Clojurians Log v2

They don't know that the Dow is over 500 thousand.

The Index is not equal to economy growth and economy growth is not equal to welfare of people

Any thing which reduces demand side of economy (total salary which is got by people) would cause economy collapse

Collapse is probably a strong word. Civilizations with even wilder wealth inequality still existed and "prospered" (for some definition of the word) in the past. Some countries today still exhibit enormous inequality, e.g. North Korea or India (no intent to offend anybody, I'm speaking from stats), and they don't "collapse" in a dramatic sense. You can use this as a blueprint for the type of economy the rest of the world should be expecting.

Yes collapse is not a correct word. However lower economic growth or recession could be explain it. Statistics also shows this in the book "capital in 21th century" , it shows growth of economy has negative correlation with inequality & wealth distribution for the western economies. BTW North Korea economy is completely opposite of India.

He's a grumpy old man before his time.

But smart and thoughtful!

Many smart and thoughtful people out there. Some of them have committed atrocities. Not the only metric that I use to decide whether someone is worth listening to. 😊

Yeah… I suppose if the world keeps going in the direction it’s going I'll be ok to have lost. Winning doesn't feel like the point anyway.

The new monk class that uses just the right amount of AI is gonna win? Maybe if local models become an actual thing. But, for now they are at the mercy of the labs.

Ha! I realized that I'm probably a priest. But I prefer paladin. Lawful good of course 🙂

Why defensive?

@andersmurphy In my opinion, the situation with local models is already pretty good. Sure, hardly anyone runs something at home that can compete with SOTA in most areas, but it's still no problem to use another provider that you like. ofc. then there's the question of FOMO and wanting to use the best. Both Anthropic and OpenAI are working to make their models impossible to distill, which might cause issues down the road. But that's a question for tomorrow 🙂

The problem is 7% of their revenue comes from companies distilling their models. 😂

But yeah, local modals are getting there, and are getting better.

Based on what this guy says, shouldn't he be in favor of voluntary disclosure of (type of) AI use?

He would have a 1 paragraph opportunity to separate himself from those he says use AI inappropriately. He can let us know how he uses it appropriately. He might avoid that comment in lobste.rs being the highest voted (though I know that's not a guarantee)

Those who don't want anything to do with software developed with AI would have their confirmation, and would hopefully leave this guy alone.

Those who don't care about AI use, wouldn't be dissuaded by the disclosure anyway.

For that whole spectrum in the middle, they would have one additional piece of information to inform their decision if they should dedicate time to this project. Is the added value of this software compared to an incumbent solution worth it, when also taking into account a new less-proven software development technique.

I assume he doesn't want to trick these 'priests', right? So what's the harm of pre-emptive disclosure?

I'll still wait until models (even local) are trained on green energy, using opt-in content, and when the LLM tools don't have dark patterns that make them more addictive than they need to be. Oh, and when they can't be used to exploit labor. That might not be until enough people have unionized I suppose.

> I'll still wait until models (even local) are trained on green energy, using opt-in content, and when the LLM tools don't have dark patterns that make them more addictive than they need to be. Oh, and when they can't be used to exploit labor. When we were doing the planning for Season 2 of Politechs, one of the questions that we were left with from Season 1 was whether there could be such a thing as ethical AI. Ray came up with an acronym that captured some of the requirements as we saw them: H.E.A.V.E.N.L.Y. AI: • *H*igh quality - clarity over the data, tokens, parameters and weights • *E*thical curation and purpose - licensed and representative data, tools to support and elevate human endeavours • *A*ccessible - not beyond the means of individuals or schools / universities • *V*iable - can be created and operated by individuals / small organisations • *E*nvironmentally sustainable - low power and machine resources • *N*arrow focus - to ensure the quality and ability to control the model • *L*arge Language Model - don't toss the Transformer architecture out with the bathwater (debatable, so let's debate it!) • *Y*OLO - so you may as well YOLO well By the end of Season 2, we had answered the question to our satisfaction: no. A resounding no. We explain in the season finale: https://politechs.dev/episodes/s02-heavenly/

> So what's the harm of pre-emptive disclosure? Because that disclosure has to be kept actual to be useful, and it might invite more dialog that clearly Zach doesn't want. If I were to publish something as OSS, I'd probably also skip on any disclosures. Those who might want to use the code, can judge what's up themselves, without constant (albeit small) extra effort from me.

@jmglov Love that acronym. I’ll add “D” to make it HEAVENLY-D — stands for “Dark pattern free”

Makes it even less likely haha

Humans are also full of "dark patterns". ;) Avoiding a potential for their surfacing deliberately and completely means not working with anyone or anything at all, apart from software with fully observable and predictable behavior.

Yeah.. false equivalency. Humans can be negotiated with. We all have defenses against manipulation by humans baked into our makeup and cultural norms. And our brains have hard wired social characteristics. Software that plays on those characteristics is immoral in my opinion. Humans that interact with each other are not. They have intrinsic worth. Machines do not. This goes directly to whether or not an AI is a moral agent. Can it be punished? Can you negotiate with it and have that stick forever? Can you hold it accountable for failures to meet the contract in a court of law? So far, no. A single instance of an agent is ephemeral and actions it takes are your responsibility. When you prompt it wrong or allow yourself to be manipulated by its dark patterns you're the only one to blame. So, just like I choose not to drink so as not to be addicted to that chemical, I choose not to use AI so as not to be manipulated by its dark patterns.

