Want to start a new green revolution? Start with plant-level data.
For more than a decade, global agriculture has been awaiting an AI revolution.
First, big data was going to change everything. 2013, Monsanto Company paid $1.1 billion As a machine learning startup, we created the first AgTech unicorn. Researchers believe that by 2019 Greenhouse using AI It will spark a second green revolution. Next, companies tried Teach a robot to pick fruit. Then ChatGPT came along and farmers were promised GenAI agronomist. Now, Booster’s claim agent AIor Even AGIultimately providing significant benefits.
Spoiler alert: None of these promises worked out. agtech fundraising is flattening This is because investors are reluctant to invest in idle land. Some bright spots are in narrow areas such as biotechnology and precision agriculture rather than innovative AI solutions.
There are two problems. For one, farms are a very difficult environment to develop AI. Second, current agricultural data are not sufficient to overcome these challenges. Until we solve that fundamental problem and start feeding AI models just data. more Although it is data, fundamentally better Data shows that most agricultural AI will continue to succumb to infestation.
Three reasons why agricultural AI fails
What makes agricultural AI so difficult? Consider the following challenges. Even one of these challenges sinks most projects.
- Feedback is slow. AI relies on rapid iteration, but breeding and testing new seeds takes time. Norman Borlaug won the Nobel Peace Prize in 1970, in part, thanks to the innovation that resulted from increasing the number of breeding cycles for crops from one to two per year. In recent decades, major seed companies have pushed this to three feedback cycles per year. Still, this is not enough for the fast feedback loops that AI relies on.
- High dimension. AI models suffer.”Accuracy collapse“Agriculture is one such environment, where a large amount of documentation is required to get an accurate answer. A simple question like how much nitrogen to apply involves a myriad of variables, from soil type to previous crops and yields, to pathogens and tillage history, to the presence of livestock on the property from decades ago. Simplifying all these variables into something that AI can handle can be virtually impossible.”
- There are many edge cases. AI models are best suited for: Handle a spherical cowor trying to predict the most likely next token, but real farms have myriad operational idiosyncrasies. To capture a broader range of edge cases, we either need to add dimensions (causing the curse of dimensionality mentioned above) or AI needs to develop something like a “world model” that goes well beyond current technology. Nothing can be generalized. Even if you had a model that solved the edge cases, its usefulness would probably be much narrower than expected because farmers’ needs are not the same. Even if they grow the same crop, they may have different technological aptitude, labor practices, access to capital, or farming philosophies. There is no AI model that will work for all farmers because there is no universal “right answer” to target.
Many people in Silicon Valley see these challenges as hurdles rather than obstacles. Inject more data and AI will eventually deliver results. There is certainly no shortage of agricultural data. The average farm has an estimated production of 500,000 data points per day. But nevertheless, there is something surprising a bit high quality data,and Garbage in and garbage out rules This also applies to agriculture. Billions of dollars have been invested to collect and organize the 200 TB (give or take) of high-quality data that powers leading LLMs. These data consist of hundreds of trillions of tokens. The corresponding ag dataset simply does not exist.
Understanding agricultural data issues
Part of the problem is that agricultural data is fragmented. Data differs from farmer to farmer. Standardization of data It’s very difficult to do without flattening edge cases.
But there is another deeper issue. Current datasets don’t provide the insights you need. AI models are trained based on external factors (weather, soil acidity, nitrogen levels, etc.) but know virtually nothing about how plants grow. respond to those factors.
Yes, temperature and precipitation affect crop growth. Yes, the levels of pathogens and certain molecules in the soil will affect certain types of plants in some way under certain conditions. However, these datasets only capture what is happening and cannot conclusively tell farmers what their crops need. outside plant. It’s like tuning a race car based solely on speed. Without access to engine telemetry, algorithms can’t accomplish much.
This does not mean that AI has no place in agriculture. Computer vision use cases such as differentiating plants from weeds and discarding spoiled fruit during processing have made significant inroads into the industry. But without plant-level insights, agricultural datasets are all noise and no signal. No matter how much we grow, we cannot overcome the challenges that are unique to this field.
Consider Gro Intelligence. The company has built the world’s largest repository of climate data focused on agriculture, raised more than $120 million, and recently closed the door. Or please consider acre valuea startup co-founded by one of us to turn farmland data into actionable insights, later admitted that its estimates were just a “starting point,” given that it’s nearly impossible to know everything about an acre of soil with a fixed list of variables.
Where agricultural AI plays an active role teeth Successfully, fruit sorting or Detection of unripe tomatoes or spray herbicidefor example, has a narrow range. Its power comes not from vast datasets, but from carefully selected use cases. Although effective, this is not scalable. The tomato sorting algorithm is useful, but it won’t usher in the next green revolution.
Looking inside the plants
Leveraging agricultural AI for its broadest and most valuable use cases requires data from “in the engine,” that is, from inside the plants you’re growing. New technology has made it possible for the first time.
For example, we are developing crops that communicate internal processes by emitting fluorescence. This summer, one of our plants contracted a fungal infectionthat immune response triggered a fluorescent signal, and for the first time in the 10,000-year history of agriculture, farmers learned about the pathogen before symptoms became visible.
This is good for farmers. Early warning means better results. But it also opens the door to new kinds of agricultural data. For the first time, AI innovators will be able to leverage data that reveals more than just what is happening. around it Crops are growing in the field, but what is happening? internal they.
This new data generation mechanism makes it possible to avoid the tedious process of inferring plant biology from external factors. Instead of building sprawling and overly complex models, experts are increasingly building lean algorithms that leverage data about a plant’s inner workings. You don’t need an infinite dataset. Gain customized insights directly from the plant itself.
the way to go
If all of us, more than 8 billion of us here today, are to avoid poisoning ourselves and cooking the planet while feeding our species, with billions more on the way by 2050, we will need agricultural AI that enables new efficiencies, higher yields, and greater resilience. But I can’t get there using the current dataset. For more than a decade, researchers have been trying to forcefully solve the agricultural sector’s challenges by accumulating more data and computing power, but they have reached a limit.
Now, armed with fundamentally better data that reflects real biological processes, we have a chance to break through the Gordian knot and give farmers meaningful insight into how to care for their crops.
companies like waymo and 1X is creating unique datasets that support extremely powerful AI models that interpret and interact with live environments. InnerPlant has built a dataset that records plant metabolism throughout the seasons across agricultural centres, allowing us to chart similar trajectories. Our sensors give you a head start. We are the only company that can detect disease as soon as a plant’s immune system responds.
There is still work to be done. Obtaining plant-level data is only the first step. But this is an important precursor to the kind of revolutionary AI that farmers have been promised for years. The agricultural AI revolution is finally here, and it starts with data extracted directly from the plants themselves.
