[hparadiz]

Metrics, profilers, architecture! Use AI to get back to basics! Wanna prove a technique is better? Use AI to make a benchmark! Learn by experimentation! That is my advice to juniors. At the end of the day AI is writing code and there may be 10 different ways to run something. Only one is the fastest in any given use case.

[chowells]

Fastest is usually the wrong metric. But you'd need experience to know that, I suppose...

[mikepurvis]

But the point is that LLMs giving you 10x the potential code output doesn't have to mean 10x the code committed. It can also be "let's look at all three possible implementations in more detail and decide which is really the best fit for our situation, and commit that one."

That's still 2-3x the velocity, but you get a better result because you went deeper on the paths-not-taken when designing.

[steezeburger]

I think it's just the wrong metric to optimize for _first_. It's not generally a bad metric to keep tabs on though. Make it work, make it right, make it fast? Or something like that.

[steezeburger]

Yeah I totally agree! I also think people should still be reading as much code as they can. That's always been true imo. It is just hard to keep up with it now because of how much code an LLM can generate for $20/month. I do think we'll move to higher abstractions of course. We won't have to understand code as much as how the systems and components are architected. It would also be nice to use our new efficiency to return to producing truly optimized and fast software.