Discussion
I use Claude and Thaura for search in my daily life.
I use Claude as a rubber duck for coding. I also let it review my code for errors or make recommendations. On occasion, I'll let CoPilot or Claude add a few lines of code directly. I don't "vibe code".
I enjoy both uses. I don't feel guilty about either.
@me Hey, Jonathan. You are not disagreeing with me.
You should go back and re-read what I said about how I use LLMs. I don't "vibe code", and I don't let LLMs generate a lot of code for me. I use LLMs for reference, rubber-ducking, and occasionally to do mechanical transformations (like changing the order of arguments to a method).
@me Also, I really recommend that you try using Claude, Codex, or another agent to learn about an existing codebase. I think you'll be pleasantly surprised how well they can answer questions like "What are the high-level groups of modules in this software?" or "Why does this codebase still depend on Python 2? What would the upgrade path be?" or "Where is the configuration option for debug logging set?"
@evan I guess my problem here is that I can't trust the LLM not to "hallucinate" the answer. I would feel the need to double check them, negating any benefit I might get because I'd have to do the same amount of work anyway. I can perhaps see some limited utility in using the LLM as a sort of search though, e.g.: "where in this codebase can I find the logic that handles $thing?"
The one point where I really get hung up is this: if machines can in fact think—which I doubt for reasons that are outside the scope of this post—would that not make it morally wrong for us to try to bend them to our will? What makes that different from slavery? I bring this up because it seems to be the argument supporting your position that LLM output doesn't constitute a derived work of its training data.
@me I will say that just for doing research and for doing code review, LLMs save me a lot of time and hassle. I do not enjoy breaking my flow to go read docs on the available arguments for a library's functions. I also catch a lot of typos and other minor errors using LLM code review. I think I make better software faster this way. And I'm a lot happier!
I'd like to setup my own openwebui and ollama server at home sometime, but I haven't gotten around to it. I think it'd be nice to use open weight models instead of cloud services. But I'm doing ok otherwise.
I don't agree with many of the arguments against LLMs I see on the Fediverse.
LLM usage is not a GHG-intensive activity, especially compared to other everyday activities.
It disturbs me when people imply that LLMs are primarily responsible for climate change. This is not remotely true, and it ignores our real responsibility for working on the climate emergency.
It hurts my heart that smart people I know and respect don't know the main sources of GHGs.
I don't use the term "stealing" for ideas, and I don't think it applies for LLM output.
I think that training LLMs on content to produce a model, then using that model to generate new content, is more akin to human learning than to verbatim copying.
I don't think that LLM output is a "derivative work" of the training data.
I don't think the fact that LLMs make mistakes is a reason to not use them. It is a good reason to get reference links, and to cross-check any output.
Install Fedizen
Get the full app experience
@evan This is where I'm going to have to respectfully disagree.
If I'm going to ship code, I need to be able to be able to answer for what that code does. The necessary level of understanding can be achieved with LLM-generated code, but it involves studging its output to be sure that every line—every statement—is understood. To me, it seems that the amount of work necessary to achieve that level of understanding negates any boost in prodictivity I may get from using the LLM in the first place, and it certainly doesn't justify the ethical cost.
I've been working on a fork of a large FOSS project lately. For me, the bottleneck hasn't been the actual writing of the code. More than half of my time has been spent studying the existing code to understand how it works so I can figure out the best way to make my edits. An LLM can't just magically give me that understanding.
Also, when it comes the actual writing of code, it's easier for me to express what I want in the unambiguous syntax of a programming language rather than the crude tool that is natural language. The latter I only use for communicating with other humans, and only because I lack a better tool for doing so. I'll freely admit that that last point is not universally true for everyone.
Also, I'll concede that the use of third-party libraries muddies the waters in terms of understanding how the code works. This is why I choose not to crucify those who do use LLMs to write code. I just choose not to participate.