Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
Using focused AI to help small farmers and reduce food insecurity
102 Plays12 days ago
What if narrow #AI, rather than imagined AGI through scaling will be what changes the world? In some places, that’s already happening.
El Mahdi Aboulmanadel founded DeepLeaf after watching smallholder farmers in Morocco misdiagnose crop disease because three distinct conditions can look identical to the human eye. Wrong diagnosis, wrong treatment, chemical residue on food.
Best case scenario? Export crops rejected at customs.
Worst case scenario? Food scarcity for communities that can’t afford it.
DeepLeaf's answer is deliberate focus: one problem, field-validated data, models trained on hyperspectral and RGB image pairs across 57 crops. The accuracy doesn't come from scale. It comes from specificity. Fine-tuned continuously on new field data. The result is less compute, faster iteration, and outcomes closer to the ground truth.
DeepLeaf has both cloud inference for large or multi-crop operations and lightweight edge models downloaded per crop for farmers running on Android phones in areas with no connectivity. The architecture fits the user, not the other way around.
We get into economic potential for farmers, and of course, the effects of the war in Iran.
This episode is about what new AI perspectives than the ones taking up all the oxygen in the West. This is technology that’s built for communities that Silicon Valley usually ignores.
Recommended
Transcript
Challenges in Identifying Tomato Diseases
00:00:00
Speaker
And to give you one simple example that anyone listening to us can try in his house to prove that this is real, just type in Google, for example, tomato brown regoes virus, and also type tomato magnesium deficiency, and also type tomato yellow leaf curl virus.
00:00:22
Speaker
These three are different problems. Two of them are viruses, one of them is a deficiency. But when you Google them and you see the images, they look exactly the same.
00:00:34
Speaker
So put yourself in the shoes of the agronomist or the farmer. of the farmer You won't be able to make difference. When you don't make the difference, you end up choosing the wrong disease.
00:00:45
Speaker
By choosing the wrong disease, you choose the wrong treatment.
Introduction to Deep Leaf and its Mission
00:00:56
Speaker
Hey, this is Bare Knuckles and Brass Tacks, the tech podcast about humans. I'm George K. And normally joined by George A., but we're tight on time to record the intro and the outro, so I'll run it solo, but the interview does include us both.
00:01:12
Speaker
So this week our guest is El Khmadi Aboul-Hmanadel, who is the founder CEO of Deep Leaf, an AI startup out of Morocco that is using narrowly focused AI to help small farmers identify crop diseases using sometimes just the compute on their local phones, even without Internet.
00:01:36
Speaker
We wanted to talk to him because in the West, we hear mostly about AGI and mostly about LLMs. This is the discourse that is being controlled by Silicon Valley and is basically taking up all the oxygen in the room.
00:01:53
Speaker
Meanwhile, around the world, there are entrepreneurs using machine learning, deep learning, and AI techniques to develop applications that can materially impact their are communities, whether that is in real economic terms or in food security terms.
00:02:10
Speaker
And so we really wanted to have someone on to have that conversation.
The Role of Expert Annotation in AI
00:02:13
Speaker
And we get into a lot of that and we also dive into the tech, how they are training, what kind of data they're using, the high quality expert annotation and validation.
00:02:25
Speaker
These are all different modes that are still popular throughout the world. It's just not being discussed over here or at least not in the proportion we think that it could be. So without further ado, here is the interview.
Impact of Deep Leaf on Smallholder Farmers
00:02:43
Speaker
fadi abli banel How are you, sir? Welcome to the show. I'm doing well. Thank you very much for having me here. Yes. So let's start in the obvious place. Why don't you tell our listeners a little bit about Deep Leaf, how it started and what you're doing today?
00:03:00
Speaker
ah So I'm an AI engineer and I started Deep Leaf three years ago, right after graduation, because I lived with farmers, smallholders specifically in Morocco, North Africa, ah because my father was a veterinarian and I've seen their struggle trying to detect pests, disease and nutrition deficiencies that are making them lose their crops.
