Almost No Skill Required to Cook a Steak (Though You Probably Can't Make a Decent One)

There's a metaphor buried in this piece that's simple enough to steal, and honest enough to sting a little. Cooking a steak takes almost no skill. Anyone can put meat in a hot pan and produce something edible. Getting a steak that's actually good, evenly pink, properly seasoned, right every single time, takes real practice. AI-assisted coding, the argument goes, works exactly the same way.

Right now developers are throwing everything they have at models: agents, prompts, tool chains, elaborate multi-step workflows, all in the hope that something great falls out the other end without anyone having to understand the mechanics underneath. Sometimes it works. Often it doesn't. The model hands back something confidently wrong, technically plausible, functionally broken, and it does so with zero hesitation, the coding equivalent of serving burnt meat and calling it medium-rare.

The natural response is to outsource the problem. Pay for a premium tool, switch frameworks, hire an agency, hope someone else already cracked it. But the piece makes a sharper point here: most of these products are running the same underlying models, the same steak machine, just with a nicer dining room around it. Cost-cutting teams tell themselves customers won't notice the difference. Often they're right, because most software only needs to be tolerable, not excellent. Bugs get shrugged off, weird UI gets tolerated, glue-code nobody understands keeps humming along in production.

The piece draws a clean line between AI as an amplifier and AI as a replacement for judgment. A model can follow instructions fast, at scale, tirelessly. What it can't do is know what you actually wanted unless you've already translated that vision into constraints, tests and feedback loops it can act on. And even then it's limited by context windows and whatever scaffolding surrounds it. Standing over it correcting every output might patch things temporarily. It doesn't turn a fast follower into someone with taste.

So the conclusion isn't anti-AI, it's pro-competence. Use the tools to move quicker, automate repetition, get a first draft going. But the responsibility for knowing what 'good' looks like, and catching the moment something is technically fine but wrong in every way that matters, still sits with the human. Skip that step and you're just hoping to get lucky with dinner, over and over, forever.