Clojurians Log v2

> A list of open-source projects that reject AI-generated code The link you put has this as top-line rationale. So “reject AI-generated code” is their own top level billing, as quoted fairly by Ben here

I have regular conversations with brib on Saturdays. They are really cool. 😊

Also Codeberg gets attacked often for their firm pro human stance. Might have something to do with their periodic downtime

Github, famously pro-machine, similarly attacked for their periodic downtime.

I really like the idea of the slopfree index. I do kinda wish that they had named it the "thoughtful codebases" index, but I get the inclination to push back hard in the name. The thing is - there are precious few projects on that list. As a (minor) contributor and newbie to the Clojure space, there isn't much to chose from. My plan: • I'm thinking of switching my editor back to gnu emacs from VSCode. VSCode seems to have becomes an LLM-first editor, and it somehow crashes more often now than a few years ago. • It seems like many Clojure projects are open to generated code and generated issues. I'll continue to contribute as can but I sometimes have a feeling that I'm rubbing shoulders with clankers, not humans. Open source isn't about me (as they say) and I am far from experienced enough to criticize, but it can feel blah. Not sure there's anything actionable here. • For myself, I've been following* something similar to the https://openjdk.org/legal/ai . I'm lucky enough that I don't have a boss dictating my policy. Do you or others have suggestions, critiques, or thoughts? (* it helps that I don't have a big token budget 🙂 ) ( Aside: I wonder how long the Java team will be allowed to have this policy. Oracle has been relatively hands-off, but considering their financial difficulty I can see pressure being applied.)

I know brib would love contributions!

They tell me they are a bit behind because they are in university

I think this is an x y problem. https://en.wikipedia.org/wiki/XY_problem There are certainly problems with how some people have used some AI at some points. It's worth digging into those problems but those problems are not inherent to AI nor are they unavoidable. One problem is that AI output is sometimes really hard to ingest as a human. It's overly verbose. It doesn't structure or order the information in a logical progression that meets the reader where they are. It speaks in sentence fragments without subject or verb. I would be offended if someone gave me text like this, and I would be just as offended if a human wrote it. Another problem is that the human meat proxy may not be able to intelligently answer follow up questions about the output. That certainly can be a problem, but the human may also be well enough acquainted with the subject that they can answer follow ups intelligently. Does it matter that the output happened to be created by an AI in that case? I don't think so. I would be just as offended if a human wrote something themselves and overlooked some obvious problems or follow up concerns. This isn't a new problem or a problem unique to AI.

So I think the idea that AI output itself is the problem is missing the point. That would be bad enough of a misconception, but I think it's poisoning the well to escalate that misconception into the moral conviction that AI output is slop by definition, and that it's disgusting to expose to colleagues. I expect more introspection, more humility, and more argumentative sophistication than that. Make a technical or practical argument instead.

So I think the idea that AI output itself is the problem is missing the point.

> My own take on AI etiquette is that AI output can only be relayed if it's either adopted as your own or there is explicit consent from the receiving party. > This is not an invitation to engage. If there are specific problems with some the content of communication that's worth engaging with.

To the extent that the article is worth engaging with, here's what I think about it: It seems to be roughly equating AI output as meaningless noise. If that were true I think it would be enough to say that sending someone meaningless noise is rude, which it is, regardless of who generated the noise or how. AI output is not just meaningless noise, at least not all the time, so a better argument deserves to be made. One could criticize AI output for being hard to digest, despite containing some useful information within the noise. When that happens it's a legitimate criticism, again regardless of whether it's AI or human generated. We should take care to produce good writing that is easy and useful to digest by its audience. Anything less is a waste of the audience's time and effort. One could say that even if some output is useful and well written, if it's AI generated then the meat proxy who pasted it is a useless node in the communication graph. If that's true then the audience should just go ask the AI themselves. I do think there is some social utility in requiring human-only communication in some circumstances. Requiring pure human communication not just accomplishes the immediate communication goal but it also reveals the mental models of the humans participating. Even with AI involved, the mental models of the humans matter hugely, and I definitely think some mechanism is needed to reveal, align, and evolve human mental models. I just don't think that's required 100% of the time, and I don't think it's fair to insist on it 100% of the time. Growing human mental models is not always the point or worth the effort.

