@brandonscript @timbray @adarsh

When I see descriptions like this, a thing which comes to mind is that covid can cause brain damage. As well as LLMs diminishing the incentives to think, they might become more tempting to rely on as cognitive skills deteriorate from repeat infections.

In other words, there might be more than one thing happening.

There's a disturbing trend of developers I've known for years now suddenly being unable to discuss even basic aspects of modern software development. They have become so calcified and unable to embrace new skills relaying on orthodoxy and conservative software development, they can't even innovate or learn new, more efficient and challenging techniques. Feels like losing friends to dementia.

I don't want to know what string of tokens the LLM free associated with the question, I wanted to know what the human I asked thinks.

Cheap to sketch, often illegibly and inconsistently.

This! I have to continuously redirect LLMs to follow conventions and place code and logic in sane places so the project structure remains solid. Otherwise, everything would be randomly scattered throughout the project.

Anyway, I've done that 3-4 times. Its not like it's my regular workflow yet, but its something I've generally been happy with.

@headius

My colleagues (I have not conducted an interview in a while) report that candidates do seem "dumber" on average, than they once were, but of course it's anecdotes and feelings not hard data.

(I use LLMs and don't feel I'm getting less capable, but maybe I am)

This works because before, I’ve planned things out in great details, with the LLM’s help. I’ve taken many big decisions up-front.

It works, but the LLM must populate its context, so the first few turns are slow. Then, it gets better.

Cheap to sketch, often illegibly and inconsistently.

Static analysis, performance monitoring, continuous integration, and ops visibility have to be up to the load of doubtful application code.

Not to leave aside the importance of specific and legible tests of system transitions and invariants.

Perhaps you will be able to avoid that slow decay by only letting LLMs focus on the small pieces while you decide how they should fit together. That's not the trend I'm seeing.