After Codex, Does 'I Can't Code' Still Hold Up?

From 'I Don't Get It' to 'Maybe I Can Try'

Having observed the AI industry for several years now, the hardest change to quantify has never been a model's benchmark scores — it's been the psychological distance ordinary people feel toward technology.

Since OpenAI integrated Codex into ChatGPT, that distance has been shrinking in a way that's understated yet remarkably substantive.

Codex's core capability isn't new — it's a model built on the GPT-4 architecture, optimized for coding tasks, capable of understanding natural language instructions and converting them into executable code. But what's truly worth noting is what happened when it was embedded into the ChatGPT interface: it stopped being merely an accelerator for engineers and began giving people with no programming background whatsoever the thought, "Maybe I could give this a try."

That psychological shift is the real product breakthrough.

What Is Codex Actually Doing?

Concretely, Codex's integration within ChatGPT allows users to describe what they need in plain language, and the system generates the corresponding code — then actually executes it in a sandbox environment and returns the results. This means the interaction no longer stops at "here's some code, go try it in your IDE." Instead, the entire loop from description to execution is completed within the conversation interface itself.

According to test data published by OpenAI, Codex scored 72.1% on SWE-bench Verified — a benchmark specifically designed to evaluate an AI's ability to solve real-world software engineering problems. The top models from the same period last year scored around 40%. This isn't incremental progress; it's a leapfrog in capability.

But the numbers are just backdrop. What's genuinely worth discussing is what this capability leap means for non-engineers.

The Division of Labor in Human-AI Collaboration Is Being Redrawn

In the past, when we talked about human-AI collaboration, the assumed script was: humans handle "figuring out what needs to be done," and machines handle "running known solutions faster." Engineers remained indispensable translators, converting business requirements into a language machines could understand.

After Codex, that line is beginning to blur.

Take a marketing professional who, in the past, wanted a simple data analysis script. Their path looked like this: clarify the requirements → find an engineer → wait for a slot in the schedule → revise requirements → wait again. Now that path might look like: clarify the requirements → describe it to ChatGPT → get an executable result → adjust it yourself.

This doesn't mean engineers will disappear — I've argued this point repeatedly in previous articles and won't rehash it here. But it does mean the threshold structure of "programming ability" has been reorganized. Previously it was binary: you could code or you couldn't. Now a third state is emerging: you can't really code, but you can collaborate with AI well enough to solve the problem.

This third state is precisely the social effect of Codex most worth watching.

A New Track for Non-Engineers

I've encountered several concrete cases of non-technical users producing real output with Codex-type tools. A legal professional used it to write a Python script that automatically organizes contract clauses. A teacher built a simple grade-tracking tool. A designer generated a command to batch-rename files automatically.

Not one of these people intends to become an engineer. They simply solved a problem that previously required waiting for someone else to solve it for them.

The flip side of this phenomenon: for engineers, Codex is also changing the texture of the work. Repetitive boilerplate code, basic debugging, documentation writing — tasks that used to consume enormous amounts of time are beginning to be absorbed by AI. What's left for engineers is more architectural decision-making, logic validation, and cross-departmental communication. The capability boundary is being pushed outward from both ends simultaneously.

The Prerequisite for Collaboration: You Need to Know What You Want

There's a premise that often gets overlooked: the quality of Codex's output depends heavily on the quality of your input.

"Write me a program" versus "Write me a Python script that reads a CSV file, filters specific columns, and exports the result to a new Excel file" yield dramatically different results. This isn't a weakness of Codex — it's the nature of human-AI collaboration. The machine amplifies your clarity, and it amplifies your vagueness in equal measure.

What does this mean? Structured thinking and the ability to articulate requirements clearly actually become more important with the rise of AI, not less. The problem-decomposition and input-output-definition mindset that engineers develop through their training remains a scarce asset in the age of AI collaboration — it's just that its owners are no longer limited to engineers alone.

Redefining Your Relationship with Code

I believe the most important significance of Codex entering ChatGPT isn't that "AI is better at writing code now." It's that it's making the question of "can you code or not" increasingly non-binary.

To those who are still on the fence, I want to say this: the cost of entering this collaborative mode right now is lower than it has ever been. You don't need to learn Python first. You just need to learn how to articulate your problem clearly first.

From my observation, the aspect of Codex most worth tracking over time is its accelerating effect on "democratizing technology" within organizations. When code is no longer the exclusive language of a specific job function, the cost of cross-departmental collaboration, the degree of reliance on outsourcing, and the speed of product iteration will all shift accordingly.

This change is already happening — and at a pace faster than most people expect.