I'm trying to find where you disagree with the article. > AI output is not just meaningless noise, at least not all the time, so a better argument deserves to be made. This is not how I would summarize the article. I think it's actually much closer to the sentiment you provided: > We should take care to produce good writing that is easy and useful to digest by its audience. Anything less is a waste of the audience's time and effort. In fact, your statement is stronger than what the article suggests. The etiquette suggested by the article also allows for folks to politely share their ideas, even if they are poorly expressed, as long as the ideas are their own and are shared in good faith. > I just don't think that's required 100% of the time, and I don't think it's fair to insist on it 100% of the time. I assume this is referring to requiring purely human communication 100% of the time, but that is not what the article is suggesting. The suggestion is human only content with two exceptions: • the content (ai or otherwise) is adopted as your own (deciding what that means is left as an exercise to the reader) • there is explicit consent to share ai content So yea. I'm struggling to see where you disagree with the article. Do you have a different etiquette in mind?

Where I disagree is I don't think it's the audience's right to require anything of the speaker beyond the content of the speech or the conduct of the speaker in the context of the conversation. I don't believe in requiring an AI disclosure. I don't believe it's rude to communicate AI output as long as the communication meets all other standards of good communication.

The article doesn't actually say anything about requiring AI disclosure. However, it is my own view that requiring AI disclosure is good for online good communities. Authorship The main reason for requiring AI disclosure is that it tells the reader who the author is. It is extremely rare for online communities to be completely anonymous, where messages have no indication of authorship (note: this is different from pseudonymous, where messages have pseudonyms that may or may not be linked to real identities). I think truly anonymous communities are rare because knowing the author of a message provides a lot of value to readers and to communities. In reality, the content of a message is almost always viewed in context of its metadata, including its author. One could argue that LLMs are a tool and don't change authorship, but I don't find that argument compelling. Ignoring ai, most generative tools actually include the name of the tool as part of their output because authorship is important metadata for content. Bots Another reason for requiring ai disclosure as a general policy is that it builds trust. Let's imagine there are two forums, one that is 100% human content and one that is 100% ai generated. I would personally not be interested in the 100% ai generated forum, even if it is indistinguishable from the 100% human forum. I would not be interested in the ai forum, even if it was better. I don't think I'm alone here. Requiring ai disclosure is a useful policy if you are the kind of person that would not be interested in participating in an online community of bots. Mistakes From a purely technical perspective, I'm not sure why ai disclosure is controversial, especially for code. The practical reason is that LLMs make categorically different mistakes than humans. Yes, humans make mistakes, but the types of mistakes humans make are different. Sure, there is some overlap, but my experience is that knowing how content is generated is extremely useful for me to know what types of mistakes to expect. Negative Externalities The last reason that I would consider ai disclosure to be good etiquette is that ai generation has negative externalities. For some people, the negative externalities associated with generative ai crosses the threshold and they don't want to be complicit. Even if you disagree about where the threshold should be, I think it is rude to hide ai usage if you know that it is important to other people to know if they are engaging with ai generated content.

I don't believe it's rude to communicate AI output as long as the communication meets all other standards of good communication.

Anecdotally there are fewer “slop grenades” in my work place after some incidents where total nonsense was presented as fact. In a couple of cases a skim read and gut reaction was enough to know. I'm still finding tickets/specs from PMs are full of inaccuracies. Having a conversation and re-writing based on what they/the customer wants has helped with that. There's been no discussion about rudeness but it does make people look silly. When something has my name on it I make sure it is accurate. I'm in an environment with a lot of pressure to use LLMs. I’ve found tight specification and QA is essential to get generated code that isn’t full of bugs. More so than with human code.

This name already caught traction, it's not Ben's idea, just do a web search.

That doesn't mean you have to keep repeating it

I'm making an allusion to an argument that if the human is merely taking queries from other humans, putting them into an AI agent, then copy pasting the answer back, they are nothing more than a proxy and are contributing no human intelligence to the process. It's not meant to be flippant. My usage of the term is shorthand to refer to that argument. If you have a better term that lets me refer to that argument without so many words I'm happy to use it. In the context you're quoting, merely saying "human" wouldn't get my point across.

