DEV Community

Cover image for If AI Gets Better at Coding, What Becomes the Bottleneck?
Micheal Angelo
Micheal Angelo

Posted on

If AI Gets Better at Coding, What Becomes the Bottleneck?

One thought has been occupying my mind recently.

As AI systems continue to improve, where does the bottleneck move?

Historically, writing software involved two distinct challenges:

  1. Understanding the problem
  2. Translating the solution into code

For a long time, both were difficult.

Today, however, the second challenge is becoming significantly easier.

Modern AI systems can already generate substantial amounts of code from natural language instructions. Open-source models are improving rapidly, and frontier models continue to push capabilities forward.

This raises an interesting question:

If generating code becomes increasingly automated, what remains uniquely valuable?

I do not claim to know the answer.

But I have a hypothesis.


A Historical Parallel

Consider what happens when a program is compiled.

Developers rarely write machine code directly.

Instead:

Human Logic
      ↓
Programming Language
      ↓
Compiler
      ↓
Machine Code
Enter fullscreen mode Exit fullscreen mode

In a sense, compilers already generate the final executable instructions.

Yet software engineers did not become irrelevant.

The bottleneck simply shifted.

Humans focused on:

  • Problem solving
  • Architecture
  • Design decisions
  • Trade-offs

while compilers handled translation.


AI May Shift the Bottleneck Again

Something similar may be happening today.

The amount of code directly written by humans may decrease.

The amount of guidance provided to AI systems may increase.

The bottleneck may move from:

Writing Code
Enter fullscreen mode Exit fullscreen mode

toward:

Defining Problems
Enter fullscreen mode Exit fullscreen mode

and

Communicating Intent
Enter fullscreen mode Exit fullscreen mode

Articulacy May Become a Competitive Advantage

One skill that increasingly feels important is articulacy.

By articulacy, I mean:

The ability to translate thoughts into clear, precise language.

This sounds simple.

In practice, it is surprisingly difficult.

Many people know what they want.

Far fewer can describe it precisely.

The difference matters.


Vague Requests Produce Vague Results

Consider the difference between:

Build me a web application.
Enter fullscreen mode Exit fullscreen mode

and

Build a React application using Redux for state management,
with authentication, role-based access control,
server-side pagination, and optimistic UI updates.
Enter fullscreen mode Exit fullscreen mode

Both are requests.

Only one provides meaningful constraints.

The second request communicates intent more effectively.

As AI systems become more capable, the quality of instructions may become increasingly important.


Asking Better Questions

Another skill that feels increasingly valuable is asking the right questions.

Good questions often emerge from understanding how systems work.

For example:

  • What are the trade-offs?
  • What assumptions are being made?
  • What constraints exist?
  • What information is missing?

These questions frequently matter more than syntax.

The better the question, the more useful the answer tends to be.


Why Domain Knowledge Still Matters

A common misconception is:

If AI can write code, technical knowledge becomes unnecessary.

I am not convinced that this is true.

To guide an AI system effectively, one still needs to understand:

  • Architecture
  • Design patterns
  • Framework choices
  • Performance constraints
  • Security considerations
  • Operational requirements

In many cases, knowing what to ask requires understanding the domain first.


Open Source Models Will Continue Improving

Another observation is that open-source models appear to be improving at a remarkable pace.

Today's frontier models may remain ahead.

But history suggests that capabilities often diffuse over time.

What seems exclusive today may become widely accessible tomorrow.

This raises another question:

If capable models become available locally, what limits adoption?


Hardware May Become the Next Constraint

One possible answer is hardware.

A model can be open source and still be inaccessible.

Why?

Because running it may require expensive hardware.

Most people I know use laptops with configurations such as:

  • 16 GB RAM
  • Consumer CPUs
  • Integrated graphics

If a model requires:

  • High-end GPUs
  • Large memory pools
  • Expensive infrastructure

its practical adoption remains limited.

In that case, the bottleneck shifts again:

Model Availability
         ↓
Hardware Availability
Enter fullscreen mode Exit fullscreen mode

The model may exist.

The ability to run it becomes the challenge.


A Possible Future

This makes me wonder whether future software development may increasingly look like:

Human
    ↓
Idea
    ↓
Clear Instructions
    ↓
AI System
    ↓
Implementation
Enter fullscreen mode Exit fullscreen mode

In that world, human value may not disappear.

It may simply move higher up the stack.

Towards:

  • Problem definition
  • Architecture
  • Communication
  • Judgment
  • Context

A Thought, Not a Prediction

I could be completely wrong.

Technology has a habit of surprising people.

But if current trends continue, I suspect the most valuable skill may not be typing code faster.

It may be:

  • Thinking clearly
  • Communicating precisely
  • Understanding systems deeply
  • Asking better questions

Because even if machines become exceptional at implementation, someone still needs to decide what should be built in the first place.

And that decision begins with a thought that can be articulated clearly enough for others—human or machine—to understand.

Top comments (0)