I believe it’s time for this community to declare some kind of policy regarding the discussion of generative AI tools (e.g. LLMs). We have recently had a few threads where projects created with such tools were submitted. Community sentiment appeared to be mostly against their submission, and I ultimately deleted those threads at the time. However, we currently have no explicit policy, so I imagine it would be confusing for any new member to work out our community norms.
I also recognise that while I myself am against the use of generative AI tools (and thus do not wish to read about it here), I do not want to unilaterally declare a community policy on this. I realise this community likely has a broad spectrum of views on this topic, as it hasn’t been directly discussed that much here so far. Please let me know your thoughts by voting in the poll below. Feel free to also reply to add further nuance if you wish. If you’d prefer to discuss some aspect of this privately, you are welcome to message me directly on the forum or email me.
I am assuming as a baseline that we all already agree that generative AI tools should not be used to write messages in our discussions with each other here, so that aspect is not mentioned in the poll.
Ban discussion of generative AI tools and projects created with them
Ban discussion of projects created with generative AI tools only, take no decision on discussion of such tools themselves
Allow discussion of generative AI tools and projects created with them
My vote would be for “allow discussions that mention generative AI tools but ban discussion of projects created with them”.
I don’t see much point in discussing vibe-coded projects. Whatever malleability they may have is only for the vibe coder.
I am not personally interested in discussions of generative AI tools, but I would be OK with simply not reading such threads. Completely banning any reference to such tools sounds too harsh. If it’s a tangent to a discussion, that’s fine with me, and I’d be happy to learn from other people’s experience even though I have personally decided not to use these tools.
Thanks for sharing your perspective. I did wonder whether to have an option like that, but I wasn’t quite sure how to phrase it at first… I have added another option to the poll just now which I believe captures your position, so feel free to vote for it if you wish.
The whole rest of the internet will happily guzzle slop, let these projects go elsewhere where they can be amongst their peers and keep this place for human agency and creativity
Assuming for a moment an ideal world, there is one use case that interests me.
Consider the case where we write specifications and programs. Both these are easy to do, the difficult part is to make sure that the program follows the specification.
If this became trivial with AI, then anyone would be able to write his own program while at the same time being sure that it is correct.
The reason why this is not harmful is because we do not care for the type of proof here, only that the proof exists.
This solves the problem of extending or altering a program while being sure that there are no harm consequences.
There are other solutions to this, like runtime types, so I don’t think of AI as a necessity to this vision, only one possible solution.
I will accept any decision, but in my opinion, it is better to elucidate our values instead of describing the technologies that will get us there.
Everyone here wants to promote human agency and creativity instead of technological social systems that are out of the grasp of the human mind.
Bret Victor, for example, points that it is imperative for democracy to have complex systems explainable, to recover the human agency we have lost.
A more thorough description of our values will filter most of the AI consence that we see on the internet today while at the same time allowing us to be flexible to discuss about the effects of AI.
Given the slop that we see outside, I would suggest that we are strict at first, even banning AI discussion, till we define our values in such a way that will block the AI slop effectively.
Let me point that people in our community are already trying to understand whether there can be good use cases for AI that most probably share our common values. @clemens
Maybe instead/in addition to a policy we should work on our definition of “malleable system” to make clear that an AGENTS.md does not quality. Contrary to the label “end-user programming”, which can reasonably be interpreted as a form of vibe-coding.
In my reading of what is being voted on (ban on generative AI or projects created using generative AI) would include using an LLM based coding agent to write code for a project. This seems mind blowingly backwards at this point to insist that every line of code is hand written.
I was summoned it seems. While I share common reservations and concerns with AI, I am also actively trying to figure out what AI, LLMs, coding agents etc. means for malleable software, and I am exploring the use of it in many forms. I am not particularly interested in projects that are purely generated with the use of AI. Not because I am against using AI in programming as such (concerned for the craft yes, ofc.), but more that the thinking that goes into those projects is typically shallow and hastened.
I’m happy to look at less slop as a pragmatic moderation choice, yet it’s hard for me to formulate a principled position flatly ruling out either of these.
We have good theoretical discussions with no implementation at all. Now, imagine one of these had an attached PoC made with (clearly disclosed) LLM help. Imagine poster making effort to respect our time, making it smooth to try and guiding what’s worth attention.
It’s quite possible that LLMs are biggest practical advance in decades toward end users wielding the flexibility of computers. That’s embarassing and I fully believe we need to figure out making traditional code more malleable anyway — but specifically for malleability, it’s weird to declare AI so irrelevant one can’t even mention it.
There are narrow shapes of LLM integration I’d find interesting. E.g. self-explaining systems one can interrogate to find out what they need to understand & change. Others might find other angles interesting?
Some people I respect here use AI, and I wouldn’t want to drive them away.
Projects optimizing for agents above humans is yet another industrialized step raising the entrance bar, and works against our goals.
This community is working on open problems, we wish to build better kinds of software than anybody currently knows. Throwing “bigger/quicker/cheaper” at it is unlikely to help. (Smaller is intriguing.)
Some people surmount a problem with software by letting AI drive it. Good, but now we have two problems: (1) reduced control & visibility into the (still bad) software (2) malleability of the AI itself. We still wish to improve the foundations.
This place is for human discussion; AI artifacts may be linked if clearly marked as such, and respectful of community attention.
I’m happy to look at less slop as a pragmatic moderation choice, yet it’s hard for me to formulate a principled position flatly ruling out either of these.
Agreed. It’s annoying for me to have to put up with bad submissions. We always had some of those, and now the volume might go up. (I haven’t seen a serious uptick here, but it’s quite possible I missed it.) To the extent we’re just trying to keep heads above water from a moderation perspective, any temporary rules we want to create are fine.
