←back to thread

AI’s coding evolution hinges on collaboration and trust

(spectrum.ieee.org)
183 points WolfOliver | 2 comments | 29 Aug 25 15:24 UTC | HN request time: 0.546s | source
Show context
jbstack ◴[29 Aug 25 17:38 UTC] No.45067107[source]▶
But what is meant by "real coder"? Does it mean an experienced senior coder? In that case, no, AI isn't ready to replace one. How about a junior coder who is fresh out of university and starting his first job? Surely that counts as a "real coder"? I'd argue that AI can and does already replace such coders, depending on the type of work they have been assigned.
replies(1): >>45067167 #
gchamonlive ◴[29 Aug 25 17:42 UTC] No.45067167[source]▶
Could you elaborate your argument? I'd say AI is far from replacing any human contributor at this point.

Context doesn't work the same way as memory + experience in humans. While humans have an impression and a flexible mental model of any single domain, AI needs hard data, which is hard to manage with context and can't really be worked around by fine-tuning in practice, lest you have to retrain the model on each and every code merge.

replies(1): >>45067557 #
1. jbstack ◴[29 Aug 25 18:14 UTC] No.45067557[source]▶
Sure. My point is that it's not unusual for the quality of work of a junior coder to be fairly low, and that work typically needs to be double checked by someone more senior before it can be used in production. Often, an LLM will produce a higher quality output than that junior developer in which case, given that the work has to be checked by a senior coder no matter what, the LLM can replace the junior developer.

Also, not every project is highly complex or involves huge codebases and, for those that are, a junior developer might be assigned to a reasonably self contained and small module within the project (i.e. something that a LLM can do better at).

Finally, although this wasn't part of my original argument: before modern LLMs, let's say that a senior developer could produce N units of work per Y units of time. Now, with the help of a LLM, they are able to produce XN units of work per Y unit of time. I don't think it's too controversial to claim that some portion of developers are now more productive than they were before. And if developers in general are more productive, then you can expect some reduction in the number of developers that you need for a given project. So, in an indirect sense, LLMs are already capable of replacing coders.

replies(1): >>45068458 #
2. seadan83 ◴[29 Aug 25 19:38 UTC] No.45068458[source]▶
The pattern I see time & time again in software development: a small set of few and usually very talented senior developers build a software project and they build examples of how everything is done. Junior developers then extend those examples, eg: "add another column to this table." Juniors are just extending the existing patterns, seniors are the ones that create them.

This pattern I see play out time and time again. Senior builds the integration to various team services, then juniors tweak & mimic those to do more.

> let's say that a senior developer could produce N units of work per Y units of time.

So, it's more the case that senior developers are building full production assembly lines, that can then be replicated to do similar things relatively easily. The cost of the units of work are not at all equal.

How does this apply to AI? Good luck getting AI to do something novel, by definition it can't. OTOH, extending an existing pattern is trivial for a senior developer to do after they have built that pattern; juniors spend a lot of time at it cause they have to learn everything.

So, who is the AI going to replace? If the junior, then you have a senior developer using an AI to do something that is already trivial for them to do. No way AI is replacing the senior.. So who? Who is getting replaced by the AI and who is using it?