Farming, Fast and Slow

Right after Michael Lewis published Moneyball, people started coming out with every possible variation: Moneyball for urban planning, Moneyball for early childhood education, Moneyball for medical treatment, Moneyball for you name it. If you haven’t read the book, it’s about how the manager of the Oakland A’s used statistics to govern hiring decisions, outperforming the expert judgement of his scouts. A pair of economists wrote a book review on the phenomenon, essentially summarized as “Duh”, because the shortcomings in the subjective judgement of “experts” had been deeply examined by two psychologists, Daniel Kahneman and Amos Tversky, who founded the field of behavioral economics. Kahneman had recently won the Nobel Prize in Economics. Everything that’s old is new again.

“Do I contradict myself? Very well, then I contradict myself.

I am large, I contain multitudes.” — Walt Whitman

Kahneman simplified our many contradictory impulses by suggesting we have just two brains. We have a fast, instinctive brain that is effortless in making gut assessments (“System 1”) and a slow, methodical brain that can eventually calculate how to send a man to the moon, but only with a lot of concentration (“System 2”). When I wake up on Saturday and realize I have no coffee in the house, I can get dressed, hop in the car, drive to the store, pick up a pound of beans, banter with the cashier as I pay for it, head home, and make a pot of Joe all without the use of my sentient brain (remember: no coffee). There is no robot or autonomous vehicle on Earth that could pull off any one of those steps, and yet for Fast Brain it is effortless. When I go to the store Sunday, my wife gives me a list to get for the week: French Ham (thin sliced), havarti (plain, no dill), pickles (the naturally brined ones), and don’t forget she’ll be gone Monday night so get something to make for dinner. It’s basically just a list, any computer could store that, but it takes up all my brain power to run them over using my Slow Brain so I don’t forget. It’s like keeping plates spinning!

I bring this up, because I observe that there is an astounding array of digital agriculture products on the market, and yet we haven’t seen anything really adopted fencepost to fencepost. For context, no farmer doesn’t use a tractor, fertilizer, or improved seeds.* The value proposition for each of these is so obvious and so large, that one taste of success and there’s no going back. A farmer doesn’t try fertilizer on some acres each year to see if it makes a difference. It does.

What is different about digital products is that fundamentally they shape decisions people make, which is to say that they engage with the human brain. Look over at Facebook and WhatsApp and YouTube and Instagram and YikYak and whatever else they’re making these days and you can see that they have not only engaged with the human brain, they have fully reverse-engineered it to do their bidding (which is to serve you advertisements). Google perfected the system that you literally can’t opt out (cite). But there is something wrong in the way that agtech products have engaged with people and their brains, that for whatever reason is not compelling or sticky or value creating. And I think it has to do with this fast brain / slow brain stuff.

This piece is a lens into what I wake up thinking about every morning: when are we going to see every acre managed digitally? I want to probe here two questions, namely:

• How technologies are marketed and sold
• How decision support tools are used effectively

First, some macro trends defining agriculture. Over the last century, the number of farmers seems to be halving every generation, and yet the same amount of land is farmed, more or less. And so, history shows that the yield gains per person are far greater than the yield gains per unit of land. What this implies is that there is more value creation in super-powering people (returns to labor) than boosting yield (returns to land). It’s valuable to see farmer adoption of technology over the last century in this light. Are GPS guided tractors something that improves yields? Or makes a person’s job easier? Are the cocktail of glyphosate + gmo soy something that improves yields? Or makes a person’s job easier? I’m happy to acknowledge that they improve yields, but the choice to adopt (and we see that the pace of their adoption was rapid {link}) was driven more by benefits to people working the land than to the land itself. Is the reason so many digital technologies don’t get adopted, is because they add complexity to human decisions (negative returns to labor) even though they boost productivity (positive returns to land)?

Maybe we should look at agricultural technologies as being part of the same trend that drives digital transformation of all other businesses, namely human productivity tools. How do we at Arable run a business with employees in multiple timezones at any given moment, expertise spread around the globe, serving a clientele in 40 countries? Productivity tools! It’s little different than so many of the multi-billion-dollar food and ag companies we serve, which have employees in multiple geographies harvesting and shipping products around the country and around the world.