00:03:23
Speaker
And to give you a general overview of this problem, the FAO farm ah the fau Food Agriculture Organization has declared that the world globally lose $300 billion dollars due to these problems specifically every year.
00:03:38
Speaker
So I built Deep Leaf in order to identify these problems early with a simple solution for smallholders, which is a mobile app where they take a picture and AI analyze it and gives them the disease name and treatments from nearby agro-dealers.
00:03:53
Speaker
And today we are helping up to 5 million smallholders from Africa, Middle East, to Europe, and we're trying to expand into other continents. Yes, absolutely brilliant. Thank you for that summary. And I think, you know, for most of our listeners, if you've traveled through the American countryside or hell, even the Canadian one, they are used to these large scale industrial farms run by maybe five, 10 people at most. Um,
00:04:19
Speaker
But so when we say smallholders, for the sake of this podcast, we are talking about small single owner farms, which do make up the vast majority of agriculture on the planet.
Focus on Specialized Agricultural Solutions
00:04:31
Speaker
So the dominant story in the West, as it relates to AI, right, is scale, the so-called scaling laws, which... I will point out, are not actually laws like gravity.
00:04:45
Speaker
But this idea that if you ingest enough data, models start to generalize, et cetera, et cetera. This has given us generative AI LLMs. However, your application seems to be at the other end of the spectrum. Can you talk a little bit about that? Because I know you and I connected initially because I was posting about ah smaller, higher quality focused data sets.
00:05:11
Speaker
and i At DeepLief, we also started like a general solution, but then we find out that in order to be a leader in agritech, you have to narrow down your approach. You have to focus on one specific problem in agriculture, especially coming from Africa, where resources are very limited.
00:05:29
Speaker
Compute is very limited and there are many competitors because as you may heard, Africa has the biggest number of youth compared to other continents. So in order to be a leader in the space, we focused on phytosanitary, which is ah disease pest management and also pesticides and fertilizers.
00:05:49
Speaker
So we built a computer vision AI model trained on 1000 anomalies labeled by experts across 57 crops.
Adapting Tech for Smallholders
00:05:58
Speaker
And we've been able to do it across 57 crops, which is the biggest coverage now in the world compared to competition because of this focus.
00:06:06
Speaker
Otherwise, we would be stuck in one or three crops. So um once once we did it, we find out that it's not easy again to implement this technology in African countries because of other problems. First thing is tech literacy, especially for smallholders.
00:06:25
Speaker
They are not used to use other apps besides WhatsApp, Facebook. So you have to build some UI that is a bit similar ah because they are familiar with the experience there. That's why we created it a chatbot that looks exactly like WhatsApp.
00:06:41
Speaker
So instead of texting your friend, you're texting an AI agent with voice, with images. And it brings you local dealers from nearby locations showing you the treatment and generating for you also advisory messages on how much you should use, where to spray, and to stay compliant.
00:07:01
Speaker
Wonderful. Wow. i That's pretty incredible. So i'll I'll drop back a little bit just to go on the the core fundamentals of the tech. Just, you know, took a look at you guys as well for the show. And it was really impressive well the work you guys have put in.
00:07:19
Speaker
You know, you built a model that runs offline um on a low-end smartphone. You know, so running inference on the edge, more specifically with low-tier hardware, that's a pretty massive technical hurdle.
00:07:30
Speaker
Behind just, you know, the model quantization, how are you handling the trade-off between latency and accuracy? You know, more specifically, are you using, are you using, know, a custom kernel or running specific printing architecture to ensure that, you know, hundred dollar Android phone doesn't overheat while trying to run a computer vision task in like 40 degrees Celsius heat?
AI Model Training and Customization
00:07:53
Speaker
So ah thanks for mentioning the specific parts because this is where we, where our R&D is focused now is bringing large models from the cloud to the edge. And since we don't have a lot of computes available in Africa, we've got the support of European countries, ah specifically Italy.
00:08:13
Speaker
They gave us access through UNDP AI Hub ah to Cineca, which is a large data center, the second largest data center in Europe actually. ah with HPC computers, hyper performance. So to give you an idea about ah how big it is and how important, ah it's 100,000 hours of GPU and each GPU in the cluster is 10,000 gigabytes of visual memory.