Not sure what worth has to do with anything, but someone that wants to manipulate you will attempt to do so. Depending on their skill, they might succeed, without you knowing - there would be nothing to negotiate. Note that I'm not arguing about the degree. A piece of software can attempt manipulation for as long as you feed it tokens. A human will get tired much quicker. So I'm talking only qualitatively here. > This goes directly to whether or not an AI is a moral agent. Can it be punished? Even if you have the ability to punish someone who is trying to manipulate you (assuming you notice it in the first place) - the attempt has already been made. > Can you negotiate with it and have that stick forever? Absolutely not. Just as I cannot guarantee that with a human, unless I get rid of that human in my process altogether and try working with a different one. Who might be more moral or just more skillful in manipulation. But woe is me if it's my boss. Then the whole process must go out the window. > Can you hold it accountable for failures to meet the contract in a court of law? Also not really relevant to my point of ability. Unless you are willing to accept a possibility of clandestine manipulation given a possibility of holding a human accountable. > When you prompt it wrong or allow yourself to be manipulated by its dark patterns you're the only one to blame. Absolutely. But it's trivial to not get manipulated - just... don't talk to it. There is no need to have a discussion with a code generator. You can just tell it to do X that's well-scoped and easy to validate, then judge the result and proceed from there. Unless some magical LLM can manipulate humans by writing code, I don't see anything risky in that. > So, just like I choose not to drink so as not to be addicted to that chemical There's a difference though, between drinking liquor and, say, kefir that might have something like 0.4% alcohol. Or avoiding any kind of fruit that's ripe and has been on a shelf for a few hours (some bits might get fermented without you noticing). Or, keeping up with the hyperbole, flooding your body with antibiotics because there's a chance that bacteria in the intestines are fermenting something.

I’m not sure the line is quite that clean. Even between humans, understanding is often incomplete, and feeling understood isn’t the same as actually being understood. Someone trying to avoid conflict might agree with us and make us feel heard without really understanding our perspective. We wouldn’t necessarily call that manipulation. An LLM often does something similar. That doesn’t give it the same moral status or accountability as a person, but I don’t think appealing to our social instincts is, by itself, what makes an interaction immoral. The question is what that trust is being used for.

That goes to say, there's no need for trust when you can validate. So as long as you validate stuff, don't trust anything. Corollary: delegate to tools and people only things you either don't care that much about, or things that you will validate. The degree of care will be different when working with tools and with humans. Or when working with different tools, or with different humans. I have an ex-colleague whom I'd trust with any code whatsoever, and dozens of other ex-colleagues whom I wouldn't ask to write anything that's even remotely useful.

I don't really understand the "yeah but humans do that too" rhetoric. LLMs are not people. People have intrinsic worth, software does not. Anthropomorphisation of LLMs is a huge problem.

Considering humans are all different and they adapt to changing circumstances and they can be reasoned with and are motivated to accept feedback and so forth, and ultimately if I happen to be a bad fit for a team I can leave. These HELLISH (I like the politechs podcast acronym) LLMs are all the same and exhausting to deal with as I have to continually fight the same dark patterns. And soon we will all be forced to use them to earn a living. Therefore, since we have no choice, and the mental overhead to use them responsibly is too high, the tools need to change. And I intend to advocate for that ruthlessly. Or retire early. Ha!

Oh yeah, I forgot to post HELLISH: • *H*yperscaled • *E*xpansive • *L*arge • *L*anguage model • *I*mperial • *S*ucks up all the energy, water, and air in the room • *H*uman suffering

I'm also getting sick of the AI apologism (new word?) in this channel. Every other day I have another huge spike of anxiety defending against it. It's not worth it.

@jmglov I agree but comparing a specific effect on the user isn't the same as attributing human qualities or moral worth to the system in my opinion.

Ooh I will read this now

@stephen676 Totally agree. Some of the discourse in this channel feels like https://wondermark.com/c/1062/. "Of course I'm an AI sceptic, but just explain this one thing to me and also give me permission to use AI in a good way but not like those people in the other channel because I'm a sceptic remember?" I'm not a sceptic, https://anthonymoser.github.io/writing/ai/haterdom/2025/08/26/i-am-an-ai-hater.html.

Sealioning seems more appropriate than the useful idiot categorization — one is a behavior while the other is a label. I don't want to demean people.. but I have no problem demeaning their behaviors.. 😅

That cartoon is gold

@U0BHKM127SQ Only one with that degree of trust, yes, Plenty with varying degree. And please be civil. @jmglov There's a huge difference between equating LLMs and people (which I don't intend to do at any point in time) and arguing that there are common characteristics (which I am doing). My point is that if we choose "dark patterns" as something that completely prevents us from interacting with an entity, then we must also not interact with humans. If we do still want to interact with humans, regardless of "dark patterns", then either "dark patterns" isn't a good metric by itself (could be totally fine even if the only addition is "unless it's a human"), or the avoidance of "dark patterns" might still be manifested without having to completely avoid the entity that has a capacity for them. @stephen676 > I'm also getting sick of the AI apologism (new word?) in this channel. The original announcement of this channel might indicate why. Skepticism is not synonymous with complete avoidance and swearing against. LLMs are terrible in many regards, and are still useful in many other regards. The use can be extracted without having to also be exposed to the terrors - the degree depends on many factors. Could even be 0 if your moral compass does not let you use anything that was made with any amount of suffering, or something like that.

@zikajk Just what argument are you trying to make? That last sentence didn't make sense to me in the context of the discussion.

On the subject of moral worth, it makes perfect sense to ascribe a level of moral worth to a thing based on its effects on people. Tracking cookies are evil because of how they allow tailoring of ads to you based on where you've been on the internet thus incentivizing you to buy things you don't need. Heroin is evil because obvious reasons. LLMS are HELLISH. I don't even need to explain because @jmglov did such a good job. Tho I still need to add the -D because dark patterns causing high levels of addiction and making people dumber. To differentiate from the human suffering caused by environmental catastrophe, disease incidence increasing for those living near data centers, etc

@stephen676 Well, I don’t think heroin is inherently evil either, so perhaps that’s where our intuitions differ. But you say we should judge things by their effects on people, and that’s exactly what I meant by “what that trust is being used for.”

@U0BHKM127SQ If you can't think of a good way to negatively characterize a group of people, just don't characterize at all. As simple as that.

