truly effective.

But it does provoke another question, which is, you know, AI is still a controversial subject. Um, you know, what, what do you feel, from a tech provider perspective needs to be educated in the mind of the public to really kind of accept and, and see AI for the opportunity it is versus perhaps as something that is more threatening.

Yeah, no, this is a good question, Adrian. You know, in the past, whenever new technologies have come up and like even if you [00:23:00] think of PCs came, the internet came, and, uh, then there was mobile, social, then there is now with, uh, with AI. Each of these has led to changes in what we do. Like with the internet, now there is Uber.

Initially there was a lot of resistance, but we all see a lot of benefits of having, uh, uh, like, you know, Uber, Lyft, and these kind of, car providing services. So I think even right now with AI, we're still in, like the middle innings, um, in some sense of, of AI with, uh, right now given where all the advances are.

We need to make sure that AI is being developed with what people want in mind. It's not just unfettered. It's more like we can guide AI to do what it should do. So when I talk to people about AI, I give them two things, two pieces of advice.

One is, you know, all of us should just start using AI. It's not just to be more efficient, it's also to be safe. A lot of time when people are on the internet, they're interacting with [00:24:00] agents in a lot of cases. They need to know what this could... what, what kind of responses you're getting. Of course, it would also benefit a lot, like the way we're-- the way you use AI, the way I use AI, we use it to make ourselves more productive, to give us more time to do, uh, to do other things.

So I think that's one. Uh, but-- and the other thing I would say is that just because you know AI today doesn't mean that, you know AI tomorrow. This field is changing at such a fast pace. For most of us, most of the people in, uh, in agriculture and food, I would say just be on top of new tools that are getting released and just try them out.

Just keep, keep upskilling yourself.

If you ask me where things will go down the line, I think AI will become a prerequisite. Like, you know, most of us need to know how to use stats, use probabilities. We all will need to know how to use AI. It'll become a tool in our, in our set of skills that will help us achieve a lot more.

Everyone. In fact, right now, uh, if you ask me and [00:25:00] my team, what do we do with coding? Have we stopped hiring people? No, my team has grown quite a bit in the last year, and the fact is we are still hiring people, but we are also generating a lot more code. The amount of code that, that AI writes is a lot more.

So I'm looking for people who can use AI to be really more productive, and that's going to be the new definition. Like imagine hiring someone in the team who did not know math before math was there, or who did not know how to write simple code, to people who can write code, right? So there is a difference.

You would hire people who can be more productive, and this just becomes a skill set that everyone should pick up. And the h- And the good thing is you don't need a PhD to pick up AI. Anyone can, and use AI to teach yourself AI. So I would just say to, to remove the inhibition. I think this is going to create a lot more opportunity for everyone around us.

I think humans are going to uplevel their ambition of what can be achieved much faster. Like the number of problems that we have in agriculture, Adrian. You know, [00:26:00] you've been in the industry for a long time. I've been around for some time. And when we look at where, uh, like the number of open problems, we haven't made progress as fast as we should.

If we could make that faster progress using AI, there's so many people in the world that would benefit. And so it's not because... I don't think jobs will go down. We need to solve that problem. It's just that we need to solve them faster, and AI will help. Uh, and pe- it should create more jobs because we need to solve them.

There is a material need, an economic need to, to address those problems.

Ranveer, with that last piece of very inspirational advice, I wanna say a huge thank you for being on the podcast. It's been great to catch up with you and listen to some of the stuff you're working on and what might come down next, uh, the technology pathway to farmers across the world.

Thank you. Thank you, Adrian. Really fun being here.