The first time I used ChatGPT to code, back in early 2023, I was reminded of “The Monkey’s Paw,” a classic horror story about an accursed talisman that grants wishes, but always by the most malevolent path — the desired outcome arrives after exacting a brutal cost elsewhere first. With the same humorless literalness, ChatGPT would implement the change I’d asked for, while also scrambling dozens of unrelated lines. The output was typically over-engineered, often barnacled with irrelevant fragments of code. There were some usable lines in the mix, but untangling the mess felt like a detour.
When I started using AI-assisted tools earlier this year, I felt decisively outmatched. The experience was like pair-programming with a savant intern — competent yet oddly deferential, still a tad too eager to please and make sweeping changes at my command. But when tasked with more localized changes, it nailed the job with enviable efficiency.
The trick is to keep the problem space constrained. I recently had it take a dozen lines of code, each running for 40 milliseconds in sequence — time stacking up — and run them all in parallel so the entire job finished in the time it used to take for just one. In a way, it’s like using a high-precision 3D printer to build an aircraft: use it to produce small custom parts, like hydraulic seals or O-rings, and it delivers flawlessly; ask it for something less localized like an entire cockpit, and you might get a cockpit-shaped death chamber with a nonfunctional dashboard and random knobs haphazardly strung together. The current crop of models is flexible enough for users with little-to-no coding experience to create products of varying quality through what’s called — in a billion-dollar buzzword — vibe-coding. (Google even released a separate app for it called Opal.)
Yet, one could argue that vibe-coding isn’t entirely new. As a tool for nonprofessionals, it continues a long lineage of no-code applications. As a mode of programming that involves less prefrontal cortex than spinal reflex, any honest programmer will admit to having engaged in a dishonorable practice known as “shotgun debugging.” Like mindlessly twisting a Rubik’s Cube and wishing the colors would magically align, a programmer, brain-fried after hours of fruitless debugging, starts arbitrarily tweaking code — deleting random lines, swapping a few variables, or flipping a Boolean condition — re-runs the program, and hopes for the correct outcome. Both vibe-coding and shotgun debugging are forms of intuitive flailing, substituting hunches and luck for deliberate reasoning and understanding.
We’ve used machines to take the load off cognition, but for the first time, we are offloading cognition itself to the machine.
As it happens, it’s not considered good form for a self-respecting programmer to engage in shotgun debugging. Soon, I came to see that the most productive form of AI-assisted coding may be an editorial one — much like how this essay took shape. My editor assigned this piece with a few guiding points, and the writer — yours truly — filed a serviceable draft that no sober editor would run as-is. (Before “prompt and pray,” there was “assign and wait.”)
Likewise, a vibe-coder — a responsible one, that is — must assume a kind of editorship. The sprawling blocks of code produced by AI first need structural edits, followed by line-level refinements. Through a volley of prompts — like successive rounds of edits — the editor-coder minimizes the delta between their vision and the output.
Often, what I find most useful about these tools isn’t even writing code but understanding it. When I recently had to navigate an unfamiliar codebase, I asked for it to explain its basic flow. The AI generated a flowchart of how the major components fit together, saving me an entire afternoon of spelunking through the code.
I’m of two minds about how much vibe-coding can do. The writer in me celebrates how it could undermine a particular kind of snobbery in Silicon Valley — the sickening smugness engineers often show toward nontechnical roles — by helping blur that spurious boundary. But the engineer in me sees that as facile lip service, because building a nontrivial, production-grade app without grindsome years of real-world software engineering experience is a tall order.
I’ve always thought the best metaphor for a large codebase is a city. In a codebase, there are literal pipelines — data pipelines, event queues, and message brokers — and traffic flows that require complex routing. Just as cities are divided into districts because no single person or team can manage all the complexity, so too are systems divided into units such as modules or microservices. Some parts are so old that it’s safer not to touch them, lest you blow something up — much like the unexploded bombs still buried beneath European cities. (Three World War II-era bombs were defused in Cologne, Germany, just this summer.)
... continue reading