To me, productivity has always been about how to stay focused, and how to quickly return to the task after interruption. In other words, it is all about managing attention span. So it is interesting to me how AI is said to increase productivity. Although AI's prompt-response interaction is poor for attention span, it makes up for the inefficiency in other ways.
A typical AI driven development falls into this workflow:
- Create a skeleton project.
- This is usually done by invoking framework-specific tooling, such as
npm create <template>or something likenpx create-next-app@latest. AI reads the "getting started" guide and runs the command shown. - Create and populate a list of tasks.
- AI tends to under-specify, while humans tend to over-specify.
- Implement each task.
- This is akin to the StackOverflow searches that developers used to do. It's made possible because someone figured out some clever tricks to get something done, and the solution happened to be relevant to another person's problem.
Since project creation and task implementation are fairly mechanical, it takes about the same amount of work for both AI and an unassisted human to do them. Humans are of course limited by how fast they can read and type, but the same limitation exists when interacting with AI. The difference is that humans tend to get side-tracked when they see something interesting or unexpected, or they like to shop around for options before settling on a decision, while AI is unabashedly laser focused.
Deciding on the list of tasks is a double-edged sword. Under-specifying tends to get things done with the least amount of work, but it may miss critical components when planning. Over-specifying can waste time, but it can also create new growth tangents. There is a balance between a business process (think like Bill Gates) vs. a creative process (think like Steve Jobs).
The AI success stories tend to happen when the developer ecosystem has comprehensive step-by-step documentation, and many people have contributed solutions that are tried and true over the years. All this information is available for search on the web. Although the unsuccessful counter-examples are not talked about as much, it would not be surprising if AI struggles with lesser documented languages or more obscure use cases. Also, procedural documentation (written as step-by-step examples) tends to be more actionable by AI than declarative documentation (written as descriptions of defined behavior). Perhaps where AI actually excels is when the documentation should have been replaced by a library function or a simple command line tool that does not require reading tens of pages of documentation to understand what it does.
In summary, AI increases productivity by being focused on the task at hand, by doing the minimal amount of work, and by following the path that is well traveled. Indeed, many people enjoy great success by picking these low hanging fruits, and AI makes it even easier. However, AI also levels the playing field because everyone ends up picking the same low hanging fruits. Ruthless pursuit of efficiency begets ruthless conformity. Then it begs the question: why bother doing the same things that everyone else is doing? When everyone is ghiblifying their photos, the art just becomes petty. We will have to think more carefully about the purpose of things we do.
