Humans still write the code, we just do it using AI tools.

If we abdicate and let all software development and operations occur through the use of fully autonomous agents, someone is still ultimately responsible for the agents work/screw-ups. Thus, there is still a human writing code, from an ownership perspective. We all know it won't be the capital/manager class holding the stick there.

> Humans still write the code, we just do it using AI tools.

That's literally the opposite of what TFA quotes in Proton's "Engineering Manifesto"

> Your job has fundamentally changed. It is no longer to write code, and LLMs are not here to help you write better code faster. Your job is to help the LLM write better code faster

It goes on to say

> Every engineer is now a team leader, a manager, and your direct reports are the models and agents you use.

They don't want their engineers and programmers to write the code. They want LLM handlers.

>If we abdicate and let all software development and operations occur through the use of fully autonomous agents, someone is still ultimately responsible for the agents work/screw-ups. Thus, there is still a human writing code, from an ownership perspective.

That isn't what "writing code" means. The responsibility is for using the AI, or being negligent in doing so, but the AI is obviously writing the code. That's the entire point of it.

Are compilers writing code? I mean, people hadn't been producing anything that a CPU is actually executing (excepting some very rare special cases) for decades now. So we never wrote code to begin with...

A compiler works much more deterministic though

"More" - yes, "deterministic" - well... Have you looked at what modern optimizing compilers do to the source? And how they behave in those places described as "undefined behavior" by the standards? Could you predict, looking at the code you've supposedly written, something like Meltdown or Spectre? We have to face it - there are many layers of complexity between what we feed into the machine and what the machine actually does, and for many of them most programmers have but a passing knowledge. They don't care because it just works, so why bother? But for somebody writing in something as high-level as, say, Python, to puff their chest and proclaim "in our time, we used to strike out bits on a rock with chisel and hammer! Kids these days!" is not exactly appropriate. Yes, LLMs work differently from even the cleverest optimizing compiler or interpreter. But they are still tools for translating programmer's intent into machine actions.

...no? That's just called being a crumple zone.

I have a feeling the same sentiment will ring true across most of the software industry in the next couple years. LLMs are cool, but they don't perform better than purpose built solutions at scale.

You can use LLMs to build purpose-built solutions. You can also use LLMs to automate what used to be human tasks.

I disagree. It works with the wow gold farmers because there is a finite set of known states in a videogame. In real life I don't think that's the case.

You can automate the first 3 or 4 nines the old-fashioned way. I guess you can escalate up to an LLM instead of directly to a human in the remainder of cases.