00:08:44
Speaker
ah You can compare it with 10 times the GPUs that are used now to train large language models. So we use that ah infrastructure ah to train our algorithm. Deep Leaf's algorithm is called Deep Leaf Net.
00:09:00
Speaker
It's a custom kernel, but not 100%. It has a backbone of vision transformers, which is a well-known algorithm used by everyone.
00:09:10
Speaker
So ah we needed to customize it just to make it specific on agriculture. We added pre-processing layers that allow us to remove the background, the noise, keep just the leaf or the fruit by segmenting it.
00:09:25
Speaker
ah Also, there is a big problem of processing data collection in agriculture. For example, an image of a tomato in Morocco is not the same as an image of a tomato in Mexico or in the US. So we had to create an adaptation layer that makes sure the image is normalized.
00:09:43
Speaker
So basically we tweak the contrast and luminosity in multiple other elements based on the location, allowing the image to be treated the same way as if as other locations.
00:09:57
Speaker
And, uh, yeah. And finally, there's a minimal, wait, so there's a minimal margin error when matching through assumptions. Like, like that, that scan detect is pretty, it's, it's asking me to take a bit of a leap.
00:10:10
Speaker
It's, it sounds very novel. ah Exactly. There is a big gap that we are filling. And yeah many people talk about AI hallucination.
00:10:21
Speaker
Well, our AI hallucinates as well. We can't say it doesn't, but we control its hallucination through ground truth evaluation. I will explain to you how how it works. When you train an AI in your computer, the metrics that you get at the end Accuracy, F1 score and all of those metrics are only ah digital metrics that gives you an idea about how your AI performs in synthetic data.
00:10:49
Speaker
For example, I will train the AI on 100 images and I will leave 30 images out of the dataset so he doesn't see it and then I will test it on those data. Yeah.
00:11:01
Speaker
Basically, that's how those metrics work. But in agriculture, in healthcare, in ah spaces that are very ah sensitive, where you cannot make mistakes, you have to do also ah evaluation in the ground, in the field.
00:11:17
Speaker
So what we do is with the collaboration of universities, with food security programs, agronomists and researchers, we go to the field, whether it's a greenhouse or an open field, and we deploy three to five people.
00:11:32
Speaker
They take pictures and these are experts who already know how to identify the disease with their own eyes. And they cross-check every AI recommendation. Is it matching the expert detection?
00:11:46
Speaker
This is the expert annotation part Yeah, it's not just annotation, but also evaluation to evaluate the output of the AI, whether it's correct or not. And to confirm, because many of these agronomists, although they are experts for years, they still make mistakes. We take those samples to the lab to do further analysis, to know whether the AI maybe is even better than that agronomist in some cases. And that's what we found.
00:12:13
Speaker
And
Risks of Misidentifying Plant Diseases
00:12:14
Speaker
to give you ah one simple example that anyone listening to us can try in his house to prove that this is real, just type in Google, for example, ah tomato brown regoes virus, and also type tomato magnesium deficiency, and also type tomato yellow leaf curl virus.
00:12:36
Speaker
These three are different problems. Two of them are viruses. One of them is a deficiency. But when you Google them and you see the images, they look exactly the same.
00:12:47
Speaker
So put yourself in the shoes of the agronomist or the farmer. You won't be able to make difference. When you don't make the difference, you end up choosing the wrong disease.
00:12:59
Speaker
By choosing the wrong disease, you choose the wrong treatment. And when you use the wrong treatment, which is most of the time chemical, us consumers who go to the grocery shop to buy, we find residues of chemicals and we consume them and we hurt our health.
00:13:15
Speaker
So this is a like a concrete example of why this technology is super important for the world. Yeah, absolutely. Thank you for that. And also I'm intrigued You sort of had the startup mentality of I have to
Expanding to New Markets and Partnerships
00:13:31
Speaker
focus, right? So instead of like general purpose agriculture chatbot, let's focus on disease. But you said that that focus also allowed you paradoxically ah level of breadth, right? You said 57 crops instead of, you know, say three or something like that. Can you talk a little bit about that and like...
