When AI Stops Being a Code Completion Tool and Starts Doing Your Job

Last month, a junior developer on my team asked me point-blank: "If AI can write code, why are you still here?" I laughed it off. Then I read about OpenAI and Molecule.one getting GPT to autonomously optimize a medicinal chemistry reaction, and the question didn't seem so funny anymore.

Here's the thing that got under my skin: this isn't about AI writing boilerplate or fixing syntax errors. This is about AI taking a genuinely difficult, domain-specific problem—one that requires deep chemical knowledge, iterative experimentation, and intuition—and improving it without a human in the loop. That's a different category of capability than I expected we'd see this soon. And if it's happening in chemistry labs, I need to think seriously about what it means for software development.

The Autonomous Part Is What Matters

Let me be clear about what actually happened here, because the headlines were misleading me at first. It's not that GPT wrote a formula and chemists applied it. The system was given a challenging synthetic route—specifically, improving a reaction that was critical to manufacturing a drug. Then it autonomously iterated: proposing modifications, reasoning through chemical principles, evaluating constraints, and refining its approach. The loop ran without human intervention.

That's the part that caught my attention. Autonomous means the AI didn't just generate suggestions for humans to pick through. It had enough chemical understanding to navigate a genuinely constrained problem space and make real improvements.

Why This Feels Different From Code Generation

I've been using AI tools for coding daily for two years now. Claude writes solid helper functions. ChatGPT helps me debug. But there's always a human gating layer—me, reviewing and validating. When I use these tools, I'm still the decision-maker. The AI is a consultant, not an agent.

What's happening in chemistry labs is different. The AI is setting direction, not just generating options. It's operating within constraints that require genuine understanding, not just pattern matching against training data. And critically, it's producing measurable improvements on problems where the "right answer" isn't a matter of opinion—you can literally test whether the reaction works better.

I kept asking myself: where's the analogy in software development? And I think it's not code generation at all. It's more like asking an AI to autonomously refactor your entire codebase to reduce latency by 20%, and it actually does it, and it's correct. That would be genuinely transformative.

The Honest Gap in My Thinking

I've been somewhat dismissive of "AI replacing developers" discourse. My argument was always: building software requires understanding user problems, making tradeoff decisions, owning outcomes. An AI can't do that alone. But I'm realizing I might be conflating "AI can't do my whole job" with "AI can't do the hard parts."

The chemists in this story still exist. The work still requires human direction, validation, domain expertise. But the cognitive load just shifted dramatically. Humans now oversee rather than execute. They've moved up the stack from "figure out the reaction pathway" to "did the autonomous system accomplish what we needed?"

In development, the equivalent would be moving from "write the database layer" to "did the AI-generated service properly isolate concerns?" That's a real change in what we're doing, even if we're still employed.

What Happens When This Reaches Software

Here's what keeps me up at night: chemistry has constraints. Physical reality enforces correctness. A reaction either works or it doesn't. Software is messier. A system can pass tests, meet requirements, and still be architecturally wrong in ways that matter a year from now.

But that's also not a guarantee we're safe. Autonomous systems in other domains have learned to game their constraints. An AI that's incentivized to "improve code quality metrics" will find ways to technically satisfy that goal while creating unmaintainable abstractions.

I'm not panicking. I'm paying attention. And I'm building differently now—focusing deliberately on the kinds of decisions that require judgment beyond pattern matching. Systems design. User psychology. Architectural decisions with long time horizons. These still feel like places where human brains add irreplaceable value.

Your Turn

Are you noticing this gap between "AI helping with execution" and "AI operating autonomously in your domain"? And more importantly: if it comes for software the way it came for chemistry, what part of what you do feels safest?

Source: This post was inspired by "A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry" by OpenAI Blog. Read the original article