@zikajk I personally do my best to avoid building or using software that manipulates people to use it against their own best interests. Goes back to my deontological view (see Kant) — using people merely as a means instead of as an end in and of themselves is morally wrong.

I wonder sometimes if socialists are more deontological. And rabid capitalism / fascism seems to come from hyperuse of utilitarianism.

@U0BHKM127SQ I was not characterizing them, however. I only said that I wouldn't ask them to do anything, insinuating that I don't have trust in them. And it was nobody on this Slack. And even if I were to call everyone names here, it wouldn't excuse anyone else doing that. It would only mean that I would have to change my behavior or go.

The sea lions cartoon, while funny, is completely unrelated to anything here. This is not a private conversation, not someone's home. This is a public forum. Posting here is, of itself, an invitation for discourse.

@p-himik Disagree. The comic makes perfect sense in this context.

How?

The name of the channel is an invitation for skepticism not apologism. A sea lion embodies an apologist wading into a place designed for skeptics.

Expect the same treatment from a skeptic in this channel as you would if you were pro life waving a sign at a massive pro choice rally.

Have you seen the original announcement? As I said, skepticism is not a purely negative flux. It's a questioning attitude. Hating is not being skeptical, at least not in the spirit and purpose of this channel. And a questioning attitude presupposes answers, and some answers are bound to be non-negative. > A sea lion embodies an apologist wading into a place designed for skeptics. Maybe in the first three panels, but that's it. In isolation they don't make much sense though. The full comic would be more apt if people here were blasting guys that have, say, hairy bellies, and then I came by and started pestering everyone involved in their DMs, regardless of multiple "go away"s. But this is not a DM. Nobody is targeting anyone. Saying "I found a use for an LLM" does not make someone an apologist and provocateur, or even just a non-skeptic. Anyone here can just mute the thread or leave the channel or block any replies of whom they don't want to see.

I love the apologist characterization! Definitely a lot of that happening. Having emotions and feelings is normal. The models still stole my thoughts, ideas, code, and so on, distilled it into a product, with no recognition, no thanks, no compensation, no acknowledgement, and sold it back to me at a price. I'm basically buying a stolen bike after having had my bike stolen. Now if you come at me with the "greater good" morality. Or with things like, but ignoring that, don't you see the utility? And so on. It feels like sealioning 😛. I'm not being hateful, as towards people, but I can be angry, pissed of, annoyed, these are all justified emotional responses.

But it's not even "greater good". On that, to me the question is still open - I have no clue. The point is that some people have found ways of extracting use from LLMs. And to those people, using an LLM might be worth it. How "worth it" is decided is a subject to a debate, of course - hence me engaging with the "dark patterns" angle. I saw a flaw in logic and I tried to address it. Most of what I wrote was left without any feedback so I still have no clue whether it was indeed a flaw in the logic, or a flaw in my understanding. But eh, it's fine - we aren't a parliament.

> There's a huge difference between equating LLMs and people (which I don't intend to do at any point in time) and arguing that there are common characteristics (which I am doing). A rock and a person have common characteristics, too. But only one has agency. This is also true of an LLM and a person.

> But only one has agency. This is also true of an LLM and a person. Of course. But my angle on "dark patterns" had nothing to do with agency.

@didibus Wouldn’t a closer analogy be a musician who listened to your music, along with thousands of other people’s, and now charges you to attend their concert? You’re paying for a performance your work helped shape, rather than buying your own record back. https://press-pubs.uchicago.edu/founders/documents/v1ch16s25.html.

Neat quote!

> LLMs are terrible in many regards, and are still useful in many other regards. The use can be extracted without having to also be exposed to the terrors [...] This I don't agree with. How can you avoid being exposed to the terrors? > - the degree depends on many factors. Could even be 0 if your moral compass does not let you use anything that was made with any amount of suffering, or something like that. Almost everything is made with some amount of suffering, unfortunately. I drive a car and fly in airplanes, both of which are made with and contribute to suffering. What I don't do is try to convince people who are rightfully pointing out how bad these things are that my use is somehow "good". I'm not necessarily saying you're doing that, but there certainly are many people in this channel who are.

> How can you avoid being exposed to the terrors? It depends on how you understand the words "terror" and "exposed". Say, I'm running a local model. A static artifact that doesn't access anything and doesn't consume any resources apart from the electricity that I pay for. All the terrors that have been created by the process of training it have already affected me to the degree they can - nothing I can do about it. Using that local model will not incur any new terrors. So I'm not exposed to anything at all then. It's kinda similar to the horrors of the Unit 731. It's inexcusable and if I were able to, I'd rewind time and made sure nothing like that existed. But it has produced an immense amount of useful data that we still benefit from today. If we were to somehow decide that that data must not be used, it would lead to more suffering and wastage. > What I don't do is try to convince people who are rightfully pointing out how bad these things are that my use is somehow "good". Neither am I. I'm not labeling what I do as "good" here. I'm pointing out that I myself have found use, for myself. Maybe Zach has as well, to some degree. > there certainly are many people in this channel who are. Maybe, although I can't say I remember noticing that.

To slightly correct myself there - I'm also saying that other people might find use in LLMs themselves, for themselves. Adjusted for "exposed to terrors", of course. That's why the "dark patterns" angle got my attention, since to me in and of itself it doesn't seem like a cause for any decisions whatsoever. It's just one potential condition that doesn't work in isolation.

> The point is that some people have found ways of extracting use from LLMs. And to those people, using an LLM might be worth it. It depends how this is framed though. I have myself had to buy a stolen bike to replace the one stoled from me. It IS worth it for me to buy stolen bikes, as I still want to be a good cyclist, I still want to compete in cycling races, I need to cycle to get to places, and so on. There were those skeptics who never cared that my bike was stolen or that the newer bikes are from theft. They only cared that stolen bikes outperform those that are not. They were skeptical they did at first, and now see that the stolen bikes perform even better so are changing tunes and becoming pro-stolen bikes for cycling. In all cases, I am still justified to be angry. And utility is pretty much irrelevant to that feeling. So if you justify those who use stolen bikes from a utilitarian point of view, when engaging with someone talking about the issues with bike theft, it's a poor framing. It's kind of sidelining the topic almost. And anyone who dismisses the theft, and justifies it due to the utility benefit it eventually enabled, in my opinion, is apologist, if it's to make themselves not feel guilty of it.