@christian767 I just pointed it out without any personal take. But since you reacted, sure, we don't have to repeat it, but where do you draw the line? It's a known term. Does the term vibe-coding bother someone? Clanker? Words and phrases have meaning and we're going to see "meat proxy" on the internet.

@ben.grabow I think the request was to use a phrase, e.g. "human proxy"

Not the same meaning.

This whole argument is backwards, "don't show AI output to people". What matters is the text. If I, through reading tons of AI text, take some mannerisms over and start writing like an AI, is my text bad? What matters is what information the text covers, and how well it conveys that information. Whether it's AI generated or not does not matter, at all. If you send me incomprehensible AI slop I will tell you "don't send me AI bs". If you send me legible, readable, factual text, I will thank you the same way whether you used AI or not.

If I, through reading tons of AI text, take some mannerisms over and start writing like an AI, is my text bad?

Please explain

interesting, I said "take some mannerisms over", so, only some, and I did not pin which, and my text is already bad? Thank you but I don't want to join your cult.

I see you edited and added bold to "start writing like an AI", but that's still not pinned. Writing like, not exactly as an AI.

Another way to approach this debate is a "failure to cite" or "failure to show your work". If I hired someone to ghostwrite a book for me, and I reviewed every page and believed it to be exactly representative of my own thoughts, I would still be in error (both morally and epistemically) to claim that I wrote the book.

There are some cases where the point of writing is to demonstrate your thoughts. Other times where the point is to "fix the bug" or "implement the feature". I don't see why we should care that much about attribution or proof of thought in those latter cases, as long as the behavior of the code is verifiable.

Can't we treat the human+AI as a single system?

Narrowing down on authenticity: LLMs are a plausibility machine. If you write a linkedin post that's just LLM output, you're misattributing the work as your own, as something that you thought about. The same is true when you use it for code.

@john.t.richardson.dev only if not disclosed though, right?

Of course, just like a book's ghostwriter (you mentioned elsewhere).

a little late to the party, but i think “meat proxy” is a great pejorative of a failure state to be avoided. Also agree that sometimes it’s important that a human solved it, sometimes important that it was solved, deliberately passive voice.

I'm laughing to myself thinking of what linkedin would look like if posts were automatically given pangram ratings. 🙂

The people posting LLM slop on LinkedIn would not be caught dead disclosing that it wasn't their own thinking. If they were magically forced to do so, they would probably stop posting altogether. The aversion to disclose LLM use is a stronger signal than the disclosure itself.

> The aversion to disclose LLM use is a stronger signal than the disclosure itself. aka shame

@U0BHKM127SQ Yeah. Writing is thinking. It's hard. It's empathetic. When you write, your "thinking muscles" get stronger.

For what it's worth, I wouldn't read a novel generated using AI. But I think I will read a technical spec generated using AI given that it transmits the details across. I think there is a difference.

> technical spec I haven't thought too hard about that, but it might have it's use. I like to use the 4 D's of robotics as a stand in for what LLMs might be good for: • dirty • dull • dangerous • dear

For open source, tech specs might be dull. (I know tech writers that would disagree about that in general though.)

@U0BHKM127SQ Yup. There are great technical books (going back to the old K&R C book) that are a good read. I'm not convinced that an LLM could give us that.

@U0BHKM127SQ I think all the power to them. I personally don't adhere to any essentialism though.

