How Vibe Coding is like a Spherical Cow

A spherical cow is an "inside" joke amongst physicists referring to tendency to develop "toy" models that reduce the problem to the simplest form imaginable while ignoring "second order" effects.

When you use AI to vibe code it feels like a spherical cow: a beautiful solution that works in a vacuum (in fact, Claude Code will insist it is “Perfect!”).

Take this example: Your product manager asks for a file upload function. Your PM strolls in, prototype in hand.

“Lovable.dev did it in three minutes. Should be easy!”

You smile politely. Inside your head:

Functional reality check (the easy part):

• What kinds of files should we allow?
• What’s the maximum size? How many files?
• If the files are shared, do we need fine grained sharing permissions?
• Can the files be viewed in the browser?
• Do we charge them for storage? Is there a free tier? What if the user discontinues the service? Do we delete the files? Keep them around hoping they come back?
• Is there a trash bin for deleted files that they can recover from?

Then comes the non-functional reality check (the physics):

• How many users? How many file uploads per second? Do we need queues? Caches? Auto scaling?
• Where do we store this? S3? Which tier? Multi-cloud? Which regions? Do we need bucket monitoring? Encryption? What strength? What about European users? 
• What about backup? Restore? 
• What level of trust do we have in these documents? Should we scan them for viruses? CSAM? How often? At what cost? What if we get subpoenaed for the files?
• We need an audit trail, probably. Where do we store this?

And then the unspoken layer — the stuff that never makes it into a requirements doc:

• Should we roll this feature into the multi-cloud migration effort we’re starting next month?
• Our cloud security person left and we haven’t found a replacement.
• Is this a new microservice? Should we take this opportunity to use a more modern language? We need to purchase the tooling for that technology then. Do we have the budget?
• Is this  a good time to replace some of those outdated React libraries?

And on, and on, and on. Compliance. Disaster recovery. Budget. Staffing.

Sure AI can help create code for some of these constraints. But the judgement and taste is up to you. The AI has no idea about your environment, your constraints, your business priorities. Heck, even you may not consciously know everything, just relying on years of experience and hallway conversations to come up with a practical solution. And even if you were able to convey all these constraints perfectly to the AI, it is not trained to make these decisions - though it may offer very confident answers.

So yes, AI will give you elegant solutions for spherical cows in vacuums.

But the real work — the engineering, the tradeoffs, the quiet judgment calls — happens back on the messy farm, with real cows, real mud, and real deadlines.

AI will keep getting better at writing code. But judgment — that ability to balance the unquantifiable — will remain the rarest form of compute we have (fingers crossed). 

And, as your team adopts AI coding tools - where have you seen spherical cows in your own stack lately?