3 ms·
"Note that querying LLMs and making the outputted code conform to specifications is probably a lot easier than writing the code yourself" This overestimates th
by imtringued 4y ago
"Note that querying LLMs and making the outputted code conform to specifications is probably a lot easier than writing the code yourself"
This overestimates the time developers spend typing text and underestimates the time spent thinking about problems, verification and locating bugs/code, etc.
Just because the output of a model is text doesn't mean it is a replacement. You would need extraordinary "single shot" functionality where the LLM always outputs correct code and knows how to deal with any potential specification. The moment it fails then you wouldn't want "paralegal" developers, you would want normal, maybe inexperienced, developers.
Single shot functionality is very impressive because humans can't do it, they build things incrementally, but once your program gets large, it can only be written incrementally, which at least implies that the AI needs to keep track of unbounded context.
- nonethewiser 4y agoThe smaller the problem the easier it is to use AI. Which makes the window where it's useful very small. Its easier to understand and test code that you wrote than code you just read. If you ask it to write a 100k line program you're going to have to trace through everything and read it countless times over to fix any issues. If you ask it to write a small 10 line snippet that you need to test you might as well do it yourself.
- flir 4y agoIt can write a ten line function faster than I can, and my code has a lot of ten line functions in it. Even for trivial code, it's a force multiplier (moreso when plumbed into an IDE)
- uwuemu 4y ago[dead]
- aws_ls 4y agoOnly yesterday, I worked on a ticket for few hours, inspecting several Mongo collections in QE and prod. And then had to make 4 to 5 lines of changes across 3/4 python scripts. And then manually run them to patch the mongo collections. Which was the root cause of the bug. I don't see GPT doing this kind of stuff anytime soon. And yes, very less percentage of the time, do we right chunks of fresh code, in which this could help. But that again to an extent. It can generate functions. But the glue code is still complicated. So it can definitely be a better Stack Overflow replacement, perhaps.
- hnbad 4y agoMuch like the adage that you need to be smarter to debug code than to write code, you need to be smarter to "touch up" AI generated code than to write it yourself. The problem with AI generated code is that it contains mistakes and because of the nature of how it works, those mistakes may be entirely unlike the kind of mistakes a human would make or may be completely undetectable unless you know what the code should be doing. The best use case for AI code generation is auto-complete. Depending on the complexity of the code, that can be sufficient to do 90% of the work of writing code, but it's very different from "solving problems" which tends to be the tricky part. It also requires the AI user to already know the code they want the AI to generate before it does it. An interesting application of AI outside of simple code generation however is what GitHub is currently experimenting with as "code brushes", i.e. "touching up" existing code, trying to fix bugs or apply refactorings. This fits in the gap of "single shot functionality" but operates on existing code rather than requiring a blank slate. But at the end of the day I think AI will fail to kill programming just like other attempts like WYSIWYG UI builders and COBOL have and for the same reason: competence as a programmer is not about the ability to write code but the ability to understand programs and systems, both at scale and in depth. A beginner will not write bad code simply because they lack experience in using the language, they will write bad code because they lack understanding of what the code does and how it interacts with its environment (and how that environment interacts with it in turn, e.g. UX).