And just like sushi, the range of quality is incredibly wide.

Really bad sushi tastes awful, is pungent, and, if it’s poorly prepared while there’s bacteria or parasites present, it might even kill you!!

On the other hand, really incredible sushi is made by a master chef, takes time and energy to prepare, required years of apprenticeship to learn the necessary techniques, demands the right high quality ingredients, and can produce once in a lifetime flavors and experiences.

Really bad AI code is no different than really bad sushi.

In the short term, it can bring down your entire platform with logical errors, poorly handled runtime exceptions, misconfigurations, and code that won’t compile.

Longer term, alot of really bad AI code can start to make managing tech-debt a nightmare: abstraction after abstraction that doesn’t all fit well together, implementations that should be apart of an abstraction but instead are one off singletons, or scope that wildly creeps out of control with no consideration for the broader architecture.

In my humble opinion, really great AI code akin to really great sushi does not yet exist and likely never will: even the most sophisticated models rely heavily on generalized patterns and semantics from their training data that often times is too general to apply to really specific use-cases. The best code out there is hand crafted and designed with intent.

Most AI code I’ve seen or reviewed (which, by my own metrics, is around 90k+ lines of code and 5,000+ messages with my agent this year alone) is somewhere between “week-old leftovers” and “grab-and-go grocery store” sushi.

Don’t get me wrong: I really love a quick and easy grab-and-go sushi lunch. Who doesn’t? It’s fast, convenient, usually tastes ok, and gets the job done.

But the industry seems to be obsessed with shipping as much grocery-store grade AI code as possible. And, just like with eating way too much low quality sushi, eventually, we’re gonna get really sick. I find that it’s no coincidence that nearly every software platform and product seems to have gotten worse and more unstable over the last 5 years: the general “enshitification” of all things seems to only be accelerated by extremely mediocre AI code and integrations.

If you need really high quality code, you need a skilled professional: someone with years of experience, a taste for how something should work, a sense of true creativity and adaptability, both architecturally and semantically, and a knack for managing the software life-cycle from beginning to end.

This means hiring skilled software engineers, mentoring the next generation of jr. engineers, and bravely adopting necessary technologies regardless of how good AI is with it yet.