Thirty-year veteran software engineer Christoph Mütze shipped a 25,000-parameter transformer that runs on a stock Commodore 64, complete with an exhaustive test harness and a stack of reference implementations that all have to agree before anything ships. He called it SoulPlayer. In return he got called a vibecoder. Same reflex as the Monet pile-on: label first, verdict next, evidence optional. His response is the takedown of the “vibecoded slop” accusation I’d been waiting for somebody to write, and it lands on a single question that nobody on the accusing side wants to answer:

If vibecoding is what you say it is, if AI does the hard part, if the human just prompts and ships, if expertise is no longer a moat, then the world should be drowning in proper software right now. Not slop. Real tools. The kind people pay for, depend on, use every day. Two years of access. Millions of people with the models. The barrier supposedly fell. …where is everything?

David Pierce, catalogued the bespoke micro-apps people are building for themselves: family budget trackers, fantasy baseball rank engines, migration logs with a total addressable market of one. That’s real, and it’s the right scale to celebrate. But Mütze is asking a different question: where is the vibecoded Photoshop? Where is the vibecoded Maya, the vibecoded Blender, the vibecoded compiler that compiles itself? If the prompt-and-ship cartoon were true, two years in we’d have an avalanche of sophisticated tools built by people who don’t know how to code. We don’t. The category is empty. Mütze’s diagnosis of why is the part I want every designer reading this to take in:

Level 1 is what the industry usually calls coding. The syntax, the loops, the years memorizing pointer arithmetic and which header file the function lives in. LeetCode-measurable. The job interview essence. The mechanical part. The typing.

Level 2 is flow. What you do with Level 1. Knowing the right data structure. Knowing which ugly pragmatic solution to ship instead of the beautiful academic one. Reading other people’s code. Taste and judgment. The reflex of rejecting solutions that almost work and shipping the ones that do. Debugging, unit testing, the quality-control part.

Level 3 is architecture. The macro decisions, made with full awareness of their consequences. What to build at all. Why this data structure and not that one. Why this trade-off and not the obvious one. Which design survives contact with the real world, and which one silently falls apart two years later. The deciding part.

The three have never been the same thing. The gate was never at Level 1. The gate was at Levels 2 and 3, where the work that holds together actually happens. AI lowered the cost of Level 1. It didn’t really touch Levels 2 or 3. The gate is exactly where it always was.

You can easily translate this framework from engineering to design. Level 1 in design is pushing pixels: the auto-layout setup, the icon nudging, the variant-matrix work in Figma that fills our days. Level 2 is the taste that picks which of the fifteen generated directions is actually worth shipping. Level 3 is deciding what to build at all, and for whom. AI is eating Level 1 in design the same way it has eaten Level 1 in code. The designers who panic about “vibecoded design” are panicking because Level 1 was the layer they could see, measure, and defend. The gate is somewhere else, and it always was.

The reason this gets so emotional is the part Christopher Butler has been pointing at for a while: AI doesn’t just replace tools, it renegotiates what made you worth hiring. Mütze says the same thing:

The accusers cannot see this. They are not at the gate. They were at Level 1. Level 1 was their identity, their hours, their proof of belonging, their reason to feel at home in this profession. When AI made Level 1 cheap, it did not threaten the gate. It threatened them. Because they bet their self-worth on the layer that just got rented out. So they call the work vibecoded. They have to.

Mütze could weaponize the accusation back. He has the receipts: the test harness, the reference implementations, thirty years on the demoscene. He refuses and ends with a call-to-action:

If you’ve been sitting on something you made with AI, ship it. Name your tools. Don’t apologize. The accusation is cheaper than the work. Yours is worth more.