In a key respect though, these aren’t really the productivity tools that move the needle in ag. Farming is far more like Google Maps: A million moving parts, interacting, with massive random perturbations. We sometimes call these “complex adaptive systems.” And like fertilizer to a wheat farmer, no professional driver doesn’t use Waze, always, even though they know where they’re going. What’s distinct about this tool from all the productivity tools up above is that now we’re in the realm of predictive analytics. It’s a tool that helps people evaluate and decide on multiple scenarios to cut a path through the haze of uncertainty and risk. Such a product is engaged heavily in the recesses of the human brain that Kahneman excavated.

Predictive analytics is such an essential dimension of digital innovation in agriculture, because there are essentially infinite combinations of weather, soil, management, and genetics, which interact, and these interactions are adaptive. We know that when traffic changes, people adjust their routes, including ourselves, and that this in turn changes the traffic itself. We know that when a weather event happens, say a heat wave, the plants respond, say in their leaf function, or root/shoot allocation, and people make adjustments to the knob, say by running a sprinkler, and this in turn shapes not only the plant, but the microclimate experienced by the plant. Or, the weather turns cold and wet, fungal pathogens start to grow, people respond by applying micronutrients, the plant is able to fight off the infection, and the fungal population drops. This is biology, not physics, and it’s why data is essential to track the current state of reality! In the face of such complexities, calculating the odds of any future state is devilishly hard, because there are so many individual factors. Because of this, we revert to stereotypes, past situations that may not be relevant to the current circumstances.

Reversion to stereotype is one of the “fast brain” behaviors that Kahneman and Tversky saw as flaws, because they bely our self-conception as rational, analytical beings. They found that humans are bad at:

• Evaluating probabilities
• Assessing the price of risk
• Incorporating new information
• Holding many things in their head
• Updating past patterns to new situations

All of these are slow brain problems, but with rare exception, we conduct agriculture using our fast brain: instinct, experience, trusted experts. It’s not hard to see why the yields of the average producer are so much lower than the best producer. It’s nobody’s “fault”, there’s just a lot of information to assimilate for each of the many consequential decisions made each year!

THE SAMUELSON PARADOX

When I think of marketing side of digital technology, I think of the lessons of Samuelson Paradox.

• I’ll offer everyone in the crowd the following deal: I’ll give you $10 cash now, OR a 1/100 chance to win $1000. The expected value is identical. Almost everyone takes the cash.
• Now I’ll offer everyone in the crowd the opposite deal: I’ll charge you $10 cash now, OR there will be a 1/100 chance to be fined $1000. The expected value is identical. Almost everyone takes their chances.

The paradox is that we behave in opposite ways to potential upside and to potential downside, even when their monetary value is identical.

That, in a nutshell, explains our attitude to insurance. People don’t like paying for it. In fact, the individual mandate to buy health insurance is one of the most acrimonious policies of the last generation. And so, across so many perils, a relatively small number of people buy the insurance unless they are required to. How many people here buy more than the minimum of car insurance?

I’m intrigued by this because most digital ag is packaged and sold as insurance. Which is to say people are asked to pay up front for a chance to avoid potential loss. Is this why we don’t see mass adoption of digital ag, because the business model is wrong?

Just on time, we’re seeing companies that are turning the traditional sales model for digital ag on its head. Instead of asking farmers to pay to insure against a potential loss, there are new business models that bring so much confidence into their predictions that they are essentially paying farmers to adopt, so that the offering company can participate in the upside. In the limit, this can be used to enhance every consequential decision: what variety to plant at what population; whether to spray and when; what new technologies to try. Xarvio’s Healthy Fields initiative removes the risk associated with making crop protection decisions, by essentially guaranteeing a minimum benefit. One could argue that Indigo’s commercial strategy is similar: by guaranteeing a premium price on the harvested grain, they have eliminated the risk of adopting a new technology that has only been in the market a couple seasons. By the expansion of this program to include payment for ecosystem services (accumulation of soil carbon), Indigo Acres pushes this even more broadly than simply marketing premium attributes of the seeds they sell.