00:13:52
Speaker
I guess, what kind of coverage you're seeing? You've mentioned North Africa, but I think you're covering in other parts of the world as well, right? Yeah, we are now exploring other markets like ah Turkey so we can enter Asia, Bangladesh. We're also exploring Mexico with one of the fertilizer companies.
00:14:10
Speaker
And once we have a customer secured in one of these spots, we are able to ah further expand into bigger countries that have giga farms as well.
00:14:22
Speaker
ah But ah the way we do it now is via API integration. So we are trying to act like a technology provider to local partners, local agri-tech companies and startups instead to stay the interface, instead instead of being the interface everywhere.
00:14:43
Speaker
For example, if you go to Croatia in Europe, you will not find deep leaf. you will find AgriV, which is a Croatian startup, 100% Croatian. They built their own chatbots, their own farm management system and so on.
00:14:58
Speaker
But yesterday in Jitex Africa, that happened in Marrakesh, which is a big tech event. We just signed a technology partnership where DeepLeaf provides them with disease detection API through their chatbots. Beautiful. So farmers there use our technology in the background.
00:15:14
Speaker
So this is ah a pivot that we have done a year ago ah because it's very hard for a small company to be able to market its own interface everywhere in the world.
00:15:26
Speaker
So instead of trying to do that, we integrate into existing platforms. And I hope someday one of the big ag techs in US as well be using DeepLeaf as a disease detection API. That was brilliant. Yes, and I think...
00:15:43
Speaker
More and more with AI applications and startups, it feels like the power isn't in the interface, the power or even like the software per se. It's in the ability to transact at the data layer. Right. And so in this case, you have Deep Leaf Net, which is going to be a value to other ah agricultural startups. Yeah,
Advanced Image Techniques for AI
00:16:06
Speaker
absolutely. Makes perfect sense.
00:16:09
Speaker
Yeah, that's a... I'm like, okay. Because I had to visualize in my head too. i was like, okay, cool. So this makes sense on paper. And this is...
00:16:20
Speaker
I've never seen anyone else try it quite like this. I think it's novel. I'm very fascinated. You know, you've talked about ah training off proprietary field validated data across specific crops and conditions, right? So every LLM um developer is currently fighting, quote, data decay or model collapse from scraping the open web.
00:16:40
Speaker
you're doing the opposite, right? You're generating high fidelity field validated data. From a strategic security perspective, i'm interested in your data pipeline integrity. That's something that always have to deal with as well.
00:16:52
Speaker
How do you prevent undesired noise from a field like, you know, poor lighting or a varied camera sensor from poisoning the weights of your specialized models? So ah first, the the way DeepLief input is built is not based on not just ah RGB images.
00:17:11
Speaker
The data set of DeepLief is a pair of RGB and hyperspectral images using a specific camera. ah But since we cannot sell that camera to farmers because it's very expensive, as you know, hyperspectral imaging is is not... ah It costs more than $1,000 each. Right. so And it's not scalable. So what we did is ah we created an AI that's trained on the pair of a hyperspectral image in near infrared, in visible red, and other spectrums, but also RGB.
00:17:49
Speaker
And then this AI have to be able to detect the disease in the RGB, only after finishing the training. ah I will try to give you a different example so you can understand how it works.
00:18:06
Speaker
So imagine you are used to ah see someone do some specific activity and then ah later on ah there is a wall beside you and him, ah but you still can hear his voice.
00:18:23
Speaker
you will be able to, based on your previous experience, which is the AI training, you will be able to still know what is he doing just based on his voice. So it's the same theory behind our ah solution, our the AI inputs that we provide to our algorithm.
00:18:41
Speaker
So we make the AI look at two elements that are similar, same disease, but different spectrums. This one shows you early signs. This one shows you what is visible to the human eye and what the camera of and any phone is capable of.