But I think it’s a mistake to be happy in the long term with any rule that uses the term “AI” and its ilk. It is not strategic to blindly accept the frame of your opponents.
For people who don’t care what we think, existing rules regarding spam are completely sufficient. For people who care what we think, any bad submissions indicate a failure to communicate what we consider good.
The best rule I’ve been able to come up with so far is, “share things you found yourself paying a lot of attention to.” Interestingly this covers some “holy cows” that everyone likes to cite but nobody has actually paid any serious attention to.
I haven’t been active here for the past year as I’ve been heads down with research, so feel free to disregard but I still wanted to chime in.
If people’s qualms are coming from ethical concerns with AI production due to intellectual property, environmental/economic impact, etc. then I think there’s not a super valuable discussion to be had (at least to me anyways). The concerns have been debated as nauseam in other forums and I think there extremely reasonable takes that are different than my own.
However, if the purpose is to remove low-effort spam-style slop content I would rather have a more nuanced take than outright banning that hopefully considered the quality of contribution rather than the means of production.
I fully agree with other commenters who have said that llm-assisted vibecoding + open source is not a solution to malleable computing. I do however think interesting work is happening pursuing malleable ends both using ai coding assistant in the production of software artifacts as well as malleable tools that have ai-powered components.
I would support any decision community chooses to go with but I would like some clarification on what constitutes ai content. I use ai tools at various points in my research process and if any of my thoughts downstream of that are “ai generated” that would make it very difficult to contribute. Similarly if I were to vibe code a mvp for something would “human-washing” it by reimplementing it by hand make discussion okay?
The remarks by @tom.larkworthy and @cben suggest that we should formulate our AI policy clearly in terms of discussion, not code. There’s very little code posted here, or discussed about. I have occasionally looked at the code behind someone’s post here, but not so often.
From our “malleable systems” perspective, I expect projects discussed here to be malleable for unassisted humans, so that’s where I would draw the line. But I suspect that everyone has different sensitivities on this question.
All that said, I vehemently disagree with @tom.larkworthy about hand-written code being “mind-blowingly backwards”.
Completely agreed. I think this should be non-controversial. Even if you like AI and used AI to create something, it seems like good sense to not expect continued access to AI. It’s just more resilient.
The use of Claude Code is proposed to be prohibited? Any project that used Claude Code is now a product of a generative AI technology and the majority of this forum support no projects or discussion that uses a generative AI (such as a coding AI)? This would be considered over-the-line?
@tom.larkworthy If you’re replying to my comment above (Discourse isn’t always clear here), what it means is: any project submitted here should be feasible for a newcomer to make changes to without resorting to Claude Code or anything comparable. (It doesn’t comment on use of Claude Code in other parts of the life cycle; that discussion is in the rest of the thread.)
Basically I would try to persuade you with an appeal to the Lindy Effect. AI is new technology. We shouldn’t yet assume we will always have ready access to it. Maybe in 5 or 10 or 20 years that assumption will be obsolete. (My own preference is the longer end of that spectrum.) But for now it still seems important to care about people without access to AI.
Another argument I can make is an analogy with accessibility and retrocomputing. We have computers with gigabytes of RAM but kilobytes are still interesting. We all have abilities we should be worried about losing in future. A wide variety of capabilities seems a crucial aspect of being human, and thinking about a wide variety of capabilities seems crucial to malleability IMO.
I am not directly replying to you @akkartik. I asked a question and I am listening to everyone that followed but they seem to be reinterpreting the vote wording.
the clause being voted for is “Ban discussion of generative AI tools and projects created with them”
@akkartik Yes, I am very much into building a system that can change itself from within, using code. Whether you use a human or agent to write the code is up to you. I personally use Claude Code a lot, therefore, half my code base it tainted, even though I started the project before Claude Code was a thing.
I think I live the ideals of malleable system, fiercely local-first, but I feel the way it is worded and some of the things @jryans is saying I would consider my work to fall into the proposed ban because some of it is derived from Generative LLM technology => we could never discuss it
I don’t think there’s any possible interpretation of the rule that would be inclusive enough to allow something like Lopecode here, as it sits quite far outside of malleability, robustness and sustainability based on the wording of the mission statement of the community.
Seeing that it has thousands of slop commits per month, it’s a prime example of a project utterly hostile to folks contributing(let alone understanding) to the code without using a proprietary pay-to-play product like Claude code.
The library that allows meta-programming on the 3rd party Observable reactive runtime was also made in 2024. The hardest function “observe” that allows a reactive runtime to stream changes from another userspace reactive variable is 2024 era too. That is the foundation upon which the userspace pluggable visualizer sits on.
Its a modular system so some modules are slopped and the key ones have more care. As models have got more good and the patterns of the systems are more established I have used more LLMs.
Of course its a malleability system. You can literally change almost any aspect of the system at runtime and save a copy to disk. That feature is complicated to be retrofitted onto a system not designed to do it, and I am trying to invest in explaining the mechanisms – which are unchanged since 2024 - (the toString trick is explained in Source-last programming). But it is complicated to understand before AI as well, its not the AI getting in the way of understanding it.
I absolutely hammer the LLMs producing slop notebooks, I consider that userspace exploration and I can get an artifact to play with to see if an idea is good or not, but the framework it sits upon its much slower moving.
sits quite far outside of malleability, robustness and sustainability
I do not personally think my work is incompatible with malleability community if the criteria it is being judged upon is malleability. Having a local-first version of Observable Notebooks that does not depend on an proprietary API server I also think is more robust and sustainable than not.
and I fully believe we need to figure out making traditional code more malleable anyway — but specifically for malleability, it’s weird to declare AI so irrelevant one can’t even mention it.
:
