Think back 10 or 15 years ago, or further, and imagine that you wanted to order a pizza for delivery. You would call a store, talk to an employee, and place your order. In some reasonably expected timeframe, the order would show up at your door, delivered by a pizza employee to whom you would provide payment and (hopefully) a tip for their service.
Now imagine that they delivered the wrong items. You discover it after the fact, call them up, and they bring you the correct items as soon as they can.
Today, may pieces of this process have changed. You’re probably ordering the food from a third-party, who contracts with restaurants to be their delivery partner. If something goes wrong, you often resolve it with the third-party, typically through a credit for your next order.
Transformations like this happen in part because of scale. If a pizza place wants access to a wider customer pool, they partner with a big company that does a lot of deliveries in the area. Customers who want more variety use third parties to get a consistent experience whether they want pizza, barbeque, or salads and wraps.
The flip side of scaling, though, is more randomness in the process, or stochasticity. A single, relatively small pizza shop can manage customer expectations carefully, because they only have so many customers in a given day. A large third-party platform has a huge amount of uncertainty to manage, from individual stores to drivers to the platform/app itself.
Pre-platforms, for the most part processes like delivery “just worked.” But now the definition of “just worked” has shifted to account for the more stochastic nature of things. Platforms bake randomness into their expectations. This is why you can get an automatic credit when you report a missing item or bad experience, they budget for it. Over time, we shift our expectations based on this new experience, and we start to expect more randomness in what used to be very clear-cut processes with known, sspecific outcomes.
Nowhere is this more obvious than with the current wave of AI, powered by large language models (LLMs). LLMs are powerful and probabilistic. They are also often billed as general-purpose AI, as opposed to prior AI technologies where models were carefully trained for specific tasks. This is another example of scaling. As the models increased their scope of what they try to do, the deterministic nature shifts to a more stochastic one. In the extreme, this leads to hallucinations; however, even when an LLM-powered AI is “working,” you don’t get a precise, exact output. You get something like a paragraph of text that “makes sense,” “looks good,” or “should be ok.”
There is a case to be made that this shift opens up a lot of opportunites for a mindset shift. However, while this stochastic mindset may open unique opportunities in the workplace, it may have some downsides for customers specifically, and humanity more broadly. Without being too dramatic, as more and more things become stochastic in day-to-day life, there are by definition fewer and fewer things that an individual can rely on to consistently act as expected. While most things will work most of the time, lingering uncertainty will act as a sort of reverse slot machine. Pull the lever and things should go fine, but there’s always a chance that you don’t get what you ordered, your AI interaction goes off the rails, or your call to customer service gets routed to the wrong department or disconnected entirely. All you can do is shrug and try again, because in the limit, things are working as expected.
While we live in a world with a significant amount of uncertainty, and can adjust expectations to deal with the uncertainty, humans also need fixed points for grounding ourselves and offloading some degree of vigilance and monitoring. If I’m worried about whether my delivery order will be correct, I’m not able to focus on my kids as they play before our meal. If I have to check an app to make sure the product I ordered arrives on time, I’m pulled away from my actual workday tasks.
Stochasticity opens many possibilities for scaling services and experiences, making things available to people who might not otherwise have an opportunity. But let’s not assume that it is a complete positive (or even a net positive, he said provocatively). In particular, from the side of the consumer/recipient/human, there is something to be said for the predicatbility of specific requests or interactions resulting in specific outcomes. And it goes beyond just pizza.