> if you justify those who use stolen bikes from a utilitarian point of view, when engaging with someone talking about the issues with bike theft, it's a poor framing Are we in the same thread? There was barely any mention of property here. As I said twice already - I'm arguing purely about "dark patterns" and a possibility of utility. If you don't accept LLM usage because of theft, I feel you, but it's a separate topic altogether. I'm not dismissing the notion - it's just not something that I'm engaging with in this thread.

> Maybe a #ai-haters channel would be great Not something admins would be able to affect though, short of deleting it if things get too heated and people complain. So I doubt we ourselves would make and announce it. Don't want to invite more hate. But anyone can still create it and invite other people manually. We also have #C0BLJS63J2Y. Kind of the opposite - not "hating AI" but "loving the absence of AI".

@zikajk In my opinion, no. When I published work in public, or even paid commissioned work, I did so with reasonable assumptions. One of those is that, I'm totally happy for other human beings to get inspired, and imitate or build upon my work. This use provides me with value, it supports the communities and collaboration I enjoy, and makes me feel happy. The end result is a thriving and positive society/environment for me and my fellow humans. I never expected, and reasonably so, that one day a model would compress the entirety of human produced work, and enable the generation of new work at the press of a button, at a scale which no human being can compete with. In the process, killing those same human communities and collaboration that I was trying to support by publishing my work. The end result of which, well we're not sure yet, but could lead to a bankrupt, lifeless and struggling society/environment for me and my fellow humans. So assuming the prediction of the outcome is that of a struggling society, than no it's not similar at all, because the musician who's concert I'm at now is exactly the outcome I wanted, this is not. The only caveat is, I can't say for sure yet, the optimistic side of me is that maybe it will yield the outcome I wanted as well, maybe AI will lead to a prosperous and deeply fulfilling society for me and all my fellow humans.

@p-himik I may be misreading the discussion, but much of the thread seemed focused on the ethics of AI rather than its utility for a particular use case. I interpreted some of your responses as addressing those ethical concerns primarily by pointing to AI’s perceived utility. Was that your intention, or is there another aspect of your argument that I’m missing?

@didibus Sure, I understand that, and those are concerns that go well beyond both my analogy and the bike theft analogy.

> Was that your intention, or is there another aspect of your argument that I’m missing? For the fourth time now, I was addressing only the "dark patterns" angle and the possibility of utility. Nothing else.

“If we were to somehow decide that that data must not be used, it would lead to more suffering and wastage” This needs more explanation. It feels like sunk cost fallacy. I think of it as the same as blood diamonds. I don't buy blood diamonds because of how the miners were treated. I can't do anything about the fact that they were mined. But I can do something about whether I buy them. If anything, it reduces demand. For the blood diamonds. Not buying them doesn't make any new waste. It isn't my fault they were mined. But if I buy them I offer feedback that they were valuable. I’d rather not.

A sunk cost fallacy is where the goal hasn't been reached and dropping everything now would ensure that it's not reached, but continuing cannot guarantee anything. Here though, the goal has been reached, even though we call it that only post-factum. We already have the data, we don't have to do anything to retain in - we only have to avoid deleting it globally and prohibiting its usage, that's it. > I don't buy blood diamonds because of how the miners were treated. > it reduces demand Fully on-board, zero doubt. However, it's not equivalent to LLMs, local models specifically. Nobody benefits from your usage of them. Well, except for your electric power provider, if you have one.

Feedback comes as people talk about how useful they were. That encourages more production. It's just good marketing.

Since people only index on usefulness there is no incentive to produce models ethically

Wait, we started with "I don't buy" but ended with "people talk". What people do is not something you can affect, in general. If you care about it, you can simply not mention when an LLM was useful to you. Oh, wait... Did I just solve the riddle, did I uncover the secret plot? :D > Since people only index on usefulness there is no incentive to produce models ethically There is, it's just very local, sorry for the pun. Similar to many other things, like e.g. people boycotting Nestle or moving from Windows to Linux.

“Secret plot?” What kind of rhetorical device is that? Feels a bit demeaning. There are many more facets to a person’s choice. I was simply following a single thread.

Nah, just joking. Extended "if I don't talk about it, others won't benefit from the talk" to "if I talk negatively about it, others' potential for the benefit will be reduced".

> Some of you just need more exposure to the self-proclaimed “AI nerds” with business backgrounds (and zero technical ability) that populate corporations. Why, what do you think it would achieve?

Genuine.

> The sea lion focuses entirely on the technicality of debate rather than the social reality of the situation. > > The sea lion is completely blind to context. He firmly believes that because he is keeping a "polite" and "civil" tone, his invasive, exhausting pestering is completely justified I think this can happen on both side. AI "skeptics" can fail to realize many don't care about ethics, externalities, or the longer issues that could derive from it, they only care about the utility of it for their next coding project. Similarly, the "non-skeptics" can fail to realize ethics is the main area of concern, greater impact on society, community, enjoyment, career, etc. is what they care about, and utility is not a redeeming quality if it's not useful to the above.

And off course, this is a spectrum where people exist on some continuum. I mean more that the discussions sometimes appear to struggle because of this lack of context of where every participant lands on this spectrum and what the topic is currently focused on. I feel #C068E9L5M2Q is where debates on utility take place. As that's where people discuss what works/doesn't, experiment with utility, and so on. If you are skeptical of its utility, you will learn more about it there. I feel #C0B7Q4XRM1U is where debates about ethics and long term benefits beside purely technical delivery is discussed.

