Software as Universal Skill

The printing press analogy

The clearest version of this prediction holds that writing software will become as broadly accessible as reading and writing, not a specialist skill requiring years of training but a general-purpose tool available to nearly anyone. Before the printing press, roughly ten percent of the European population was literate, a share often employed by rulers who could not read themselves, reading and writing on their behalf. After the printing press, more literature was published across Europe in the following fifty years than in the prior thousand combined, the cost of a book fell by roughly a hundred times over that period, and global literacy eventually reached around seventy percent over the following few centuries, to the point that no one today needs a formal credential simply to read and write, even though professional writers remain a distinct role. The prediction is that software follows the same arc, but far faster than the centuries literacy took to spread, driven by AI coding agents rather than a printing press.¹

The domain-expert corollary

The most counterintuitive implication is direct: "the best person to write accounting software I think maybe even today is not an engineer, it's a really good accountant, because they know the domain really well and coding is the easy part. It's knowing the domain that's the hard part."¹ This inverts the standard framing of AI-assisted coding as primarily a productivity tool for engineers: the larger shift is that domain expertise becomes the scarce resource while coding becomes commodity infrastructure underneath it. An accountant who deeply understands tax law, edge cases, client workflows, and failure modes can write better accounting software than an engineer who has to learn all of that domain knowledge secondhand first.

Evidence it is already happening

Palmer Luckey supplies a complementary data point from the hardware side: as a self-described weak software engineer, he argues that even rough AI-generated code is better than the nothing a domain expert such as a hardware or opto-mechanics specialist would otherwise ship, since the relevant comparison is not an AI agent's output against a senior engineer's, it is an agent's output against the nothing that expert would have produced without it.

Direct organizational evidence backs the same claim. Inside one AI company's own product team, the engineering manager, product manager, designers, data scientists, finance lead, and user researchers all now write code themselves, a functional role that was previously gated almost entirely to engineers.¹ A company-wide hackathon built around a recent flagship coding model produced winners who were mostly non-engineers, including a doctor, an electrician, and a carpenter with no prior coding experience, one of whom went on to build and sell a small company as a direct result, and before a friendlier interface existed, a meaningful share of usage came from non-technical people installing a coding agent in a terminal for the first time in their lives, simply because the tool was useful enough to be worth the friction. That willingness to push through developer-tool friction is treated as the real market test: are there a lot of people using this every single day and continuing to use it.²

A large music-streaming company's internal prototype tooling is the most concrete production example of the same corollary at work inside a large organization. Infrastructure built so that anyone can describe an idea in natural language and get a working, end-to-end prototype across the real mobile apps and backend within one to two hours, shared afterward through an internal app store where colleagues try it against real data, has been used from individual contributors up through one of the company's own co-chief-executives, and the resulting prototypes are described as good, because senior people had always had domain ideas they simply could not previously get built, their engineering teams occupied elsewhere. The unlock matches the domain-expert corollary directly: the person with the idea no longer has to convince a team of engineers to build it, they can build and test it themselves in a day instead of weeks or months, playing out inside a large existing organization rather than only at new startups.³

Why the constraint moves to knowing what you want

A separate argument gives this corollary an economic floor. Software has historically only gotten built above a certain minimum number of uses, and AI agents drop that threshold sharply toward a single user, which reframes the abstraction ladder from vacuum tubes through assembly language through modern programming languages as one that keeps climbing past code itself. It is not that everyone learns to program in the traditional sense, it is that the act of programming stops being the bottleneck activity. What remains scarce is specification: knowing precisely what you want built and being able to say it clearly, exactly the domain expert's advantage this concept identifies.

Timeline uncertainty

The prediction is directionally confident but deliberately vague on exact timing, framed only as much faster than the roughly fifty years the printing press took, itself a fraction of the centuries literacy took to reach seventy percent of the globe, a spread constrained by farming economies, missing education infrastructure, and the genuine difficulty of learning to read from nothing. Software learning faces fewer of those structural constraints, since models lower the barrier to entry dynamically and the surrounding infrastructure, smartphones and internet access, is already close to global.¹