Comment by onlyrealcuzzo
4 hours ago
My personal take is that if you want to scale in engineering you really only have 2 options.
1) you start acting like a senior staff level engineer and you stop caring so much about implementation and start caring about the things that actually matter, the architecture, the performance profile, the scalability, the test coverage, the monitoring, the velocity of development, etc etc.
Or 2) you don't scale.
LLMs can write 10k lines of code per day like its nothing. You can't review that with a jade handled magnifying glass and keep all that in your mind. After a few months, you could easily have 250k lines of dense production code + much more in tests and tooling etc.
You either have a team of people to watch over that, or you don't. You're not scaling yourself to be able to review all that and keep it all in your mind.
If other people are scaling, and you're not, you're unlikely to do as well as them. BUT, it's not impossible if you're really good. Being really good scales way less than using LLMs effectively.
Google doesn't work because they rely on hope that every person they hire doesn't do things they shouldn't. It works because they make it nearly impossible to do things you shouldn't. You need to build systems as well. Or you don't. Just do it the old way.
Pick your poison.
The problem I see is that in a traditional company a senior engineer would push back on complexity and know when to do due diligence.
AI seldom pushes back, embraces complexity, does a lot of due diligence (however often superficial), and then completely misses the point and lays a trap on the code.
It’s not nearly as trustworthy as an actual team of humans.
> you stop caring so much about implementation and start caring about the things that actually matter, the architecture, the performance profile, the scalability, the test coverage, the monitoring, the velocity of development
What architecture doesn’t care about implementation? What performance profile doesn’t care about the details of the implementation? What does the test coverage is about if it’s not the implementation? What are you monitoring? The specs or the implementation?…
Not caring about the implementation seems like not caring about the materials when building an house, with eyes stayed only glued to a novel’s like description of the house.
I think you misunderstand - it’s not that you shouldn’t be capable of evaluating and improving the implementation, but that it’s usually irrelevant. A very large majority of the time, “good enough” is fine. Fix things people complain about, and use the rest of your time to do more important things.
and this is why everything needs 32+GB of RAM and an assload of storage.
LLMs don't have human attention to detail, IME