The Dawn of Infinite Code
San Mateo, 2005: three founders add an upload button to their video-hosting website in an office that smells faintly of pepperoni.
Everywhere, 2025: a person types an idea, presses enter, and a network of agents compiles, tests, and executes.
In The End of Software, I posited that large language models would drive the marginal cost of software towards zero, just as the internet did with content. A year later, that forecast is playing out. Code is beyond trivial to generate. The question now is how we deal with the oncoming flood.
The pattern is familiar. When YouTube launched in 2005, its first video, Me at the Zoo, seemed immaterial. This was the harbinger of doom for the media industry? But that was the point—it proved that content had become too cheap to meter. Now, the dam has broken with code. Prompts produce prototypes that once required a team and a sprint. As transformative as broadcasting from your bedroom was, anyone can now ship their ideas.
Web 2.0 gave everyone a single verb: upload, and it vaporized newsrooms everywhere. AI gives us a new one: describe, and the 500-person single feature company goes poof. The work shifts from syntax to intent. Instead of writing software line-by-line, we define what it should do and let models handle the how.
But what about distribution? Don’t software companies have robust sales and marketing teams? Yes, as did newspapers with paper routes and circulation numbers. When free competes with free for distribution, paid distribution loses badly.
At first, these models will rely on a set of default environments and stacks determined by a legacy training set. But defaults change. As options multiply, the decision of where and how code runs will be handled by algorithms.
We already accept this logic elsewhere. TikTok decides what we watch. Spotify decides what we hear. The system is opaque. The results are good. So we scroll.
When the flood of code arrives, we won’t ask for explanations–we’ll reward speed. Convenience will outrank transparency. The systems that earn our trust through responsiveness will hold the power once reserved for operating systems.
As cost curves come down, we chase scarcity up the stack. It starts with compute, still mostly centralized but already seeping to the edge as local models take up residence on laptops and phones. Above that lies context, the habits, preferences, and history that let a model anticipate us better than any public corpus ever could. And finally, trust, credibility earned over time determines whether an agent has the right to choose on our behalf.
Once we arrive at full trust, active choice becomes a burden.
Children born after 2010 have never opened a newspaper. Soon they may never open applications.
History shows the vessel always fades once the payload can flow without it. Newspapers yielded to websites; CDs to streaming. Software, in its current circumscribed form, is the next container on the chopping block. When logic can be summoned and stitched together on demand, the container evaporates and only capability remains.
That’s the real end of software: not the death of code, but the moment its container dissolves. Utility unboxed, everywhere, all at once–infinite capability, infinite code.