The UN Endorsed a New World Map — Then Someone Vibecoded the Proof

A map projection just became UN policy

On September 4, 2026, the UN General Assembly voted 164-1, with six abstentions, to adopt a resolution encouraging member states to use the Equal Earth projection instead of the 16th-century Mercator projection that still dominates classrooms, browsers, and news graphics. The resolution was put forward by African Union member states and the Bahamas; the United States was the only country to vote against it. The complaint is a real cartographic distortion, not a stylistic quibble: Mercator inflates land area near the poles and shrinks it near the equator, so it renders Greenland roughly the same size as Africa — when Africa is in fact about 14 times larger. The vote is non-binding; it doesn't force anyone to change a map. But France announced the same day it would abandon Mercator in its own materials, which is the kind of signal that tends to move institutional defaults.

The 48-hour vibecoded rebuttal

Three days later, Simon Willison — a developer well known for documenting how far AI coding tools have come — got curious about what Equal Earth actually looks like next to Mercator and decided to just build the answer. He prompted GPT-6 Astra (medium) inside ChatGPT Work to produce an interactive animated map built on D3 that morphs between the two projections on a slider or a play button. No repo scaffolding, no manual D3 debugging session — a conversation that ended in a working, shareable tool, published the same day as a short write-up.

Why this is a useful case study, not just a novelty

It's tempting to read "someone vibecoded a map" as a cute aside. It's more useful read as a data point about where AI-assisted development is genuinely reliable today. This task had properties that matter: it was narrowly scoped (one interaction, one visual idiom), the domain — D3 projection math — is heavily represented in public code the model has seen, the output was independently verifiable just by looking at it, and the blast radius of a bug was zero (a wrong animation, not a production incident). That combination is exactly where current frontier models close the gap between "describe it" and "ship it" fastest.

It's also a fair contrast case for anyone building trading, fintech, or on-chain systems: the same prompt-and-iterate loop that produces a correct map in one sitting does not automatically produce a correct order-execution path or a correct token-approval flow. The difference isn't the model, it's the verification story — a map projection is visually self-checking; a smart contract or a risk engine is not.

What "GPT-6 Astra (medium)" signals

The naming — a mid-tier reasoning setting inside a workspace product, rather than a raw API call — is itself a small signal of where frontier labs are pushing: tiered reasoning effort exposed directly to non-coding-focused seats, aimed at exactly this kind of one-off builder task rather than long-running agentic pipelines.

The takeaway for teams evaluating AI-assisted dev

• Look for tasks that are narrow, visually or logically self-checking, and low-stakes when scoping what to hand to an AI-assisted, low-oversight workflow.
• Treat a clean one-shot build like this map as evidence about *scope fit*, not as evidence a model can replace review on financial or security-relevant code.
• The fastest wins right now are exploratory tools and internal utilities — exactly the pattern worth practicing before applying the same workflow to anything customer-facing.