The Butt Ugly Weblog

(Posted this on Bluesky, so the story is a bit jagged, as it was originally composed into 9 separate posts.)

I've been using AIs as a programming aide as I am delving into a slightly unfamiliar environment, and I'm realizing I'm not getting as deep of an understanding of the underlying system as before.

I've seen people lament this. "AI creates bad code and programmers don't know how to program anymore."

I don't think it's a bad thing necessarily. We've seen this before when we moved from assembly and low-level languages to higher level languages.

I remember people lamenting how programmers no longer understand CPUs and how they treat CPU and memory like infinite resources. And how owning your own hardware is much better price/performance ratio than going cloud.

Yes. We (engineers) exchanged performance for development speed. We evolved the software engineering discipline to deal with the situation. Most of us now think in components and microservices instead of L2 caches.

Some people still live close to hardware and understand how it works. But it's a specialty now, unlike in the 80s.

So now I have a feeling we're doing the same kind of a jump: We're dumping performance again for efficiency.

Now, compilers for high-level languages like Rust (or even Java) have pretty much caught up with the modern processor architectures and they're not actually that terribly far off from hand-optimized code.

So I think this is going to happen with AIs as well.

Possibly it's going to need more than LLMs, but right now LLMs are like compilers that turn your intention into maybe okay code. We are still badly lacking in the necessary engineering disciplines to deal with all this I think, starting from reproducibility, QA, ops, or even version control.

But I think we can sort of see the light at the end of the tunnel. Power requirements for individual models are going down, and specialized hardware is being born. Millions of engineers are thinking about these things every single day, as opposed to thousands of data scientists two years ago.

Engineers are a powerful, fairly unstoppable force that tends to flow through any crack if it sees a treat on the other side of the wall.

With AI, we see the treat. We just need to see the threats too.