00:18:56
Speaker
And then we remove this and we only give you this and we tell you, can you tell us what is that disease? And then you will be able to have a better ah detection, earlier detection and works even in 200 pixel images ah coming from feature phones.
00:19:14
Speaker
ah This is how we fixed the inputs of our AI. It was trained on like pairs of images, allowing the AI to be able to work also in very limited ah camera.
Flexibility in AI Deployment
00:19:27
Speaker
and But this is in terms of input. Now in terms of the resources that ai requires to be running in a mobile phone, That was the challenging part. You mentioned quantization, you we did that as well.
00:19:42
Speaker
ah But to be honest, now most of ah users are still using the cloud-based solution. And users who use ah offline edge models our users who are who have like three crops or four crops.
00:19:59
Speaker
ah So you download the crop that you need. Ah, okay. Basically. Yeah. But those who have large farms or multiple ah crops still use the cloud-based solution.
00:20:12
Speaker
Yeah. I dig that flexibility. Yeah. Yeah.
00:20:18
Speaker
After the break, we talk about the real economic impact to small holders and small farmers with respect to getting it right when it comes to identifying agricultural diseases. And it goes a lot deeper than you might think. And of course, we have to touch on food security and the geopolitical tensions currently roiling the Middle East and snarling agricultural supply chains.
00:20:48
Speaker
Let's paint a picture for our listeners here. Again, most of our listeners are in the West. We do have some in Africa and in Europe, but
00:20:59
Speaker
down to the ground level, you've mentioned ah when people, by just looking at something, pick the wrong disorder, right? They pick the wrong solution.
00:21:14
Speaker
What does that mean in economic terms for smallholder farmers? Okay. ah So, you know, this question. What are the consequences, I guess? What is the the solution that you're trying to solve ostensibly? If we were looking at Silicon Valley, they would say like, okay, disease detection. for But I think when we talk about companies in that way, we're ignoring like the downstream impact on the actual ah user or in this case, farm owner.
00:21:45
Speaker
Yeah, I understand. i will answer you first for the smallholders, the impacts on their economics, but I will also mention how it's relevant for large agribusinesses in Jenga farms.
00:21:57
Speaker
So and it's a food security matter, not just a matter of economics. So when you detect the wrong disease and you use the wrong treatment, ah you don't just ah harm the consumer, but there is an entire regulation that you are no longer liable to it.
00:22:15
Speaker
So you may face some legal charges after using the wrong treatment.
Regulatory and Compliance Challenges
00:22:21
Speaker
And the the problem is not just using the wrong treatment sometimes. It's also using expired treatments because they expire.
00:22:30
Speaker
Most of the times, especially in Africa where people doesn't know sometimes how to read, especially small older farmers. The agro dealer tricks them. He sells them expired treatments because they don't know.
00:22:42
Speaker
And sometimes they sell them the wrong treatment for the wrong crop. And besides this, there is also usage. If you don't know ah the exact dose that you should be applying and when, that's also non-compliance.
00:22:57
Speaker
And there are strong regulations, especially in Europe and also in the United States that tells you ah here is how much dosage you should use for each active ingredient.
00:23:09
Speaker
And this information is available for the public. You can just go through the food security program that you have in any country online and you will find a list of active ingredients and how much you should use.
00:23:22
Speaker
So large agribusinesses from the West and also from Europe and so on have access to this because they can read it. But from the global South, illiterate farmers and people with different languages cannot access it. And even if they do, they won't be able to understand what is said there.
00:23:43
Speaker
So what DeepLief does in this case is providing them with adaptation layer in their own language via a chatbot. Every disease detection is connected with an authorized treatment from the government's authorized active ingredient with the exact dose that should be done that time.
00:24:03
Speaker
And to reflect ah to you the impact on the economics, farmers, if they don't use this solution, if they fell in this problem of using the wrong treatment, they will be stuck in the local markets and in the unstructured markets.
00:24:19
Speaker
Like in Africa, we have many weekly markets that are unstructured. Governments have no control over them. if you are selling in those markets, you are selling at a very low price and you have very low trust from the consumer.
