As a agentic n00b, I’ve used both Sonnette and Opus in the past couple of weeks. I’ve definitely felt validated that I was never in any danger of “falling behind” as I can get it to write pretty good code so long as I ask it to (ie, give it direction on implementation details). I still haven’t done any “serious” work with it, though, and not allowing myself to use it all time (as I touched on above).
sodapopcan
paulsullivanjr
I continuously iterate and refine how I prompt. AI can create some slop but then again I’ve seen plenty of slop before AI became a thing in all kinds of languages, including a couple Elixir projects.
2
tfwright
Neither what is good nor bad about AI is the quality of what it produces, what defines slop is the speed at which it is produced. Plenty of teachers will tell you that mediocre AI generated papers on any topic are much, much better than what many (maybe even most!) of their students have produced in the past. But AI brings the cost of mediocrity down so low that it massively disincentivizes aiming for quality (again, definitely agree has always been an uphill battle for those who care about it enough to fight). Students who might have otherwise put the moderate extra effort to actually think about the topic (and therefore learn something, above all how to think) now have a much different calculation they are running. Same in software. That is the threat.
5
Tyson
Well put. That is much scarier than many of the operational questions we’ve been juggling.
0
mudasobwa
Creator of Cure
Well, my experience differs. Also, saying “something that is not a developer” is like saying “someone who is not taller than 6 feet.” “Developer” is not a blanket term/hypernym. The code of different developers might differ hundreds of magnitude times more that the code I produce from what my LLM produces.
Yeah, as SQL did, as COBOL did, as Java did.
0
mudasobwa
Creator of Cure
Reading through the thread again, I noticed that the single obvious hack has never been mentioned.
Before using LLM to produce code in language XYZ, I always feed it with best exemplary pieces of code in that language. Once I met the antipattern in the outcome, I immediately put it into the list of dont-do-that.md. That works for me (besides other techniques, as using home-baked MCP over RAG system working with AST tree instead of plain text.)
That said, having files GOOD_PRACTICES.md and ANTI_PATTERNS.md, listed in your AGENTS.md as such, is a must.
2
martosaur
It is though. All developers have human brains, were kids once, can read and write, and wake up every day with memories of the past. They have quite a lot in common! LLMs are very different, so I think it is very reasonable to expect LLM-generated code to have some unique properties to it.
Do you work primarily alone or in a team setting?
0
mudasobwa
Creator of Cure
Both. Besides that I have a ton of commits to OSS, to Elixir core, to “standard” Elixir libs, to 30+ not-so-standard libs in C, Ruby, Java, Rust.
0
mudasobwa
Creator of Cure
This logic leads to “grader removes snow in a completely different way than a dude with a shovel” statement btw.
1
tfwright
I don’t think so, for the same reason the analogy earlier in the thread is oversimplified. We can plausibly draw a distinction between tools that amplify human intention/capabilities in a form of labor and what AI appears to do, or at least claim to be able to do, which is to replace it entirely. In the former case a person is still using their judgment to make decisions and resolve ambiguities. In the latter, those are resolved in advance by the LLM model. In the former, a person must be present, must oversee, in the latter, the human involved can just step away and come back later. In the former the person is still more like a worker, in the latter they are more like a manager of a new “type” of worker. Framed in that way, I don’t think the merit of the distinction is all that debatable honestly.
I think the real question is whether the latter use of AI is actually desirable, and if not, what steps can be taken to avoid it or at the very least mitigate the harm. What a lot of people in this thread say is that they personally don’t use AI to generate slop. We all care about quality and value our own craftsmanship too much for that. We use it more like the power tool and I’m guessing you are in that group as well. But it doesn’t seem to be what’s happening in most of the industry, including on teams we are ourselves on. Even on this forum, we are seeing a sharp uptick in AI generated code of the latter kind, so it’s a valid question with real implications for this community as well, regardless of how we personally use AI.
0