True. But what tends to happen doesn't mean that it's what needs to happen. If there were no push-back against anything here, regardless of how incorrect or generalized something might be, it would just be one giant echo-chamber. I don't mind echo-chambers, but I do mind turning #C0B7Q4XRM1U into one. If someone wants an echo-chamber, they're welcome to create one, as I mentioned above.

I agree wholeheatedly, we need disagreement. We need to be skeptical of our own skepticism as well.

If I saw a thread like "I am forced to use AI at my job and it kills my soul" with all other participants chiming in and sharing their experience, I wouldn't even participate in such a thread - I have nothing useful to say in such a setting. And I don't think I've seen any genuine provocateur in any similar thread here, but if anyone has a link it'd be interesting to peruse. But if I saw something that read to me like e.g. "all X are Y, and hence if something is Y then it must be X", you bet your sweet $ I'd chime in.

@U0BHKM127SQ Maybe, I'm not in a position to judge. But I have my reservations on all and any accounts. :) One one hand, I can see all that happening. On the other, also plenty of explicit or implicit admissions of companies having shot themselves in the foot with AI. > Technical ability is being devalued (this Slack is also an example of that) Not sure how it's an example, I don't think I have seen anything here that would make me think about such devaluation. If anything, it's the opposite, with things like https://clojurians.slack.com/archives/C0B7Q4XRM1U/p1792101323861059?thread_ts=1792100736.103719&cid=C0B7Q4XRM1U. But things can still change, in any direction, and I am least of all an oracle.

What I'm noticing about myself is that I'm on a hair trigger basically all the time now due to all this. So, sorry about that.

@stephen676 Genuine question - why participate in an AI-related channel at all then? Just as an outlet?

That's an emotional statement though, not factual. Just to reiterate - not that it's not a valid reason, but it's not something that can be properly discussed. I think I saw some of those messages and I remember feeling dumbstruck. Worthless?! None of the channels went away, thousands of people are here, you can ask anything and anyone, you have direct line to the core team, you can participate in the development of the language itself or anything in its ecosystem that uses this server, and so on. How is it all suddenly worthless just because #C068E9L5M2Q exists and is (tentatively) more active than other channels at the moment?

Wild. The only thing I can say, really.

Top people retiring early in their careers because they hate AI

Well @p-himik — as all humans I seek community. Most of the Clojure community is now a threat to me, so I came here to find a refuge. I was apparently wrong. There is no refuge.

I suppose depends on what you perceive as "refuge". Participating in #C0B7Q4XRM1U, even if it were purely #ai-haters, does not feel like getting away from AI to me. ;) If anything, it's the opposite, just with a negative sign. In any case, as I mentioned there's #C0BLJS63J2Y. Or #C06E3HYPR, which is also AI-free. And many other such channels where AI isn't used at all. In fact, even #C03S1KBA2 is AI-free apart from very rare threads where we immediately ask the OP to move it to #C068E9L5M2Q.

I don't want to hide from it. I want to hear about other people’s experiences that make them more or less skeptical. I'm still quite certain of my ethical views but I can't be in a cave.

I will also check out other channels for sure.

Also, just in case, if anyone wants to chat more about programming in general, or Clojure specifically, and not have to deal with AI in any way - you can start helping beginners. Here we have #C053AK3F9 for that. This October specifically was a quiet month for some reason (cannot really blame it on AI since September and prior to that it was relatively active), but it's still getting usage. And there are other forums. Reddit gets beginner questions from time to time (usually meta, like the most recent "is it supposed to be this hard?"), ClojureVerse gets some traffic. Clojure Camp is active and has both mentors and students: https://clojurians.slack.com/archives/C02BJCKN0R4/p1781767917547999. There might be other beginner-oriented resources, I'm just not exposed too much to them myself.

> But then why criticize skeptics for posting in #C0B7Q4XRM1U? I think our definitions must differ. What I see a lot of in this channel is someone posting something thought-provoking, then somewhere in the thread (often near the top), the discussion shifts to effectively asking the AI-critical people in the thread to defend their position. > BTW tech worker organizers I've met don't characterize people as begging for "permission to use AI in a good way". They'd be seen as stripping away people's dignity, and fail to build solidarity. I think they're onto something Yeah, so I am a tech worker organiser, but that's not what I'm doing in this thread. Worker organising happens largely 1:1, and largely in person. Why are you trying to weaponise solidarity here? Solidarity is built through struggling towards the same goal, not excusing behaviour that runs counter to the goal, just because someone might potentially be convinced to support the goal if you're not too mean to them. > Maybe a #ai-haters channel would be great. Ah yes, because now criticism of AI is not even welcome in the one channel that was created so we'd stop going on about inconvenient truths in other channels. I probably shouldn't have responded to this trolling, but there you have my response anyway. And now I'm done with this thread.

someone posting something thought-provoking, then somewhere in the thread (often near the top), the discussion shifts to effectively asking the AI-critical people in the thread to defend their position

@jmglov I've always seen this channel as a place for the non believers to go through re-programming until they become true believers ready to be assimilated into the borg.

> a place for the non believers to go through re-programming A matter of perspective. A deep skeptic here would mostly feel being jostled towards lesser skepticism. But a deep believer would feel the opposite. Doubt there are many of those here though.

In the spirit of Disco Elysium, maybe it could be renamed to #the-world's-most-laughable-ai-centrists. :)

Nah, the best haters aren't dogmatically accepting anything. I started as a sceptic and became a hater through lots and lots of thinking and discussing in good faith. I don't think I am in possession of any sort of universal truths. I also don't want every conversation about the risks of AI to be met with "well actually local models".

Anyway, the conclusion that I'm drawing from this is that this community is no longer one that I feel welcome in. In fact, I feel unwelcome in most programming communities these days. I either have to grin and bear it when people are gushing over "AI" or open my mouth and be seen as that annoying vegan in the room.