CLOSING THE LOOP

Another strategy for leveraging the fast brain is in reducing the complexity of calculations down to red/black up/down hold/sell decisions. Our car just got hit; the repair is more than the car is worth. Take the money. A recession hits, and the value of the house is less than the mortgage. Wait until the market returns. Your car is veering left. Turn the steering wheel right. The bag in my left hand weighs more than the bag in my right hand. We make these assessments more or less instantly, without thinking.

When it comes to digital ag, the number one complaint (after too many logins to remember) is that there are too many numbers on the screen: just tell me what to do. Instinctively the human brain is looking for a good/bad high/low now/later binary discriminant, and yet so often what we provide is multivariate story that varies in space and time. Those are really difficult stories to build a narrative around! What I am seeing as successful is in closing the loop between expectation and reality.

We did a trial in processing tomatoes in 2017–2019 that pulled together a staggering array of data, including energy balance models of water demand, soil moisture, flow monitors, crop models, and yield / quality measurements, in order to drive a valuable decision around deficit irrigation to improve brix and thus price per ton. But initially, the presentation didn’t raise any eyebrows: we are providing a runtime recommendation, but the existing irrigation scheduler also provides such a recommendation. Separately, we showed run times, but they have their own records of run times. Yawn. Put these two together though and WOW suddenly it pops out that the irrigator put on 14 hours of water when they were only supposed to put on 4. In fact every day for weeks, the irrigator put twice as much water as they needed to. Nobody in ag got fired for over watering, so if a little is good, a lot must be better. (This is subsurface drip, so the saturation is not visible to the naked eye). Suddenly, the farm manager can see what he’s looking for: a binary discriminant of high/low more/less good/bad that forms the basis of a stern conversation.

It’s why I call this blog The Expected and The Observed. In the gap between these two lies our ability to learn. Without a model (expected) we don’t know how to interpret data. Without data (the observed) we don’t know how to interpret the model. In grad school, a friend summarized it thus: “Theory without data is science fiction. Data without theory is science nihilism.” In agtech we have tended to have separate communities that offer models without much data, or data without much context or interpretation. What is thrilling in the present moment transitioning digital ag to computational ag, is that we are seeing that the union of data (“labels”) and models (“features”) is now becoming the norm for how products are developed and released.

DEDUCTIVE, INDUCTIVE, ABDUCTIVE

When Betty Crocker first sold instant cake mix in America, it was a flop. The marketing department was puzzled. Why were women resistant to making this easy cake and earn the endearment of their family? The answer was psychological (link): The women didn’t feel like they were adding their own talents to the work, that it was not really “cooking”. After uncovering this insight, the product designers changed the cake mix to add one ingredient (an egg) and one instruction (to beat it). And the result was a wild success! The interpretation to this has always been somewhat Freudian (in fact, the focus group consultant was Freud’s nephew), but for me the lesson is that nobody likes to be prescribed exactly what to do. I always tweak the recipe. And farmers are no different: it would be a fool’s errand to be prescriptive about corrective action when you don’t know someone’s goals, their budgetary or machinery constraints, their risk/reward profile.

We are aware of deductive reasoning (if X happens then Y will happen). We are aware of inductive reasoning (if Y happened then X most likely happened). But we are less familiar with abductive reasoning (if I want Y outcome, what X is implied?). This is also known as goal seeking or scenario planning. And goal enabling is where I think the center gravity for recommendations or prescriptions in digital ag needs to be.

What does abductive reasoning have to do with Betty Crocker cake mix? Every farmer has a sense of what their goal is, their strategy, their secret recipe, but they often don’t have enough computational tools to assess how to reach the goal or implement the strategy. The deficit irrigation irrigation is a perfect example: every tomato farmer knows about and wants to reduce irrigation by some amount to improve fruit development, but they most likely don’t know when in the season to begin (it’s a complicated crop model) or how much water exactly to apply (the ET calculations are complicated). But with a simple slider, a grower can set irrigation at 80%, 70%, 50%, some fraction of ET and implement the strategy. It would be irresponsible for me to suggest a precise value for the run time, but I should make the farmer’s life easy by teeing up the data to allow goal seeking. I make the recipe book, the farmer is the chef. Talent has always been in the strategy and execution, not the dumb calculations.

* Well — sometimes they do. I used to belong to a CSA, Windborne Farm, that grew heirloom grains with drought power fertilized by manure. But let’s discount those five acres.