Clojurians Log v2

I would ne interested in understanding the desire for this. Is it a distrust of the quality of anything produced with ai, or something else? Ai is a tool, and we don't often announce all our tools, so what makes this different? I obviously use AI a lot, but am constantly looking at how to improve usage of AI to get better results. Does the feedback you give on a project vary according to whether AI was used or not?

AI is not just a tool. This seems like a deliberately narrow view to protect yourself from criticism.

That’s an ad hominem. And unconvincing, as anyone participating in this channel engages in criticism, and the person you’re responding to announced a lib on testing AI coding agents

I am genuinely interested - declaring it a tool is perhaps oversimplifying things, but exploring extremes is one way of learning. I realise I have a different viewpoint than many people, and would like to understand concretely how others' views diverge.

feedback does vary. I am much more concerned about people than their code or tools. If a person has taken time to learn and create something, it's worth my time to support them. If they used AI, it isn't.

Ia he concern that that there is likely nothing to learn from the project? or that someone announces a "drive-by" project, which they have no real interest in maintaining or improving? Does that presuppose that there has been no effort in the design of the project or the steering of the AI?

it means I don't care - personally - how much effort you put into guiding the AI.

it's not interesting to me. Others - who accept the use of AI - will differ

Fair enough - may I ask why?

http://politechs.dev

It will give you a more complete understanding of my position than I can offer here

Thanks - I had not seen that before. I suspect I will actually agree with a lot of the opinions there, and I suspect that my viewpoint will differ in that I believe it can still be used for good. But I'll have a listen.

Thanks. I used to think that was possible too.

If not a tool, then I think it can otherwise be a peer or god. A friend manages at a company where programmers were ordered to vibe-code, and it seems the programmers were reduced to brainstems that provided priors/biases for AI I think that's considered a bit odd in the Clojure community, which has a culture that views technologies as tools; and programs as data that can be generated by a machine

What constitutes AI usage? Is it just when the AI writes code directly to your project, or is it any adjacent usage? For instance, if I use it to search through documentation to find a particular class name, does that count as AI usage?

You can judge for yourself

Are things like AGENTS.md sufficient signs of AI use, rather than expecting announcements to reliably document it?

Does the feedback you give on a project vary according to whether AI was used or not?yes, very much so (obviously depending on how much was used). feedback is a human discourse and I feel like if you didn't bother writing it, then I'm not going to spend time reading it in enough detail to give actually good feedback. so even just saying something like "idea and design is mine, but the detailed execution is LLM" would give the other party a chance to better orient themself and everyone benefits

I am a first time open source developer (I have been a developer of sorts for over 30 year), just never published anything open source. I am using Claude, and it is mentioned clearly in the README.md of the project and in all commits. In this particular case I learned about some technology that seemed a good approach for a problem I was facing a few years ago. I read about it, and build a few components myself. I ended up with a good understanding and a total lack of time to build more (as the library I imagined required a lot of components to be useful). When Claude arrived I realised it might be a helpful tool to scale up what I had thought about. And through experimenting a bit I think I managed to do just that. It isn't super smart, it isn't horrible. It is just a tool that can be used to accomplish tasks. Some better than others, and it does require a lot of guidelines, babysitting, reviewing and improving. Regardless of that, it is clear to me I would never have had the (free) time to build the library I ended up creating. So for me it is useful. Whether or not anyone else should want to find it useful, or want to use a library like the one I created is not up to me. I have no issues marking releases with a Claude comment. And if you prefer to ignore that type of release, that is fine with me. I am not here to convince anyone, just experimenting with wha tI can do in the free time I have.

Like Ray said, this thread has gone far off track. It is not about whether AI is useful, it is not about how much AI is ok. It is about encouraging transparency, and creating community norms that allow for that transparency. But it seems no one here is interested in discussing that, so I suggest we leave it here.

Heres how I use AI to get its maximal benefit 1. Write code my self with industry best practice in mind + write tests for the code. 2. Only use AI to create a similar code in which AI can use my old code as a reference. Give AI specific test scenarios to test and ask them to write tests (AI references test that I wrote and writes new test in similar pattern) 3. Review code written by AI. The goal is to keep the generated code as small as possible and review them 4. Commit changes. And being explicit in the commit who wrote teh code, was it AI or me etc 5. Repeat step 1(2) - 5 AI is a pattern matching calculator. It is not intelligent. The quality of the output depends on the quality of the prompt + input data. Its mostly useful for generating code that are highly repetitive (web apps - components, new endpoints etc)

@plexus I'm curious, how would you use such disclosure?

@delaguardo It's fundamentally a different way of working because it removes fundamental constraints (time, effort, code length etc) that lead to different outcomes. It has a whole different class of failure modes, the mistakes LLMs make are very different to the mistakes humans make. Disclosure helps because it requires a completely different mindset to audit.

@andersmurphy I mean this literally — let's say something is labeled "AI used", what then? Does your thinking change while reading this? Or do you just pass it by without reading at all? I thought about this too a while ago and couldn't figure out what I should do differently after seeing such a disclosure.

It seems to me without clarifying the problem space (what we're solving for and why), we're unlikely to come up with a good solution.

@delaguardo I'd definitely say I'm more cautious. I check project size, number of commits etc. If it's terse an well written I'll read more. If it's massive I'll probably pass. I guess I make a judgement call on how feasible I think it is for the author to have actually read and understood the code that has been generated. All of this is much higher bandwidth.

It's like a more extreme version of how I treat non LLM projects, because LLM projects have a much higher chance of wasting my time (my experience so far).

Thanks! Isn't that more demanding to the review process? From my personal experience I started to spent much more time reading through a new project to understand it in details than a few years ago.