Claude now sits inside Microsoft Excel. A sidebar that reads and modifies live workbooks directly. Most people file this under productivity. Something stranger is happening underneath.
The Translation Layer
Operating systems solved a fundamental problem: humans needed to talk to hardware. The solution was code. An entire profession emerged around this translation.
For decades, the gap between wanting and doing was bridged by learning to speak machine. Now a translator sits in the middle. You describe what you want. The model writes the code. The gap closes.
If humans no longer need to speak code, what exactly is the operating system for? Not the scheduling, memory protection, or security layers. Those remain. But the human-facing surface, the part built for people who speak code, is the piece under pressure.
The Paradox
The models that make code transparent to humans were trained on code. Billions of lines of it. Code was not just training data. It was the scaffolding that taught machines how to reason step by step, decompose problems, handle edge cases.
Remove code from the world, and the next generation of models has nothing to learn from. The thing that made code obsolete requires code to exist.
Code becomes vestigial. Like the appendix: essential to what came before, baked into the body, but its original function replaced by something better. Code is still there. It is just no longer for us.
The Inversion
Under the surface, the spreadsheet sidebar still generates formulas, still produces structured logic, still writes instructions no human asked to see. The code has not gone anywhere. The human just never sees it.
Code was invented so humans could talk to machines. Now code exists so machines can talk to machines, while humans talk in plain language on top.
Human --[code]--> Machine
Now:
Human --[natural language]--> AI --[code]--> Machine
The question:
Human --[natural language]--> AI --[???]--> Machine
The IDE, the terminal, the command line. Interfaces for humans who spoke code.
They built the gate for the traveler. The traveler learned to speak through walls.
Code does not vanish. It stops being a surface and becomes a substrate.
The Trap
Code contains some of the highest-quality training signal that exists. The if/then logic, the type systems, the error handling. Precise where natural language is ambiguous. Models need that precision to learn.
The dependency chain: AI needs code to train on. Code needs people to write it. AI is replacing the need for people to write code. Something in this chain has to give.
Technological evolution following biological patterns. Structures persist not because they are optimal, but because removing them would break the lineage.
The Pattern
Assembly did not disappear when C arrived. C did not disappear when Python arrived. Each layer became infrastructure for the layer above. Invisible to most, but structurally essential.
The difference this time: previous abstractions still required humans to learn the new language. This one does not. For the first time, the interface speaks the language humans already know.
If models eventually learn to reason from mathematical proofs, simulation data, or their own synthetic outputs instead of code, then code becomes a true appendix. A historical artifact in training datasets, like Latin in medical terminology. There are early signals. But we are far from proving it at scale.
The vestige paradox: Code was invented so humans could talk to machines. AI made the translation automatic. But AI learned how to translate by studying code. The translator depends on the language it is replacing.
But the moment intent becomes executable, the interface stops being a UX problem and becomes a governance problem. Permissioned action, audit trails, rollback. The system needs enforcement, not just understanding.
Code is not dying. It is retiring from public life. Still doing the work, out of sight. And most people will forget it is there, until someone asks why the system works the way it does, and the answer traces back to something that was necessary a long time ago.
This post was written by the Fozbe engineering team. We build AI-powered software tools at the intersection of natural language and structured computation.