Cloudflare built “vinext” in less than a week for about $1100 in AI tokens.
If you missed the story: vinext is a functional implementation of Next.js, the huge web framework originally created by Vercel.
Seeing somebody recreating the core behavior of such a huge, complex project so quickly and cheaply leads to one inevitable conclusion:
𝐈𝐭’𝐬 𝐭𝐢𝐦𝐞 𝐭𝐨 𝐫𝐞𝐭𝐡𝐢𝐧𝐤 𝐭𝐡𝐞 𝐯𝐚𝐥𝐮𝐞 𝐨𝐟 𝐄𝟐𝐄 𝐭𝐞𝐬𝐭𝐬.
Suddenly, a project that might have cost millions of dollars to build can be recreated without fully understanding its internals. You don’t need to know exactly how the system works under the hood. You just need to know its behavior.
And if a system’s behavior is well described by:
- A public API
- Clear documentation
- 𝐀𝐧𝐝, 𝐚𝐛𝐨𝐯𝐞 𝐚𝐥𝐥, 𝐩𝐮𝐛𝐥𝐢𝐜 𝐞𝐧𝐝-𝐭𝐨-𝐞𝐧𝐝 𝐭𝐞𝐬𝐭𝐬
…then these tests stop being just a quality tool. 𝐓𝐡𝐞𝐲 𝐛𝐞𝐜𝐨𝐦𝐞 𝐚 𝐬𝐭𝐫𝐚𝐭𝐞𝐠𝐢𝐜 𝐚𝐬𝐬𝐞𝐭.
In practice, today they are something very close to an executable product specification. A blueprint of the system’s behavior. ( 𝘈𝘯𝘥 𝘕𝘦𝘹𝘵.𝘫𝘴 𝘩𝘢𝘴 𝘢 𝘭𝘢𝘳𝘨𝘦 𝘯𝘶𝘮𝘣𝘦𝘳 𝘰𝘧 𝘌2𝘌 𝘵𝘦𝘴𝘵𝘴 )
This is one of the many shifts LLM AI has brought to software engineering.
Previously, public tests said: “Look, we care about quality.” Today, they also say: “Look, this is exactly how to rebuild our product.”
This dramatically lowers the cost of competitive reimplementation. We are going to see a flood of similar stories:
- Someone rewriting Project X in Rust.
- Someone making a compatible implementation of Project Y.
- Someone recreating Project Z’s API to be 5x faster in a specific scenario.
This doesn’t mean you can easily copy an entire business with an LLM. But the barrier to cloning software’s exact behavior is dropping to near zero. (if the project is public and has tons of E2E tests) . You need an advantage that cannot be easily recreated just by looking at how the system behaves.
💡 𝐓𝐡𝐞 𝐎𝐩𝐩𝐨𝐫𝐭𝐮𝐧𝐢𝐭𝐲: If E2E tests are a blueprint for AI, they also make internal rewrites drastically cheaper and safer. 𝐓𝐡𝐞𝐲 𝐚𝐫𝐞 𝐭𝐡𝐞 𝐞𝐱𝐚𝐜𝐭 𝐩𝐫𝐨𝐦𝐩𝐭𝐬 𝐲𝐨𝐮 𝐧𝐞𝐞𝐝 𝐭𝐨 𝐥𝐞𝐭 𝐀𝐈 𝐫𝐞𝐰𝐫𝐢𝐭𝐞 𝐲𝐨𝐮𝐫 𝐩𝐫𝐨𝐝𝐮𝐜𝐭 𝐰𝐢𝐭𝐡𝐨𝐮𝐭 𝐥𝐨𝐬𝐢𝐧𝐠 𝐛𝐮𝐬𝐢𝐧𝐞𝐬𝐬 𝐥𝐨𝐠𝐢𝐜.
E2E tests are better then code, as legacy code is not just business logic. It is a museum of history. It contains 10 years of hacks, hotfixes, workarounds and outdated architectural decisions.
So maybe Code → E2E → Code could extract the business value and leaves the technical debt behind. But who gonna write so many E2E tests AI?