I have a surprisingly good feeling about this, though. It seems likely to me that this will have a "hormetic effect" in the long run, resulting in society placing much more value in real insight, quality, and truth than we did even before this recent wave of slop. I think of the way advertising has changed. The old ways of deceitful in-your-face advertising (Magic Himalayan energy-bracelet! Last chance! Buy now! Get a six-pack in 3 weeks!) doesn't really work as well anymore. We have learned; we demand proof and honesty, and we are much more skilled at detecting deceit. I think that no one is more tired of AI slop than the young TikTok brainrot generation who have suffered from it their whole adolescent/adult lives.

maybe not. But LLMs can give you quite tailored tools to help you learn things. Here’s something i had it whip up recently for me: https://dpsutton.github.io/execution-engine-learning/ blog post style with 7 lessons building up to the bytecode format that sqlite uses. interactive visualizations, code done in both clojure and go, examples that you write the code yourself and it runs tests to help you out, golang code is compiled to wasm and run in ghostty-wasm in the website so you can run it there without a local env. and then connects to modern database source code so you can see “real” examples of it in the wild. I balance this amazing benefit against all of the downsides. It’s true there are downsides, but it’s a very useful tool as well in my experience

> Another way to approach this debate is a "failure to cite" or "failure to show your work" LLMs don't change our values like honesty, professionalism, devotion. • If someone generates a book for me, doesn't read it, the book is horrible, and the author did not warn me it's unread AI slop, I'll feel offended. • If someone writes a book for me about quantum physics, but fails to mention (s)he has Mr. Bean level knowledge of the topic, I'll feel offended. See, the world has changed, but our values did not.

requiring attribution (Assisted-By: <your-llm-here>) is a per-project choice. Because it's really not black and white. For example, even the slop-free index projects have various takes on this. I've read https://ziglang.org/code-of-conduct/#strict-no-llm-no-ai-policy, which prohibits any use, even for spell checking, or brainstorming without using any prose, or findings bugs. So they take offense if I use an LLM. For example the LLM could help me find a critical bug, and I would write the prose, the tests, the fix, the PR, everything. It's still prohibited by the CoC.

Policing asynchronous LLM use is obviously an exercise in futility. With a zero-tolerance attitude like that, you are better off focusing on your company culture rather than the CoC. Although this might be an open-source project so maybe that's why they are in such a weird situation.

> https://dpsutton.github.io/execution-engine-learning/ As someone who used to work in this space, I'm getting flashbacks to SQL Explorer docs. The content certainly seems plausible, and the interactivity is a neat addition. As part of this thread, I'd add: attribution?

@xifi > For example the LLM could help me find a critical bug, and I would write the prose, the tests, the fix, the PR, everything. It's still prohibited by the CoC. zig may have other reasons to forbid LLMs (fossil fuel use, copyright concerns, ownership, etc). I'd understand that. IMHO: I'd lean to your use case being OK.

yes, I understand they might be taking a stance for other, legitimate reasons. I am not judging, you do you. What I don't like is statements like "showing AI output is rude" - this is judgmental, and way too broad a stroke. It depends.

It always "depends", I wouldn't be offended if you used an LLM for spellchecking. It doesn't mean we can't make any statement

@xifi Do you see a spectrum? On one end, I find LLM LinkedIn posts quite rude. I find 10kloc LLM PRs quite rude. In the middle, unattributed blog posts generated by an LLM are mildly rude. On the other end, someone may have used an LLM to prototype/research an idea, stashed the code, thought it through, wrote tests and the code themselves, and had an LLM review it. I don't find that rude.

> Why should I read this if you didn't think it? > I can think of many reasons why I might want to read something irrespective of the author's mental state. I might learn something if the content is true, for one. Even if the content is not true it might provoke me in a useful direction, a la brainstorming.

Rich's reply seems to conflate "a human didn't think it" with "the product is going to taste bad" which I think is an obvious logical mistake.

That's the kind of LLM output you would read when prompting yourself. You wouldn't want to read my LLM output forwarded to you as my message to you.

@soltanzadehramin Plus, then your own LLM output would be tailored to exactly what you wanted to know

That's a load-bearing insight and worth being precise about.

My AI output is not your AI output since I'm putting many of my own design opinions into the prompts. An AI proximally generating the text doesn't mean my voice is absent.

This also turns on whether we're talking about LinkedIn posts or bug fix pull requests. I have no desire for a unique voice in the latter.

very much agree with that. It’s not just one-shot whatever it generates. There’s a shaping process, tools, guard rails, invariants that all mean there’s orders of magnitude difference in the output.

> I find 10kloc LLM PRs quite rude @john.t.richardson.dev I find the PRs themselves rude when AI is involved - don't care the # of lines. I think this will go away hopefully.

