The Next Productivity Leap
In 1947, a paper called “Preparation of Problems for EDVAC-Type Machines” introduced the subroutine, arguing that operations which recur frequently shouldn’t require re-coding from scratch each time. At that moment, programmers wrote every instruction their task demanded. It was slow, error-prone, and confined to a narrow group of specialists.
That paper marks the start of a long arc in computing. The history of programming is largely the story of removing impediments between an idea and its implementation: high-level languages, libraries, syntax highlighting, IDEs, and code completion all reduced the overhead of translating thought into functioning software. Each step widened the pool of people who could meaningfully build with code.
Why AI Changes Who Programs
Large language models that assist with writing, debugging, and modifying code are the latest step in that trajectory. They address an oddity of traditional programming: the practitioner must be fluent in arcane details—correct semicolons, function signatures, loop boundaries—while simultaneously holding the architecture and data flow of an entire system in mind.
That kind of cognitive load is a barrier. It’s part of the reason Scratch, the children’s language, was designed the way it was: to reduce small mistakes so that learners can focus on structure and logic rather than syntax. AI assistance offers the same possibility for everyone. If models can absorb the minutiae, more people can concentrate on what they’re actually creating.
We’re not at the point where AI eliminates the need for traditional skills entirely. We are, however, at a point where AI is a powerful assistant—one that helps more and more people become effective programmers. Platforms built primarily for experienced developers are beginning to embed these assistants so that newcomers can accomplish what previously required deep familiarity with a system.
Platforms Must Treat Data as Fluid
AI has another implication beyond assisting with code: it should be a core affordance of the developer platform itself. A platform without integrated AI capabilities is on its way to being inadequate, much as one lacking floating-point arithmetic or list handling. This is clear from user expectations. When phone photo libraries can classify and search images by content—tap in “cat” and see every cat photo you’ve taken—that's a capability users expect to build with, not reinvent.
That vision rests on an infrastructure assumption. AI thrives on enormous volumes of data—gigabytes to petabytes—and, unlike a database where records sit largely dormant, model training and inference demand constant movement of data. To support AI, developer platforms must allow data to flow between systems freely. Discouraging that flow with transfer fees or egress charges treats data as a lock-in mechanism, not an engine of innovation. A healthy AI ecosystem needs platforms that don’t hoard the data they touch.
A Broader Change
It’s tempting to dismiss the enthusiasm around AI, but the scale of this shift is hard to exaggerate. Andrew Ng has compared AI to “the new electricity,” and the analogy holds: electricity changed daily life so completely that a supply failure reads as an anomaly. AI is already pervasive in that quiet way, through translation systems, voice agents, and the automated defenses that keep pace with the internet’s speed.
It’s more than a convenience for technology workers. Consider the game Go. When AlphaGo defeated the world-class player Lee Sedol in 2016, it might have been demoralizing. Instead, human players who trained against AI programs got measurably stronger, absorbing strategies they would never have discovered on their own. If AI can make elite Go players better, it has real potential to raise capability—not just efficiency—across fields, programming included.
Forty years into a programming career, it’s difficult not to see that future as the most promising in computing. The subroutine unlocked reusability; the language and IDE revolutions unlocked speed and accessibility. AI’s capacity to bridge the gap between intent and implementation points to a fundamental change in how software gets made—and by whom. It’s fair to suspect that a 2047 paper, perhaps titled “Preparation of Programs for NEURAL-Type Machines,” will again describe a way of expressing programs that current practitioners can only glimpse.
