Post

Colin McMillen
@mcmillen.dev

co-signed. a robot might be able to type rote code faster than me, but *typing speed* is not the limiting factor in my effectiveness as a software engineer, and never has been. *understanding* the code (and the pre-existing human and computational systems it interacts with) is the hard part.

"Online Rent-a-Sage" Bret Devereaux@bretdevereaux.bsky.social

I'd also identify a third position, "I see or imagine there are meaningful uses for LLMs, but that they are relatively few (not 'everything') and relatively small (+10%, not +10000%) and so may not support the cost of model building and inference or the valuation of these companies."

7 reposts28 likes

Colin McMillen
@mcmillen.dev

& the human judgment of what software *should* do is the most important thing for the success of a project. so yeah, maybe with reliable AI coding assistance i could have make the easiest part of my Staff Software Engineer job go slightly faster, but that's not 90% of what i was getting paid for.

GoldenAgeWeaver
@goldenageweaver.bsky.social

LLMs are already better at understanding whatever amount of code can fit in context. Tools like Codex are decent for understanding today. And it's only going to get better. I wouldn't bet against AI becoming able to ingest and convey understanding for codebase sized amounts of data

GoldenAgeWeaver
@goldenageweaver.bsky.social

The hard step as I see it for creating a software engineer LLM is allowing it to evolve software in a way that preserves understandability. Do LLMs have enough theory of mind to be able to design systems that humans can understand? If you think exclusively in "best practices" you will make a mess

Propriety
@proprietous.bsky.social

Writing mediocre code is easy! Truly understanding code is very very hard! It's really baffling to me that so many software engineers are acting like a machine that writes a high volume of mediocre code is good. It's not! It makes our jobs HARDER, not easier!!!