Honestly i think that attitude will become economically non-viable, and in a security context, impractical or even irresponsible soon.

Not saying this normatively, or it’s a desire, but just where i think the industry is heading. Quite possibly wrong. But that’s my guess. New tools are changing things

> Do you see a spectrum? Sure. I judge based on the content. And when I generate something, I am open about it, as well as how much effort I've put in, how much of it I understand, how much is tested etc.

It's gotta do with bottlenecks. PRs are a bottleneck.

> You wouldn't want to read my LLM output forwarded to you as my message to you. If you are someone I trust, and you decided it's worth it to trust that particular LLM output with me, I'd definitely read it.

Insofar as the idea being communicated is trivial, the LLM output and the handwritten output are going to be similar (e.g. adding a missing semicolon). Conversely, insofar as the idea is non-trivial, your voice is drowned under that of the LLM. I think your premise is that the work consists of a lot of trivial things, and I think Rich and others don't share that view.

@xifi Agreed, but in order to not betray your trust, I would have to put so much effort into analysing and verifying the output that I would effectively be communicating manually, just with extra steps.

if that's your reference frame, I can accept that. It's not everyone's, though. And your initial statement is then falsified. It's not that I wouldn't read your LLM output, it's that you'd rather write it yourself, which conflates facts with opinions.

No, it's not falsified. The premise of forwarding LLM output is to save time, and I am arguing that you don't want me to save time in my communication to you. That would be a case of me offloading work to you.

@dpsutton I don't mean to be rude, it's just my experience. The developers are assessed on lines of code as a metric. So developers push more lines of code using AI. The PR reviewers rubber stamp the PRs while relying on bots to catch errors in the code. This is the state in my company now and I don't see this reversing. This has been going on for months now.

yikes that’s a bummer. Not my experience, nor what i see people talking about online. I’m sorry you are experiencing that.

but that would be disheartening and i’m sorry that’s your workplace

I'd be interested in a survey trying to figure out how common that is. That was also my experience at my previous workplace. I have a strong feeling it might even be a majority of companies.

I was gently told off by manager for not producing enough deliverables compared to the AI credits I was using just last Friday. So, more code it is against my better judgement. I think it is difficult quantify the performance increase due to AI otherwise in a large company.

> You wouldn't want to read my LLM output forwarded to you as my message to you. @soltanzadehramin we concluded I would, therefore it is falsified. Whether it saves time for you or depends solely on you and your choices. I will not dictate how you prepare your message, and I don't want you dictating me what to read or not. If you want to dictate me how I prepare my text, that is where we will disagree. Because who are you, or Stu, or anyone here in Slack to tell me how I write. Will we ban Emacs next? Spell checkers? Will be require PGP signing of everything? Again, if it's your project, your community, your forum, it's up to you to come up with a CoC and enforce it, and I will choose to join or not. But when people want to force others how to do their work, I mean, why? Just let me work.

that sounds like an XY problem. Manager demands deliverables so you respond with more code against your better judgment

> I would have to put so much effort into analysing and verifying the output that I would effectively be communicating manually, just with extra steps. this is your experience. I accept it, and am not forcing you to change your ways. Why are others so trigger-happy to change the way of others?

@dpsutton In this job market, I'd rather keep my job and do the thing I know to be wrong.

it’s the general thing. they are asking for deliverables, not lines of code.

Sorry, my bad. I meant he had the contribution dashboard open and he showed me my "contributions" in the form of PRs merged to main.

@xifi I'm pretty sure no one here has demanded anything of you, nor suggested that anything be banned. I for one am happy to see a lot of irresponsible companies show their true faces and produce before unseen amounts of slop; I think it gives the good guys a competitive advantage. The last thing I would want to do is to force them to stop.

what is "showing AI output is rude" if not social pressure to not use it?

A wake-up call? An invitation for debate? An exaggerated statement to balance the excessive amount of slop out there? Talking with your mouth full of food can be considered rude, and publicly stated as a strong opinion by someone. That doesn't mean they seek to control how you eat or talk in their absence.

