Shape of you
Reflections from Stanford's AI in Agriculture Immersion Day
Last week, I was invited to attend the AI in Agriculture Immersion day organized by Stanford’s Human-Centered Artificial Intelligence (HAI)1 and McKinsey. Today’s post is a set of reflections from the event.
Dr. Peter Norvig, eminent AI scientist and professor at Stanford, Google researcher, and the author of one of the most widely read textbooks on AI, Artificial Intelligence: A Modern Approach, and I have something in common. If you see Dr. Norvig’s photos online, he is always wearing a colorful shirt. I do have a small collection of themed shirts for agrifood events, with the tractor-themed one2 being among the most popular!
Dr. Norvig was the opening speaker at the AI and Agriculture Immersion Day at Stanford’s Computer Science and Data department. Stanford’s Center for Human-Centered Artificial Intelligence (HAI) and McKinsey organized the session. So they had much better food for lunch than what I experienced at World Agritech 2026 in San Francisco on Tuesday and Wednesday.
The audience included a bunch of food and agriculture executives, McKinsey partners, Stanford researchers, and a random lowly newsletter writer!
I had gone to the session with two specific questions in my mind.
What is the shape of problems that are most tractable for AI?
If the cost of experimentation is low, should we focus on use cases or just brute-force-apply AI and reimagine workflows?
I’m in love with the shape of you
My first question was answered right off the bat by Dr. Norvig’s presentation. He highlighted a class of non-physical problems: data-rich, with messy, distributed data, executable or simulatable, easy to verify, and semi-personalized and custom.
He highlighted sectors ripe for transformation based on his criteria. These sectors include wealth management, banking, paralegal, compliance, customer support, healthcare administration, real estate and insurance, drug discovery, and supply chain management. Software engineering fits in very well in this class of problems. There are tons of code samples (data), and the most important part is the ease of verification. Your code either works or it doesn’t.
AI adoption in agriculture faces challenges in verification and attribution. You provide an AI recommendation to plant a particular variety of seed or to apply the right timing for spraying your fungicide. The farmer often has to wait for a few weeks or months to see whether the recommendation worked. Even if the result is good, it is difficult to attribute the outcome to that specific recommendation because of numerous confounding factors, such as weather, soil conditions, and management practices. (I wrote a whitepaper with Tenacious Ventures on this topic in 2024)
I asked Dr. Norvig about Moravec’s Paradox and the challenge with physical tasks. Humans often do these tasks. For example, labor for harvesting specialty crops accounts for 60% of total production costs in California. Dr. Norvig was cautiously optimistic but felt we might be a few years away from robots picking apples or strawberries.
Dr. Norvig also presented a list of how not to fail with AI adoption (this bulleted list is taken directly from his slide)
Al is not regular IT!
Go bottom-up
Have a clear use case
Provide training and support
Make relevant data available
Optimize processes, don’t necessarily look for revenue
Problem-solving is a contact sport
There will be uncertainty
Make it everyone’s business
Encourage communication
As you can see from this list, most AI adoption challenges have less to do with technology and more with human factors. According to Dr. Norvig, we have done a great job of focusing on the easy parts of the job.
Dr. Norvig presented Amdahl’s law, which helps us consider how certain jobs can be sped up or automated in any field, and its implications for the food and agriculture system.
In 1967, Gene Amdahl3 made a simple observation about parallel computing. The speedup of a program is limited by the fraction of the computation that must run sequentially. Amdahl’s Law highlights that the maximum performance improvement in a system is constrained by its most significant bottleneck, making it essential to focus optimization efforts on the components that contribute the most to inefficiencies. (I had highlighted how bottlenecks and shifting bottlenecks have a huge impact on the overall process performance in my cotton article.)
Amdahl’s law is very important to think in terms of its impact on automation and which jobs are under threat. Dr. Alex Imas broke this down in a fantastic Substack article by looking at the dimensionality of a job.
On the other hand, if AI automates all of the tasks—let’s say your job only involves two tasks, and they both get automated—then yes, human labor will get displaced. Importantly, the fewer the number of tasks (what we call the dimensionality of a job), the greater the company’s incentive to automate it in the first place. This is the part much of the analysis on automation misses: adopting AI into an existing organization is costly, so the firm will be more likely to invest if it can automate the job, not just the task.
If you think about the activity of harvesting a specialty crop like lettuce or strawberries by a human, it requires a large series of tasks. Many tasks require experience and decision-making within a split second. For example, is this strawberry ripe enough for picking? Where should I cut the lettuce stem to get the product off the ground without damaging the lettuce? How many lettuce leaves should I take off, and which ones, before packaging the head of lettuce? People who have studied this more closely have told me that a farm worker has to make more than 20 decisions and take actions to get a head of lettuce from the ground and into a plastic bag, while maintaining the quality expected by the end consumer.
Even if we can automate some tasks in this process, it will be challenging to eliminate human labor unless we are willing to compromise on other dimensions, such as quality, food waste, or other aspects.
When in doubt, use brute force
My second question for the evening was about how indiscriminately we should try to apply AI in our day-to-day lives and businesses. Should we apply AI to as many use cases as possible?
Dr. Norvig had already mentioned the idea of having a clear use case for AI. He also presented the concept of Exaptation. Exaptation is a concept from evolutionary biology, but it could be very useful for AI in general and for AI in agriculture in particular.
Exaptation4 is a feature that performs a function, but that was not produced by natural selection for its current use.
The direct implication of exaptation is that we should have ideas floating around within the organization. We should think of AI as a general-purpose technology (GPT) rather than a specific tool. A 2023 research paper, “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models.”
In real life and especially in the online writer community, many people have also called it “increasing the surface area of serendipity.” The basic premise is that if you put your ideas out in public, you increase the chances that interesting people will engage with them. The paper referred to GPTs (LLMs) as GPTs (General Purpose Technologies).
The concept of exaptation should not be unique to AI. Still, given the innumerable possibilities of AI, it makes sense to have ideas about AI floating around within your organization, so that it increases the chances of someone else finding novel, not-designed ways to solve existing problems or solve new problems.
The presentation also alluded to Ken Thompson’s (inventor of the UNIX operating system) epigram, “When in doubt, use brute force.” Ken Thompson was implicitly saying that simple solutions are often better than complex solutions. If a problem is not excessively large, a simple, brute-force approach can be written and verified quickly, rather than wasting time on a “clever” algorithm that may be harder to maintain or more prone to bugs. It often works better when human time (programmer time) is much more expensive than machine time. (highlight by me)
What is the direct implication of this principle? I am not 100% sure.
When we find a domain or a set of problems that fit Dr. Norvig’s criteria, should we apply AI indiscriminately if running an experiment is not very expensive? We can potentially afford to do this because of AI’s general-purpose nature, provided the cost of experimentation is low. A great example of this approach went viral on X/Twitter when someone used ChatGPT to find treatment for their dog’s tumors. (This is not generally applicable or available.)
Please let me know if you have any thoughts on when and how AI should be applied.
Touch the grass
The second part of the presentation included a talk by Dr. Riitta Katila. Dr. Katila is a very energetic and engaging presenter, and the
Professor of Management Science, Faculty Director of the Stanford Technology Ventures Program, and HAI Sabbatical Scholar at Stanford Institute for Human-Centered Artificial Intelligence at Stanford University. Her research is in the intersection of technology strategy and organizational learning, using machine learning, statistical analysis, and mixed methods.
Dr. Katila’s presentation went deeper into the shape of problems, where AI is well-suited. Given the jagged nature of AI5 (and humans as well), it is a highly context-specific question, and AI can often be more creative because it might not consider constraints, whereas humans can implicitly consider implementation constraints.
Dr. Katila shared data from a paper by Brynjolfsson et al. (2025), which showed that using GenAI in a call-center setting improved the performance of the lowest-performing humans by more than 15% when the AI model was trained on inputs from the highest performers.
Even though AI in the call center boosted efficiency and customer satisfaction, highly skilled workers follow AI suggestions more often, even when those suggestions slightly lower the quality of the conversation. Because top workers are contributing fewer original, high-quality solutions, the AI has no “new” expert data to learn from. There is a risk that future AI iterations may be less effective at solving new or complex problems because their own average output is replacing the human expertise they mimic. There couldn’t be a more classic definition of “regression to the mean”!
In this use case, if you optimize AI agents for growth, you may underinvest in novelty and exploration. Dr. Katila recommended protecting innovation through human “exploration lanes”, incentives for original contributions, and deliberate non-AI inputs (real users, field experiments).
Touch the grass!
Where are you drawing the shape?
Discussing different contexts and problem shapes, she also presented data from Uber’s Q4 2025 earnings, which showed why San Francisco is a uniquely favorable market for AV (autonomous vehicle) deployments.
And how Uber’s network of human drivers allows Uber to have the highest AV utilization, when AV demand is highly variable based on the day of the week and the time of the day.
It highlights that not only is it important to consider the shape of problems for AI to tackle, but also to deeply understand the conditions in which they exist to understand adoption challenges and the economics behind them.
Is it time to cancel my medical insurance and pay for ChatGPT?
She presented data showing that an AI model specifically tuned on medical research data significantly underperformed a frontier model like ChatGPT-4o (We are on ChatGPT-5.4 now) in a clinical setting. OpenEvidence, the fastest-adopted software product in hospitals, uses a fine-tuned model to assist doctors with diagnosis and decision-making. I have to admit, I had never heard of OpenEvidence before. OpenEvidence’s performance is significantly worse than ChatGPT’s.
The research paper indicates that,
A notable strength of OpenEvidence is its integration of citation-linked, up-to-date references, an asset for clinicians involved in academic writing, literature reviews, or manuscript preparation. This feature makes OpenEvidence particularly well-suited for research-oriented tasks. However, its reliance on well-structured and precise queries may limit its practical utility in real-time clinical scenarios. In contrast, ChatGPT-4o exhibited strong contextual awareness, consistently interpreting vague, incomplete, or grammatically flawed questions with ease. This advantage enhances its usability in fast-paced, real-world settings where flexibility and immediacy are essential. ChatGPT-4o demonstrated higher accuracy, significantly faster response times, and greater interreviewer consistency, suggesting it may serve as a more reliable and efficient tool for point-of-care education and clinical decision support.
Once again, the problem’s shape, the data used to train your model, and the context become extremely important. Also, note to self. During my next doctor’s visit, ask the doctor to turn off OpenEvidence!
Understand shapes, contexts, and make choices
The AI immersion day6 highlighted the importance of understanding the shape of problems when applying AI, recognizing its general-purpose nature, practicing exaptation, recognizing human creativity, and thinking from first principles.
As humans have always had to do throughout history, we have to make conscious choices.
AI is absolutely creating anxiety throughout the economy. Parents like me with high-school and middle-school-aged children are thinking and worrying about what an AI-powered future will look like.
I want to leave you with some sane words from Ada Palmer, a University of Chicago historian,
The dangers of ChatGPT and its successors do not lie in the technologies themselves. They lie in the fact that we must now make choices, good or bad, about how we help those navigate the rollout effects.
Knowledge empowers; experience empowers; expression empowers; each of these twenty generations has been more powerful than the last.
That’s why, as AI stuns us with its possibilities, we need to take a deep breath, zoom out, and remember that information revolutions are the normal state of human life.
I have always enjoyed the programming put on by HAI, including Jeffersonian dinners and the screening of the documentary Thinking Games about the work of Demis Hassabis of DeepMind in San Francisco, which included a Q&A session led by Prof. Erik Brynjolfsson and researchers from Anthropic and OpenAI. As the researchers from OpenAI and Anthropic were forecasting the end of software engineering jobs in the next few years, I ended up giving parenting advice to a young Stanford professor sitting next to me, who had also worked in economic policy at the federal level in the US. The irony was not lost on me!
If you look at my profile picture closely, you can see the Tractor-themed shirt!
Evolutionary biologists Stephen Gould and Elizabeth Vrba proposed vocabulary to let biologists talk about features that are and are not adaptations. For example, feathers might have originally arisen in the context of selection for insulation, and only later were they co-opted for flight. In this case, the general form of feathers is an adaptation for insulation and an exaptation for flight.
Dr. Ethan Mollick’s article “The Shape of AI: Jaggedness, Bottlenecks and Salients” is a great starting point to understand the jagged nature of AI.
I didn’t attend the entire day. The second half of the immersion day featured presentations from McKinsey, which, frankly, were a bit boring and gave examples I had seen 20 years ago. So I was rude and left in the middle of the day.
In response to tractable problems to solve, what are shapes of economic vs physical problems in ag?…
What would Nick Horob or Jake Joraansrad say are problems for AI to solve?
The economics of ag maybe more ripe for improvement by AI? Profitability problems may match Dr. Norviq‘s criteria than production problems.🤷♂️