TL;DR: The biggest threat to your career is not LLMs but it's younger engineers that will adapt to the new tools. --- My personal take is that LLMs and near future evolutions thereof won't quite replace the need for a qualified human engineer understanding the problem and overseeing the design and implementation of software. However it may dramatically change the way we produce code. Tools always begat more tools. We…
I'm actually interested in what you think AI will do as this is quite vague. > "AI" tools offer an opportunity to change the way we do those things. The way I see tooling in programming is that a lot of it ends up focused on the bits that aren't that challenging and can often be more related to taste. There's people out there eschewing syntax highlighting and code completion as crutches and they generally don't seem…
I also agree with your assessment that issue trackers and other aspects of the workflow are as important as the coding phase.
I don't know yet exactly what can be done to leverage LLMs to offer real help.
But I think it has the potential of transforming the space. But not necessarily the way it's currently used.
For example, I think that we currently rely too much on the first order prose emitted directly by the LLMs and we mistakenly think that we can consume that directly.
I think LLMs are already quite good at understanding what you ask and are very resistant to typos. They thus work very well in places where traditional search engines suck.
I can navigate through complex codebases in ways that my junior colleagues cannot because I learned the tricks of the trade. I can see a near future where junior engineers can navigate code with the aid of ML based code search engines without having to be affected by the imprecise summarization artefacts of the current LLMs.
Similarly, there are many opportunities of using LLMs to capture human intentions and turn it into commands for underlying deterministic software which will then perform operations without hallucinations.
Building those tools requires time and iteration. It also requires funding and it won't happen as long as most funding is funneled towards just trying to make better models which would potentially leapfrog any ad-hoc hybrid.