The tone conveyed signifies (to me) that people have already made up their mind about the moral valence and want to impose their norms on the community. I don't feel invited to debate.

My gratitude goes out to all those who do want to engage and debate.

If it is anything of what you said, great. If I did not get that feeling, it means that part of the message wasn't conveyed (to some at least).

I hate AI slop as much as the next guy.

I think the tone is the same as when we say "mutable state is bad". I think most of us say it with a made up mind seeking to impose it as a norm in the wider community. I can see that some OOP devs might see it the way you just described, as arrogant and not inviting. As far as evidence goes, I think Stu has been active and engaging in the debate here. No?

For what it's worth, I think a lot of the strong opponents of AI-generated code are in a state of disappointment and despair. It's not easy to participate in fruitful debate from such a place, so I think that results in more "public bursts of opinion" rather than invitation for discussion.

Stronger developers will go through the five stages of grief later than the average developers. And we are in a Clojure sub which has historically prided itself on Simplicity, brevity, mastery, etc. My 2 cents.

People are active and engaging, yes. That doesn't mean there's a high signal-to-noise ratio in our communication though. I sympathize with the people on the wrong side of the problem, where they get spammed with AI content, be it code, PRs, or other requests that require their attention. I understand they need to vent and may exaggerate. But if we want to have a fruitful discussion, we have to strip the exaggeration and anger, and have precise claims and goals. Reducing AI slop, I agree. Shaming all AI content, I disagree. I also don't care for the framing. AI magnified an already existing problem where the receiving end had to sift through and respond to poor quality content. The premise has not changed though. So, instead of someone saying "don't use AI" I'd propose "Low quality pull requests may get closed without further notice or explanation."

I don't think people want AI disclosure to shame people. I am certain that they genuinely wish to avoid AI output, even if it is likely to be well-engineered by traditional standards.

I appreciate the engagement in this slack channel. But Stu's talk still remains in a much more visible place than here, and I think much of the presentation sets a hostile tone.

> I don't think people want AI disclosure to shame people. > I've seen messages from AI opponents this morning in this slack channel expressing intent to shame.

https://clojurians.slack.com/archives/C0B7Q4XRM1U/p1793108497248539?thread_ts=1792857617.587339&amp;cid=C0B7Q4XRM1U

Someone claiming rudeness is a sign that they feel attacked. They might attack in return, as one might do with rude people. Naturally, it is a confusing situation if the supposed rude person did not feel like they were rude and that they are being berated for nothing. I know first hand that it has become common practice for two people to do e-mail communication as meat proxies in some workplaces (AI to both read and respond to e-mails). The problem is not isolated to LinkedIn. I would bet that this is the kind of rudeness Stu and Rich have in mind. AI emails, AI music, AI art, etc. Not ten lines of AI-generated unit tests. Regarding verbose AI slop, I think you are right regarding shaming, and I think it's justified. Regarding well-crafted AI output, I think there is just a general wish to be able to not have to see it.

I am far from the strongest of opponents. In fact, I chat with Claude all the time. Still, I know that if I could magically have an accurate AI detection filter installed as a browser extension, I would have it on at all times and avoid any such content, be it a Medium article or a Github code base. Maybe that will change in the future, but the fatigue is real. I know some people want disclosure for this reason alone.

If they only mean certain specific subsets of AI output it's incumbent on them to specify. Lots of people say quite clearly they intend to shame AI output regardless of quality. I appreciate the even-handed way you consider the situation but there's a real problem of tone that exists.

You are probably right there. I would expect those people to have political reasons for that, which I expect is a loud minority.

I thought Stu's tone in the whole presentation was less aggressive than I would have expected, actually. Are you perhaps focusing on the particular statement regarding rudeness rather than the whole talk?

I still am not sure where in this thread you have seen people wishing to shame people for any form of AI output.

I apologize, I don't have the time to recount from where I'm getting this impression. Maybe you should run a poll in the channel to see what people think.

In Stu's talk there are several moments where the message seems to be "AI users are careless, insensitive, not doing real engineering, etc"

Maybe I'll watch it again. Funny how differently it can be interpreted.