Two talks at Carolina Code Conference stuck out to me today. Two programming ...

Two talks at Carolina Code Conference stuck out to me today. Two programming ...

Two talks at Carolina Code Conference stuck out to me today. Two programming languages off most people's radar... one genuinely obscure (Pony), one niche but beloved (Elixir).

They essentially gave the same talk. And it's the one nobody's giving while everyone argues about which AI model is smartest.

The argument: the AI that writes good code isn't the one with the biggest brain . It's the one you've boxed in so it can't do the wrong thing.

Redvers Davies showed an AI writing his niche language correctly about 28% of the time on the first try. (There's almost none of it online, so the model borrows patterns from other languages and guesses.) Then he put the compiler in the loop. Pony already refuses to compile anything that breaks its rules, so the AI had to keep trying until it produced something legal. Correct jumped to 100%. And the cheap, "dumb" models scored just as well as the expensive ones.

His line stuck with me: your docs can say "please don't." Your instructions can say "you shouldn't." But you need something that says "no, you will not."

Paul Sullivan came at it from the opposite side: why Elixir quietly tops the AI coding benchmarks. Not for any AI reason. It has no hidden state for the model to lose track of, so there's simply less to get wrong. The same things that make it easy for a human make it easy for a machine.

Same lesson, two directions: the leverage isn't the model. It's the guardrails around it.

If you're rolling out AI at work, that changes the shopping list. Stop asking "which tool is smartest?" Start asking "what in my setup says no?" The review step nobody gets to skip. The test that has to pass before anything ships. That's the part worth paying for, and it's the part the demos never show you.

The smartest model in the world is still guessing. The boring constraint around it is what makes the guess safe.

(I'm giving a talk on a cousin of this idea tomorrow... that the real moat left in software is judgment, not horsepower.)