o3-mini on a Codebase Is Genuinely Unsettling
It knows where the bodies are buried before you finish the sentence.
o3-mini on a codebase is still unsettling. The wall hasn't come yet. The question of what o3 full looks like on real engineering tasks is still the most important unanswered question in AI tooling.
Ran o3-mini on a real codebase today. It just — knew things. Instantly. The kind of knowing that makes you look around the room.
My previous setup was o1 plus Sonnet in Windsurf, which felt fast until today, the way a dial-up modem felt fast until you got cable. o3-mini makes that combo feel like it's chewing its food.
And I'm waiting for the wall. There's always a wall. Some moment where the model confidently hallucinates a function signature, or does the thing where it rewrites your entire architecture to fix a three-line bug. I'm sure it's coming. But it hasn't come yet, which is the unsettling part.
The thing nobody's really talking about is what o3 full — high compute, no distillation, no hedging — actually looks like on this kind of task. o3-mini is the diet version. The commercial version. The one that had its edges sanded down for throughput and pricing. It's already this. So what is o3 on high? What does "blazing fast and knows everything" look like when you uncap it?
That's the mystery beast. That's the thing sitting in the room we're not allowed into yet.
Counterpoints
Push back, extend the argument, or sharpen it. New counterpoints go through review before they show up here.
No approved counterpoints yet.