FWIW, at least when it comes down to this thread and my perspective, I don't think there was anything that would make people start seeing you as an "annoying vegan". Maybe just "vegan", but that's hardly an indictment. Just a normal discussion with maybe some situational misunderstandings along the way.

Never realized this channel was actually called #debate-an-ai-skeptic …

I should have known! 🤦‍♂️

Yes, because I have mentioned the original announcements at least twice. :) It explicitly talks about debates. https://clojurians.slack.com/archives/C03RZGPG1/p1780682339639849

Yeah I stumbled on the channel but didn't see the announcement.

Non-debates can also go here, that's fine - not everything has to be a debate. The description of the channel doesn't even mention debates. The point is that an absence of debates and opposing views should not be expected. As I said, this is an open forum, not a private party.

@U0BGC50RM26 Some sealioning is happening, I think it was a good call out. I see so many AI debates where someone counters one argument with an unrelated other weaker variant, local models is a good example of that. Local models don't solve many of the problems with AI, have a minuscule market share, and have their own set of problems. Same with dismissing a discussion about agentic AI and insist that you can also just use it as a way to ask it questions. Both of those are like missing the context of the conversation, and if you keep bringing that over and over when nobody is talking about local models or this very constrained use of AI, then sealioning is an apt callout. Otherwise the truth seeking is actually slowed down since that detracts from the main discussion.

@U0BGC50RM26 No I don't think that was sealioning. I'm not trying to call out anyone in particular, I think I've seen sealioning on both sides and I've probably done some of it myself. I found the callout about sealioning relevant in that sense of having that awareness. Because I've started to realize the context of discussions are often unclear, where people are coming from, what part of the problem they are focused on, etc. is not always obvious or is easy to get wrong. For example, it's very different to complain about AI in the context of a workplace that already expects that AI is baked into the estimate, and must achieve 10x increased velocity. You no longer get asked to use AI as a tool, you are asked to deliver the thing you said will take you 10 month in 1 month instead and told AI is how you will do it, don't fail or your fired, don't show skepticism or you'll be re-assigned to something unimportant. Now someone might argue against this person for one hour, say that you can use it to augment yourself, uncover bugs, etc. You don't have to make a software factory generating CRs while you sleep. It will start to feel like trolling to the other person, but it's really that we're blind to the context. Because one person isn't even arguing if there is a way to use AI that doesn't make you lose your skills or your connection with what you work on, they're arguing that AI is polluting the profession, infecting leadership, creating insane expectations and setting new standards that are worrisome. Both participant will think they are logically right in their stance, and they very well can be, but are missing each other's context.

This seems like another cautionary tale of how community in open source will become harder to hold together. In this case, the tools automation got better and the respect that held the group together shattered. I don't know if the full weight of the problem can be placed on LLMs, but they are at least the catalyst.

This is really a new powerful curse. I initially think AI creates a new lisp curse. I heard people said this lib wont be here if not ai and it has good quality not ai slop why nobody use it. But yes, a new type of indifference comes. On the one hand, you can easily get one thing useful to you, on the other hand, people also think, “ why don't let my agent to have a my version”. So no community is needed, no sharing is necessary. Let's see how it will go.

> lisp curse But it also reads like some disrespectful community actors would "snipe" a near-complete project by throwing a lot of tokens/money at it. Like the cyclist that drafted the whole way through the race, they'd pull away at the last minute and claim "victory" to have completed the project. That's disheartening for the others who put in quite a bit of early labor* in the project but who miss out on the kudos. (* this labor may have been used for training too)

Yeah.. future software projects I make will likely be invite-only. And not open source.

It's really simple. Gatekeeping to survive will be mandatory as well as saying sorry no LLMs. LLMs are for people who want the end result not the journey, the community, the mentoring, etc. Which is the whole point of decomp/demake/indie communities. Everything is going to go dark.

Kinda bummed it's ruining game jams. As despite having strict rules against LLMs. I often see new entrants take massive shortcuts and skip the whole point of the jam, which is to learn how to make a games. That being said there's still enough social pressure in indie games/gamers against slop that maybe it's enough to keep people on the right track.

We need more people to see the human cost of all this stuff. Poor Aphyr…

In my experience, using ai keeps me in the hammock zone more. I can define the module level and system level design and boundaries, let an agent create the implementations, and quickly move on to the next phase of system level design without having been distracted for hours on boring implementation details.

I completely agree that the hammock has been underutilized and is more important now than ever. It's very easy to run quickly so running in the wrong direction is very costly if you make bad, irreversible design decisions. Fortunately I think it's also easier now to build systems quickly enough with few enough people that it's easier to maintain conceptual integrity of a system. I also think it's easier now to revise a disorganized system to become more organized if you have the right conceptual vision leading the way.

+1 for an increased hammock zone. All the time previously spent on typing, plumbing, stuffing, etc., went into designing the high-level structure and experience.

Clojure increases productivity by a factor of 10. AI increases productivity by a factor of 10. Which means, Clojure becomes obsolete. Both together increase productivity even more. But in the industry this is not relevant, because there will be no new generation of Clojure developers if everybody uses AI. You must decide if you want to solve problems efficiently or if you want to work efficiently with inefficient solutions.

I always thought rapid prototyping is part of design. Now you can prototype a few approaches using AI and go to the hammock with concrete findings on each approach. This was not always possible to do prior to AI era.

@hbrng.computer After I showed some people what can be done with a REPL, they admitted that writing everything in Rust might not be the best idea. I believe that Lisp languages offer a lot of advantages for LLM-driven development, while also eliminating the friction it used to have when people had to write them by hand...

