Making Food Programmable

“Programmability is eating the world.”

declared USV’s Nikhil Raman. USV’s Matt Mandel introduces programmability

The Information Age was a consequence of two developments: the ability to work with information programmatically and the democratization of that capability.

USV sees programmability driving impact in energy and biology (and money and intelligence), but in studying their posts on programmability I found myself longing for a precise definition of programmability.

What makes something programmable versus, say, controllable? Control theory is well over a century old––surely programmability refers to something new! A precise definition could also help make it clear what else could be programmable in this new AI Age.

Defining Programmability Link to heading

Let’s say something is programmable if

it has cheap observability and cheap compositional control.

A system is programmable when you can read its state cheaply and issue novel instructions from a grammar that allows new combinations without re-engineering.

For instance, cruise control is controllable but not programmable. It takes limited inputs (current speed) and uses them to maintain a target state.

In contrast, energy companies like Base Power and David Energy are programmable. Both might control multiple primitives mark_appliance_critical, set_optimization_goal, sell_energy each taking many parameters. Sunday Robotics’ Memo is clearly programmable.

Economies of Scope Link to heading

Controllable and programmable systems both rely on feedback loops, but programmable systems are distinguished by the compositional richness of their control system grammars.

Because of this, programmable systems possess economies of scope. They unlock efficient product diversification for a fixed capital investment.

Economies of scope traditionally refer to business operations––it’s cheaper to produce two similar products together than separately. But this concept readily extends to consumer hardware. A single device with high observability (sensors or vision) and compositional control (AI + integrated external tools) can perform many variations of many tasks.

This is the consumer form of Andrej Karpathy’s Software 2.0: edge-AI hardware that can improve for anything verifiable:

It should seem then that for programmable systems with significant economies of scope, we should want deployments, per China’s Moonshot AI, that achieve an “optimal conversion from energy to intelligence.” That is, we should want AI deployed to the edges with the greatest economies of scope.

Sensors are the precursors to programmability and autonomy (as posed by Peter Zakin). In the edge AI case this likely means deployment of sensors hooked up to AI and actuators or integrations that deliver maximal economies of scope for the least cost––cost denominated in hardware, time, energy or compliance.

It’s clear how these considerations roll into energy and robotics. But what about the home or the pantry?

Programmable Food Link to heading

Taking programmability as cheap observability (cameras in the fridge and weight sensors on the shelves) and compositional control (APIs connecting to grocery supply chain and logistics) the kitchen becomes the driving node in the programmable refrigerated supply chain.

Just as programmable energy relies on a rich grammar of control, a programmable kitchen would utilize its own set of primitives: log_consumption(item_id, mass), broadcast_demand(sku, quantity), prioritize_perishing(expiry_date), or optimize_cart(budget, dietary_constraints). This unlocks knowledge and action––personal inventory tracking and actions via integrations with retailer APIs, conveniently now being built for agentic commerce.

The refrigerator is the ideal place to start because the demand observed is the most valuable and dynamic. Unlike the pantry, where a can of beans lasts years, a carton of milk demands timely observability to prevent waste. Indeed, the entire US refrigerated grocery supply chain is in service of demand expressed and observed in the home refrigerator.

“Making markets programmable allows the cost structures to significantly shift and, ideally, give that value back to the customer.”

noted USV’s Rebecca Kaden. In the United States food waste is estimated between 30-40 percent of the food supply. When you go to the grocery store, the cost of expired food is rolled into the price you pay, in part because of the challenges in precise demand planning.

On the other side, the average American family of four loses $1,500 per year to uneaten food.

Programmability effectively delivers a ‘double dividend’ to consumers:

1. A programmable supply chain induced by observed demand signals lowers the price on the shelf by reducing retail waste.
2. The programmable fridge reduces the volume of food purchased by relieving you from food purchases that eventually go to waste.

Grocery is 3x bigger than energy Link to heading

Given these efficiencies and at a time when food prices are higher than ever, the opportunity is substantial.

The average American family spends 3x more on groceries than they do on energy. Programmable energy is a well developed vertical but programmable food is not.

Centering sensors at the point of consumption––so that nothing other than the act of consumption itself drives replenishment––has strong economies of scope relative to other tracking options.

While one might imagine general-purpose home robots performing this task, they provide only episodic and clumsy observability—a robot only knows what is in the fridge when it opens the door, even supposing it can actually see or access what’s in the back of the fridge. In contrast, a programmable fridge provides continuous observability, maintaining a real-time state of the world without interruption. This continuous observability is essential because consumption itself is a function of changing circumstances of time and place that cannot be predicted from purchase data alone. You need Hayek’s “compute on the spot”.

As a static, powered agent positioned at the center of consumption, an edge-AI fridge is ideally positioned to close the loop on the food supply chain and unlock programmable food.