00:24:35
Speaker
But when you have the compliance, when you're using the right treatment with the right dose and so on, you have certification that allows you to sell in better markets, including supermarkets, For example, we have Margen, Carrefour, and many big chains. And you can also be exporting. You will be eligible to export to European Union, to the American continents, and also to Asia because you are becoming compliant and you have certifications.
00:25:04
Speaker
ah For example, here in Morocco, tomato, cherry tomato specifically, is one of the biggest exported crops. We are the second in the world in 2026 in exporting cherry tomatoes to Europe.
00:25:18
Speaker
And imagine they make $1 billion dollars every year of this exportation. It seems big as a revenue, but they lose twice that number every year, just because most of those tomatoes doesn't follow the compliancy and they get rejected by the customs when they arrive.
00:25:38
Speaker
ah So this is the economic impact beyond the money that you waste on the wrong inputs. This is also on the rejection and
Geopolitical Impacts of Agricultural Practices
00:25:49
Speaker
so on. But there is also another impact, which is now very important, especially with the tensions and the but political problems that the world is having.
00:25:58
Speaker
So ah I will give you an example, which is sometime when, oh for example, Moroccan strawberries was exported to Spain.
00:26:09
Speaker
ah Some random people said that this is not just ah a mistake, but this was on purpose and it may have caused diplomatic problems.
00:26:20
Speaker
So you see, it's very sensitive subject because food is very important for the world. So it's not just economic. That's actually a perfect segue into what I was going to ask. So like looking at the geopolitical landscape, food security is becoming a data security issue and of itself now.
00:26:40
Speaker
So as Deep Leaf scales, do you see yourself moving towards, we'll say, a multi-model approach? for example, like I could see you easily integrating satellite telemetry or IoT ah soil sensors into your edge models.
00:26:54
Speaker
Like... Is it that or is it the holy grail is simply perfecting the vision only diagnostic on the cheapest possible hardware? like Like, how do you see the scaling taking place and really working through the world of events that are in your region?
00:27:09
Speaker
So we are trying to do what you exactly said, trying to stay on digital ah solutions instead of moving to hardware. But at the same time, if hardware is crucial for some specific ah parts of agriculture, then we don't use do it ourselves, but instead we build the software that can be used by local IoT companies.
00:27:31
Speaker
For example, instead of manufacturing our own IoT devices for irrigation, we will just partner with, ah for example, if we want to go to Tunisia and we want to help them solve drought problems, instead of building IoT sensors on our own, we will just build the orchestrating software behind it and let local company do the manufacturing.
00:27:55
Speaker
ah This way we can be highly scalable. But ah for the data part, ah many governments now are speaking about data sovereignty and they are trying to keep data, especially in healthcare and healthcare,
00:28:12
Speaker
and the agriculture, they are trying to keep it in their country. They don't want it to exit for some political reasons.
Government Use of AI for Agriculture
00:28:20
Speaker
So I can give you an example why.
00:28:23
Speaker
So in Qatar, for example, which is one of our customers, the the government use our solution to provide subsidy ah to the farmers. So ah you know what is subsidy, right? yes yeah Yeah. So when a farmer wants to use inputs or pesticide, he can get subsidy from the government to buy that, to get it for free or just to get a coupon, like 50% or something.
00:28:51
Speaker
And they've been doing it randomly in almost all countries. Even in first, like in Europe, in the West, they also been doing it randomly most of the times. But in Qatar, since they have a lot of money to invest in technology and they have crazy data centers there now, they was thinking they were thinking, why not use AI?
00:29:12
Speaker
to know which farmer really need that input ah subsidy. So now every farmer have to take a picture with Deep Leaf to have access to subsidy and he won't use it on random treatment. It won't be recommended and selected by Deep Leaf AI as well.
00:29:29
Speaker
But let me tell you what is the ah the data behind it because you may think it's just helping them distribute subsidy correctly, but it's not just that. It's even helping them to become self-sufficient so they don't rely on other countries around them that's with whom they have some political problems.