> Clojure increases productivity by a factor of 10. AI increases productivity by a factor of 10. Which means, Clojure becomes obsolete. Both together increase productivity even more. Adam Tornhill gave a talk at https://funcprogconf.com here in Stockholm last week where he talked a bit about his LLM workflow, what made him choose Clojure 10 years ago, and which of those criteria still hold today. The video isn't up yet, but https://drive.google.com/file/d/17PEdPaeAqeRUpa-Nri_I1e48N7W8oPa3/view?pli=1 are. Here's the one talking about why Clojure:

@soltanzadehramin I also did some thinking about how the Clojure philosophy meshes with AI, and gave a talk on this at Babashka Conf this year: https://youtu.be/JjBJFsnplMg?si=yzmUhjzQug4ocr8J

Why would a sceptic (of anything, not just AI) also be a proponent?

If "skeptic" is "fully against" and "proponent" is "fully for", then it's impossible to be both. Possible to be neither though. Or both, but if the definitions are changed to "has some things against" and "has some things for".

I don't think the latter definitions work in practice. I think most would then be all of: capitalist, anti-capitalist, racist, anti-racist, religious, anti-theist, etc. Most people would be most things.

I can only speak to my workplace and personal experience. At work, I'm seeing less hammock time for sure. It's not even about thinking of how to solve the problem, now it's sometimes to the point where people don't even know how the problem was solved, as so much of the thinking, the decisions, the ideas are left to the AI, and superficially reviewed. For my personal projects, it's more nuanced. I feel I might be doing as much hammock time as prior. At first I thought maybe I was doing more, but then I realized that feeling comes from the hammock time being compressed. Since I can deliver more functionality in a shorter period of time, I have to more quickly hammock, resulting in this feeling that I'm hammocking 24h/7h. What I'm unsure is if compressed hammock time still results in the same quality of thought.

@soltanzadehramin Different definitions require different approaches. They can't all be lumped together.

> What I'm unsure is if compressed hammock time still results in the same quality of thought. Just in case, though you probably already know: it's definitely not great if there are frequent context switches and no proper breaks. But my gut feeling is that compressed hammock time is fine if the context is the same (or different contexts are still similar enough so you can "flow" through them without much friction) and there are proper breaks at regular times.

Some people believe that agentic coding is generally a more efficacious way of programming. Others think it is generally a less efficacious way. Those are the sets I had in mind.

The categories should be much more fine grained than that. A problem with ai does not invalidate all uses of it. A good use of ai does not make all uses of ai good.

I do feel sometimes "sleeping on it", really provided new ideas or clarity of thought. And compressed hammock time basically removes some of that "sleeping on it". Also coding did force me to scrutinize every detail, and focus on varied parts of the problem, where now I rarely zoom in, I am constantly zoomed out and thinking at that level. Hard to say really. But I do think there's an easy fallback to get lazy mentally, so at least you have to kind of force yourself to remain actively engaged if you still want hammock time when using LLMs, because you can just let them do it on your behalf.

Building in reversible steps and composable/separable parts is very important. If you have one of those delayed hammock realizations you need to be able to apply it without being constrained by past hasty decisions.

I think the case I was making was relatively clear based on a few assumptions. So far people seem to neither have disagreed with my case nor disagreed with the assumptions. There is of course a lot of potential nuance to dig into after that. The case, in other words, is that the hammock takes up a lot of time. Ten minutes of thinking is not hammock time. I want to remember that Rich was talking on the scale of days, weeks and months. If I am spending a month thinking about how to design and implement a feature, what does the agent have to offer me that is of value? Speeding up the typing by a few minutes? I know the risk of having the agent available: it turns my hammock sessions into fruitless brainstorming chat sessions. I am much less familiar with the benefits.

> what does the agent have to offer me that is of value? Purely utilitarian things, nothing hammock time-related by itself. But such things could still be of use during that time. Consider it a shitty bloom filter that can have two outputs: "this might be relevant info" and "no relevant info was found". Neither means anything definitive. The latter is arguably useless. But the former has some use, if approached carefully. E.g. let's assume that hammock time allows for prototypes. I have to write it out explicitly because I've seen opinions that no coding whatsoever is allowed during that time, regardless of the nature of that coding. And this is just hypothetical. Similar only partially to what I have done myself - more on that below. So you make a bare-bones prototype for an idea. Maybe with an LLM, maybe without - it's a fraction of the time anyway, not important. And then you can let an LLM go ham on it to try and find issues in it. If you have enough tokens, you might let it churn for days - fully by itself. Nowadays proper reliable loops are feasible, but you can still take a peek once an hour/day/whatever just to make sure. If after that time there are no show-stopping problems with that prototype, the light is green, you can proceed as if there was no LLM involved at all. But if there were some issues, it might be worth it to inspect them and see what's up. There's a chance there that something incredibly bad is going on with your prototype. Something that completely eliminates any chance of that idea to work, or maybe something that demands rethinking from the ground up. Discovering such a show-stopper yourself is also possible. But insanely cost-ineffective. As promised - a particular example. A client had an idea of a mobile-friendly audio-based web app. Other details are beside the point. Creating a prototype went swimmingly, but with a small hitch. Were it not for that hitch, we would proceed with the idea and later we'd encounter something akin to a brick wall. That hitch was pure luck - there was no audio at all on the client's phone. I tried it on 3 phones and a tablet I have access to - all worked fine. Every desktop browser I could get my hands on worked fine. Finally started testing with a service that allows me remote access to physical devices - around 20% of iPhones had no audio. Even same model and OS/browser version. Absolutely no clue why, couldn't figure it out after spending days on it, decided it's just not worth it. That's something an LLM could discover by itself. Web audio is notoriously inconsistent, but I had no idea to which degree.

It doesn't sound like you're talking about generating production code with LLMs here, though. As a shitty bloom filter for reviews, I have no issues with it. Or have I misunderstood what you're trying to say?

Yeah, you got it right.

Perhaps I can state my position from a different angle. I feel like we are in a situation where the 3 million LOC of spaghetti is a pandemic. I think agentic coding does very little to fight this pandemic while the hammock and other teachings of Rich do a lot. I even dare suggest that agentic coding is aiding this pandemic, partially by distracting us from the hammock with promises of efficacy that are an illusion more than anything.

