Every conversation about "what AI can't do yet" assumes the current interface is close to final. It isn't. We're in the DOS phase — the era of the command line, the config file, the thing you have to know how to talk to before it'll do anything for you. Chat boxes, prompt templates, context windows you have to manage by hand: this is early tooling, not the shape of things to come.
The DOOM problem
The original DOOM shipped on a stack of floppy disks in 1993, and getting
it running was a project in itself. You were negotiating conventional memory
against expanded and extended memory, hand-editing CONFIG.SYS and
AUTOEXEC.BAT to free up enough low RAM for the engine to load, maybe
juggling a memory manager like QEMM just to get past the boot screen. None of
that complexity was the game. It was the toll you paid to reach the game.
Twenty-five years later, DOOM runs on a calculator. On a smart watch. On a pregnancy test, at this point, as a running joke about how little hardware it actually needs. Put that next to Borderlands or a modern Call of Duty — open worlds, physically based rendering, netcode holding sixty people in sync in real time — and the distance is staggering. But here's the part that matters: nobody today thinks about IRQ conflicts or expanded memory to play either one. All of that friction didn't get fixed. It got absorbed, layer by layer, until it was invisible.
Where AI actually is
Prompting is our CONFIG.SYS. Context management, retrieval pipelines,
picking the right model for the right call, wrangling tool schemas — this is
the toll we're paying to reach the thing underneath, not the thing itself.
It's real work, and right now somebody has to do it. But it is not the
permanent interface. It's scaffolding for a moment in time, the same way
memory managers were scaffolding for a moment in PC hardware.
The mistake is building as if today's friction is load-bearing — designing a product around the prompt box, a workflow around the chat window, a business around the fact that someone has to babysit the context. That's building a game around the memory manager instead of around the game.
What that means for building now
The systems worth building are the ones that survive the interface changing underneath them: clear separation between what a system is trying to accomplish and how it happens to be doing that this year, evals that measure outcomes instead of prompt phrasing, architecture that assumes the model call is a swappable part, not the foundation.
The DOS phase isn't a plateau. It's the part of the curve where the infrastructure is still visible. Our job is building for the part where it isn't — where the AI is running underneath the product the way the engine runs underneath the game, and nobody has to think about it to get to the part that matters.
Get in touch if you're building for that horizon instead of this one.