M
Max Shestov
Guest
Yesterday I watched DHH’s Rails World 2026 keynote, and one line stuck with me:
A little over a year ago, DHH was still writing about how much he enjoyed programming and had no desire to give it up to AI. Now 37signals has effectively gone "pencils down" on handwritten code, and DHH says he has "retired from being a professional programmer."
Maybe it is time to admit something we keep avoiding:
Software development is no longer the same thing as writing code.
Yet we still behave as if it is.
We still interview engineers with LeetCode, even after Microsoft Research described the technical interview process itself as broken. We teach computer science as if producing code is still the central output. We keep trying to fit AI into rules created for a world where producing the implementation was the difficult part.
I am doing a master’s in Computer Science at Georgia Tech right now, and this gap is especially visible in education. Courses take years to design and update. Even if universities start adapting now, parts of the new curriculum may already be outdated by the time they arrive.
And adding an "AI course" will not fix that.
What needs to change is what we consider evidence that someone knows something.
If an engineer can define a problem, decompose it, get an agent to implement it, identify when it is wrong, test it, and improve it, why should we care whether that engineer could manually type the same implementation?
We already accept abstraction everywhere else. Programmers do not manually translate source code into machine instructions. We use compilers, runtimes, frameworks, databases, and operating systems.
Maybe code itself is becoming another layer of abstraction.
The same problem exists in writing.
Stanley Druckenmiller was criticized after it became known that he used AI while preparing a Wall Street Journal opinion piece. His response was essentially: the views were mine; AI helped express them.
Would the reaction have been the same if a speechwriter had done it?
We have accepted ghostwriters, editors, research assistants, and communications teams for decades. AI makes us uncomfortable mostly because the collaborator is a machine.
Academic publishing exposes the same contradiction. AI can improve grammar and readability without disclosure under many policies. But if I dictate my argument and AI turns it into readable prose, did AI "generate" the text?
Where exactly does editing become generation?
Maybe the distinction between "AI-generated" and "human-written" is already becoming useless.
A more useful distinction is between AI as the source of the thinking and AI as a tool for expressing or implementing human thinking.
There is obviously a huge difference between raw AI-generated slop and something carefully edited, verified, and reviewed. But we will not develop good tools and norms for showing that difference until we first accept a more basic principle that we should stop treating the tool itself as something that needs justification.
Write with AI. Write without AI. Code with an agent. Code by hand. Dictate your thoughts. Use an editor. It should not matter.
What should matter is whether the result is good, whether you understand it, whether you verified it, and whether you take responsibility for it.
Eventually, we will stop labeling text as "AI-generated". Saying that something was produced with AI may sound as strange as saying it was typed on a keyboard or written in Microsoft Word.
The same should happen in software.
Understanding code still matters. Fundamentals still matter. But fundamentals have always moved upward as abstractions improved.
Maybe writing code is starting to move downward.
Before we can redesign how we hire, teach, test, and evaluate people, we first need to admit that the shift already happened.
And there is an irony here.
I did not type this article from scratch. I explained these thoughts out loud. AI helped organize them and turn them into readable prose. I reviewed the result, changed it, and I am responsible for what it says.
Did AI write it?
A little over a year ago, DHH was still writing about how much he enjoyed programming and had no desire to give it up to AI. Now 37signals has effectively gone "pencils down" on handwritten code, and DHH says he has "retired from being a professional programmer."
Maybe it is time to admit something we keep avoiding:
Software development is no longer the same thing as writing code.
Yet we still behave as if it is.
We still interview engineers with LeetCode, even after Microsoft Research described the technical interview process itself as broken. We teach computer science as if producing code is still the central output. We keep trying to fit AI into rules created for a world where producing the implementation was the difficult part.
I am doing a master’s in Computer Science at Georgia Tech right now, and this gap is especially visible in education. Courses take years to design and update. Even if universities start adapting now, parts of the new curriculum may already be outdated by the time they arrive.
And adding an "AI course" will not fix that.
What needs to change is what we consider evidence that someone knows something.
If an engineer can define a problem, decompose it, get an agent to implement it, identify when it is wrong, test it, and improve it, why should we care whether that engineer could manually type the same implementation?
We already accept abstraction everywhere else. Programmers do not manually translate source code into machine instructions. We use compilers, runtimes, frameworks, databases, and operating systems.
Maybe code itself is becoming another layer of abstraction.
The same problem exists in writing.
Stanley Druckenmiller was criticized after it became known that he used AI while preparing a Wall Street Journal opinion piece. His response was essentially: the views were mine; AI helped express them.
Would the reaction have been the same if a speechwriter had done it?
We have accepted ghostwriters, editors, research assistants, and communications teams for decades. AI makes us uncomfortable mostly because the collaborator is a machine.
Academic publishing exposes the same contradiction. AI can improve grammar and readability without disclosure under many policies. But if I dictate my argument and AI turns it into readable prose, did AI "generate" the text?
Where exactly does editing become generation?
Maybe the distinction between "AI-generated" and "human-written" is already becoming useless.
A more useful distinction is between AI as the source of the thinking and AI as a tool for expressing or implementing human thinking.
There is obviously a huge difference between raw AI-generated slop and something carefully edited, verified, and reviewed. But we will not develop good tools and norms for showing that difference until we first accept a more basic principle that we should stop treating the tool itself as something that needs justification.
Write with AI. Write without AI. Code with an agent. Code by hand. Dictate your thoughts. Use an editor. It should not matter.
What should matter is whether the result is good, whether you understand it, whether you verified it, and whether you take responsibility for it.
Eventually, we will stop labeling text as "AI-generated". Saying that something was produced with AI may sound as strange as saying it was typed on a keyboard or written in Microsoft Word.
The same should happen in software.
Understanding code still matters. Fundamentals still matter. But fundamentals have always moved upward as abstractions improved.
Maybe writing code is starting to move downward.
Before we can redesign how we hire, teach, test, and evaluate people, we first need to admit that the shift already happened.
And there is an irony here.
I did not type this article from scratch. I explained these thoughts out loud. AI helped organize them and turn them into readable prose. I reviewed the result, changed it, and I am responsible for what it says.
Did AI write it?