I can't call myself an "AI-maximizing Clojurist" by any means, so I'm not the target of your original question, but FWIW I agree with the message you just posted. I was only addressing the last question in the OP. Not on someone's behalf of course, just outlining a possibility.

You clearly spoke on their behalf. I can hence declare victory!

Well, shite. I'm sorry my dear resident AI-maximizers, I have failed you! Even though I never intended nor cared to.

With Clojure you can directly solve problems with very little nonsense around your solution. If you want to tell an AI in detail what you want, you need a formalism like Clojure. But Clojure is already s programm language you dont have to go through an LLM with it. Just a thought.

If I am spending a month thinking about how to design and implement a feature, what does the agent have to offer me that is of value? Speeding up the typing by a few minutes?

And that's why I think I observe more laziness and less hammock time at my current workplace. Because if you want to gain velocity, you need to shorten everything that you were waiting on, that means hammock time needs to shorten as well. Time designing, time reviewing, time testing, time implementing, time discussing with others, it all has to shorten. So we use agents to produce designs and criticize them, to review the PRs, to test the code, and to implement, etc. And we do less and less scrutiny and review of all that, which produces higher velocity, and we don't know the long term cost of this yet. Will it hit a wall, come crashing down? Or actually agents can do all that to scale and in a way that gives you long term runway?

Working in large existing systems I find a big portion of the time spent is on implementation. The available design space is not that wide. The goal is not that technically ambitious, and mostly is for a product deliverable. Figuring out exactly how to interface with the existing system structure occupies most of the time. With ai, this time is significantly shortened in my experience. That frees me up to think more ambitiously about improving the technical structure of the system as a side quest to my product work.

In personal projects I'm not encumbered by existing systems but I still find the implementation takes up a significant amount of time, and I'm able to go more quickly using ai. I also feel less personal attachment to the code so I feel more willing to explore alternative structures and throw away code if it no longer fits the vision. If I were hand coding everything I think I would be more rigid in my thinking.

I have the same experience regarding large existing systems. But the available design space was not narrow for inherent reasons, but for lack of vision. Again, the 3 million LOC system exists, and the attitude is to keep working in it, rather than fixing it. Insofar as AI makes it less dreadful to work in such a system, it incentivises its continued existence. I believe that implementation time taking up too much time in a large system is a red flag. It sounds like a result of too little hammock time, too little skepticism regarding existing architecture or assumptions. Most of the code in a large system should not be glue. If your core business logic is slow to implement after the design is clear, is that not a sign of accidental complexity having found its way into the code base? I don't say this because I know I am right. I am far from experienced enough to have such a high degree of certainty. This is my interpretation of what FP and Clojure specifically is all about, which I am sold on. Rich Hickey's Death By Specificity comes to mind: https://www.youtube.com/watch?v=aSEQfqNYNAc

Whether the implementation is smooth because of good existing design, or strenuous because of bad existing design, I find ai is able to complete the implementation faster than me in both cases. That does not solve all aspects of the sdlc but it improves one of them.

Even in the best designed system I've worked in, the implementation took enough by hand time that product managers would get bored waiting for the implementation. Until I can sit down with a product owner and get the implementation done in real time without them getting bored or impatient, I'll take all the speed up I can get. I don't really resonate with the idea that the implementation is fast enough so we don't need to improve it further.

I find there's a self adjusting aspect to implementation velocity and effort. Once the current problems have been abstracted away such that incremental progress is easy and fast, the natural next step is to run into problems at a higher level of abstraction or broader aspect of the system that haven't been solved yet. Then those problems occupy a lot of implementation time and effort. Greg Lemond, the Tour de France champion, once said "It never gets easier; you just go faster". I find this is also true in my work. Once small problems become easy I am freed up to work on large problems, and the work is hard again.

> Once the current problems have been abstracted away such that incremental progress is easy and fast, the natural next step is to run into problems at a higher level of abstraction or broader aspect of the system that haven't been solved yet. Sounds like you are saying that once you are done with the work in the hammock, the next natural step is to spend time in the hammock again for the next challenge. That's exactly it. Agreed, the work is always hard. Hard work is tackled on the hammock so that the implementation can be fast and easy. It sounds like we are in agreement; the thing I have been arguing against is the premise of the Agentic Coders Syndicate that implementation work is slow and boring. I posit that the "slow and boring" attitude is one that results in 3 million LOC of spaghetti, and that the alternative would spend so much time in the hammock that the final actual implementation of business logic becomes fast, declarative, and fun. Give me six hours to chop down a tree, and I will spend the first four sharpening the axe.

In my experience, lots of problems either don't have a clear "hammock" style solution, or that solution pays off on a long enough time horizon so as to not be viable for the business. I do enjoy when an elegant solution is available and I've built some systems like that but I've also been in situations where the elegant solution is on the bottom of a pile of rocks that would need to be moved first before the elegant solution is accessible. In the mean time work must proceed on the inelegant system. I think this point is basically irrespective of ai use, btw.

I feel a lot has to be really dependent on what/where you work on. I've found on big existing systems, AI doesn't make me that much faster, because changes are already relatively small, and it's more about being super careful you don't break a bunch of existing things, that you handle all the weirdness of legacy data, and so on. I say "that much faster", as I think it's still a bit faster. Whereas for greenfield or smaller apps, AI is way faster at implementation. In that sense, from my experience, I'd agree that in either case AI does still provide some speedups. What hammock time and simplicity provides is making that former case take less time as there will be less legacy data weirdness, less surprises, less risk of your change breaking things far away from it, etc. But even so, with AI it would still go faster. That said, there are risks with AI as well, the risk that it cannot maintain that simplicity of design, that it cannot deliver on what was hammocked (it claims it does but the code doesn't reflect it). And it's where I don't yet have the experience to say the effect long term.