Maybe we were mostly actually doing the exact same stuff and these things are trained on a whole lot of the same. Like with react components, these things are amazing at pumping out the most basic varieties of them.
Maybe we were mostly actually doing the exact same stuff and these things are trained on a whole lot of the same. Like with react components, these things are amazing at pumping out the most basic varieties of them.
Meanwhile people who work on things that are even slightly niche or complex (ie. Things which aren't answered in 20 different stack overflow threads, or things which require of few field "experts" to meet around a table and think hard for a few hours) don't understand the enthusiasm at all
I'm working on fairly simple things but in a context/industry that isn't that common and I have to spoon feed LLMs to get anywhere, most of the time it doesn't even work. The next few years will be interesting, it might play out as it did for plane pilots over relying on autopilots to the point of losing key skills
I'm currently increasingly often confronted with overly bloated codebases that, with the added speed of llms / coding assistants, have reached levels of unmaintainability much faster than traditionally possible.
Now, fixing those repos with LLMs fails miserably in my experience, so you have to invest hundreds of hours of senior engineer time to make the projects workable again, regardless of the novelty of the problem solved by that codebase.
It also frequently makes the engineers pulled in to help miserable.
But then there are those idiosyncrasies that every company has, a finicky build system, strange hacks to get certain things to work, extensive webs of tribal knowledge, that will be the demise for any LLM SWE application besides being a tool for skilled professionals.
There’s not much to ponder on here as the majority of human thought and action is unoriginal and irrelevant, so software development is not some special exception to this. This doesn’t mean it lacks meaning or value for those doing it, since that can be self-generated, extrinsically motivated, or both. But, by definition, most things done are within the boundaries of a normal distribution and not exceptional, transcendent, or sublime. You can strive for those latter categories, but it’s a path strewn with much failure, misery, and quite questionable rewards, even where it is achieved. Most prefer more achievable comforts or security to it, or, at most, dull-minded imitations thereof, such as amassing great wealth.
If you think about it, lots of professions are already divided in two. You often have profession of inventors / designers and a profession of creators. For example: Electrical engineers vs Electricians. Architects / civil engineers and construction workers. Chefs and line cooks. Composers and session musicians. And so on.
The two professions always need somewhat different skill sets. For example, electricians need business sense and to be personable - while electrical engineers need to understand how to design circuits from scratch. Session musicians need a fantastic memory for music, and to be able to perform in lots of different musical styles. Composers need a deeper understanding of how to write compelling music - but usually only in their preferred style.
I think programming should be divided in the same way. It sort of happens already: There's people who invent novel database storage engines, work on programming language design, write schedulers for operating systems and so on. And then there's people who do consulting and write almost all the software that businesses actually need to function. Eg, people who talk to actual clients and figure out their needs.
The first group invents React. The second group uses react to build online stores.
We sort of have these two professions already. We just don't have language to separate them.
The line is a bit fuzzy, but you can tell they're two professions because each needs a slightly different skill set. If you're writing a novel full-text search engine, you need to be really good at data structures & algorithms, but you don't really need to be good at working with clients. And if you work in frontend react all day, you don't really need to understand how a B-Tree works or understand the difference between LR and recursive descent parsing. (And it doesn't make much sense to ask these sort of programmers "leetcode problems" in job interviews).
ChatGPT is good at application programming. Or really, anything that its seen lots of times before. But the more novel a problem - the more its actual "software engineering" - the more it struggles to make working software. "Write a react component that does X" - easy. "Implement a B-tree with this weird quirk" - hard. It'll struggle through but the result will be buggy. "Implement this new algorithm in a paper I just read" - its more or less completely useless at that.