00:29:50
Speaker
yeah I will give you an example which is which may be a bit sensitive right now in terms of what is going on. We've never shied we've never shied away from tough topics, so go for it. Okay, that's that's perfect. So, ah ah you know, in Iran, they have a very good agriculture and they have better climate than the other countries.
00:30:14
Speaker
like ah Qatar, Saudi Arabia, and all of those countries in the big Arabian island are almost desert. ah So most of them imports fruits and vegetables from Iran because of, ah not only because of their nature, ah resources and environment and climate, but also because it's very cheap.
00:30:35
Speaker
It used to work very well, but now with these problems that are happening between those governments, they are trying to become self-sufficient. So they don't rely on that importation anymore. ah And it's not easy to do. like ah You need to have enough data to understand what is and the activities or what are the things that you do as a Minister of Agriculture that can allow you to stop depending or at least to reduce the dependency on importation in terms of food ah for your nation.
00:31:09
Speaker
So with our solution, we are now moving beyond disease detection, but adding yield estimation. Every farmer, when they take that picture to detect disease, they can also take a quick video scan of some plants.
00:31:23
Speaker
And then we tell them, this is how much your nation national farmers are producing. And then based on other data coming from the population, how much they consume, for example, in tomato, or how much they consume in different crops, you will be able to balance. You will be able to know whether by the end of the year, I will be able to produce enough food for my people or I will need to import ah from another country.
00:31:49
Speaker
Yeah. I mean, that takes a that takes a lot of policy planning and, you know, agricultural supply lines are not things that are built overnight, but that's an interesting use case in understanding and forecasting.
Towards a Comprehensive AGI for Agriculture
00:32:01
Speaker
But now it's starting. Yes, of course, it's going to start this conversation. um So I want to close out here with understanding you've mentioned a little bit about this new feature in terms of yield estimation, but...
00:32:14
Speaker
What do you think, what are your future plans near term for DeepLeaf? It strikes me that you are constantly retraining your models and maybe different climate implications will change those data sets. Yeah, yeah. They have to be refreshed. Two months retraining is aggressive. I'll say that.
00:32:33
Speaker
Yeah. Mm-hmm. So we don't do retraining, but we are doing fine tuning, which is less so compute consuming. um In terms of future developments, as I said, we are developing the yield estimation solution.
00:32:49
Speaker
And we are also building an all of these tools that I mentioned, yield estimation, phytosanitary recommendation, and so on. We are trying to make them available to every agritech company.
00:33:01
Speaker
And we want to brand it as an AGI for ugly agriculture. So ah we want to be like the like anthropic or open AI, but for agriculture.
00:33:12
Speaker
So we have our own interface that users may use, but also other companies can use our tools, our APIs and so on to build their own use cases that fits their country's needs and climate and everything.
00:33:26
Speaker
brilliant So that's our direction. Great. Well, thank you again for joining us in the afternoon time from Morocco and sharing DeepLive's story. It's been really fascinating. Always good to hear about new kinds of AI developments and innovation rather than the stuff that we just see here in the West in the news literally every day.
00:33:49
Speaker
Thank you too. Thank you very much, George. I really appreciate this opportunity. Thank you.
00:33:58
Speaker
We talked about food security in this episode. So let's extend that metaphor and talk about media diets. Instead of leaving you with a question in particular about the subject matter, would leave you with this question of what don't you know about AI developments today?
00:34:18
Speaker
And if it is just the LLM narrative of the West, How can you diversify your media diet to learn more about some of the applications out there that are really changing lives, be it in healthcare, material science?
00:34:35
Speaker
Is there a different paradigm than large city block-sized data centers? Is there a way that we can involve more people in the design of these applications?
00:34:48
Speaker
So take that forward. We will see you next week.
00:34:56
Speaker
If you like this conversation, share it with friends and subscribe wherever you get your podcasts for a weekly ballistic payload of snark, insights, and laughs. New episodes of Bare Knuckles and Brass Tacks drop every Monday.
00:35:09
Speaker
If you're already subscribed, thank you for your support and your swagger. Please consider leaving a rating or a review. helps others find the show. We'll catch you next week, but until then, stay real.