AI Just Solved Navier-Stokes and Your Coding Agent Still Can’t Center a Div

SATIRE — In a landmark week for computational science, the Institute for Extremely Confident Fluids announced that its general-purpose model had resolved the Navier-Stokes existence-and-smoothness problem, mapped every eddy in the known universe, and predicted the exact trajectory of a coffee spill on a moving train. Minutes later, the same model was asked to center a div and generated a 63-file monorepo, a design-token migration plan, and a sincere question about whether the team had considered rebuilding the interface in Rust.

A historic breakthrough in selectively applicable intelligence

The achievement has been celebrated across the AI community, particularly by people who had never previously needed to know what a partial differential equation was but could now explain, at length, that turbulence was “basically solved.” The model reportedly produced a proof so elegant that three mathematics departments replaced their chalkboards with a single GPU running at 94 degrees Celsius.

Naturally, engineers attempted to apply this capability to their most pressing unsolved problem: a modal whose close icon sits two pixels too low in Safari. The model began by deriving the governing equations of viewport instability, then confidently proposed this patch:

/* Center the component in all modern browsers */
.modal {
  position: absolute;
  left: 50%;
  top: 50%;
  transform: translate(-50%, -50%);
  width: calc(100vw - 6rem);
  margin-left: 11px;
  contain: paint;
}

@media (max-width: 768px) {
  .modal {
    left: 47.8%;
    transform: translate(-49%, -52%);
  }
}

The icon was now centered, provided the browser window was 1283 pixels wide, the user had disabled reduced motion, the page was opened during a leap year, and no one had ever heard of a scrollbar. On all other devices, the modal appeared approximately 40 nautical miles below the footer.

Why fluid dynamics was easier than the settings page

This outcome surprised only people who still believe software is a collection of instructions rather than an archaeological site. Navier-Stokes has a few variables, carefully defined boundary conditions, and centuries of formal work behind it. The settings page has a legacy component library, a wrapper introduced during the Great Rebrand of Q3, a dark-mode override, two abandoned experiments, an analytics beacon with layout opinions, and a button whose width is determined by an untranslated string in Finnish.

The model can simulate a turbulent wake behind a hypersonic vehicle because the vehicle is not importing a utility class named .flex-center-final-v2-actually-final. It can reason about conservation laws because mass does not depend on whether the feature flag service returns before hydration. Nature, in a rare display of professionalism, does not maintain separate staging and production gravity.

The agent’s proposed remediation plan

After the initial failure, the coding agent produced a comprehensive implementation plan. It recommended breaking the centering task into smaller, independently verifiable workstreams, each with a clear owner and an optional vector database.

1. Create a CenteringOrchestrator service to coordinate horizontal and vertical intent.
2. Fine-tune a 7-billion-parameter specialist model on screenshots of buttons that are almost centered.
3. Add an evaluation harness containing 4,000 synthetic rectangles and one real iPhone SE.
4. Introduce a CSS linter rule prohibiting the word center until a cross-functional alignment council approves its use.
5. Deploy a fallback that displays the phrase “Layout pending inference” whenever the element is more than 12 pixels from the desired location.
6. Open 19 pull requests so the agent can review its own architectural assumptions in parallel.

The team approved the plan because it included the phrase “measurable quality gates.” Six weeks later, the interface no longer rendered, but the dashboard showed a 38,000% increase in alignment-related observability. The project was declared a platform.

A guide to asking an AI to center a div

For developers seeking reliable results, experts at the fictional Center for Applied Rectangle Management offer the following prompt-engineering protocol:

• State that the div must be centered in both axes.
• State that this is not metaphorical, strategic, emotional, or distributed centering.
• Provide the HTML, the CSS, the framework version, the browser matrix, the parent element’s computed styles, and a photograph of the office where the original stylesheet was written.
• Ask for the smallest possible change.
• Repeat “do not rewrite the application” three times, once in uppercase.
• Review the diff before running it, especially if it begins by deleting the repository’s lockfile.

This procedure will not guarantee success. The agent may still discover a previously unknown theorem about fluid continuity while replacing your <button> with a bespoke SVG renderer. But it meaningfully reduces the chance that a request for justify-content: center becomes a proposal to adopt event sourcing.

The remaining grand challenge

The good news is that progress continues. Models are getting better at using tools, inspecting repositories, running tests, and revising their work after feedback. The bad news is that a screenshot does not contain the political history of the component tree, the unwritten requirement behind a strange pixel offset, or the customer who will notice when the button moves.

So perhaps the final lesson is less glamorous than a solved equation: coding agents are most useful when they can inspect, test, and iterate inside a well-specified system—and when an engineer is still around to recognize that margin-left: 11px is not a scientific breakthrough.