My heart goes to the stack overflow community which has always been very kind and helpful, essentially working for free.
As a self-taught developer since the age of 8, I literally grew up learning how to code through SO, asking hundreds of questions and answering many more.
So many bugs that would take 2-3 days to fix would eventually find their answer through it.
But now ChatGPT does that in minutes… so it’s for the best!
The presumption is that things will improve over time, but the big difference in my experience is the assistance I got from SO _worked_ the vast majority of the time, whereas various LLMs I have used generate unusable, misleading, or unreliable results pretty regularly, increasing as complexity or rarity arises. As human-driven knowledge bases backed by actual experience are replaced by inference from models who rely on such inputs, I am concerned about the medium to long term impact. A lot of people grew frustrated with SO for various reasons and went back to unhelpful behaviors that SO resolved at its zenith (rather than dead ends and flame wars in newslists and irc channels, they do it in random subreddits and discord servers instead). Now what if we circle back after degenerative LLM experiences only to find there’s nothing to circle back on?
I think this is the first time I’ve ever heard someone describe the stack overflow community as “kind”. Usually it’s the exact opposite: “I asked a question and just got 30 questions asking why I’m trying to do this” or “my question was closed for seemingly no reason”.
It’s literally the most blunt and aggressive website I’ve ever been on that wasn’t a straight-up troll site like 4-Chan.
I answer questions for a few tags, know the styles of the other people who answer on those tags, and I do consider my fellow answerers as kind. All of them.
That doesn't mean that you will think we're kind to you personally. We're there to build a searchable Q&A knowledge base and spread knowledge. Some people who ask questions misunderstand and think we're there to help them, personally. To work for free for that single person, and we're not there for that. We write answers for the tens, hundreds, thousands of people who will search for it.
Askers who misunderstand will come across as overly entitled.
In terms of practical effects: People who misunderstand don't tag their questions, or tag them incorrectly. They post screenshots full of text. They don't look for similar older questions in the existing knowledge base, or they insist that even slightly different questions are significantly different. All rather offputting, and often puzzling. How can you ask for a subject expert's help and simultaneously insist that you know better than the expert whether your question is a duplicate of another?
There was a time when it was really good. Like legitimately good and useful. But over time it ended up becoming exactly what you describe. But there are still countless examples of the usefulness of SO in Google results. I stopped asking questions in 2012 and stopped answering questions in 2015. Before that though, it was a very useful tool.
Agree! A decade ago it wasn't like this. But it has devolved into a community of vandals who seem to take glee in criticizing the manner in which a question is asked, rather than contributing a solution.
Yep, believe it's a direct result of Atwood's iron-fisted no-bullshit policy. To some extent it is great... don't want it turning into Yahoo Answers, do we? I think folks forget about that part.
But, as you mention they just went too damn far with the medicine.
No, you can't fix this misspelling, isn't there something else (with more characters) that you can improve as well? WTF, for realz? :-/
Mixed feelings on SO. It was helpful, but it was also a website you dread having to post on because it was filled with the most intolerable people of the internet who you just had to endure abuse from if you wanted help.
Now chatGPT gives you the same help without the abuse.
The next AI totally needs to be more snarky to make it feel more like we're dealing with actual "thinksperts", people that think they are experts even if their answers are demonstrably wrong.
Why does any LLM need new information to do fundamentally the same thing?
And what makes the data outdated? New code? It can train on that. That, or there is simply nothing new to learn, just new ways to express the same thing.
I've asked dozens of questions on SO, and never had a single one closed. I hear your sentiment often, but have no idea whether my experience or yours is more common.
I've had 3 deleted by Community bot as abandoned, but since they were over a year old when that happened, I couldn't care less.
I've got valuable advice from SO over the years. There's overlap with LLMs, sure, but it's frequent to have questions that have no answers published anywhere on the web; SO brings people who know out of the woodwork, who create an explanation that didn't exist before. A couple days ago, someone in retrocomputing got to bank-switch a 1983 Radio Shack box... that kind of stuff wasn't published anywhere, until a guy who used to write games for that box answered that question on SO.
These models can figure out syntax and language features they haven’t seen before. Try it with a few code snippets of your own made-up language. It’s a little freaky.
The point of an LLM is that it can take your problem as input, along with answers to previously asked questions (perhaps implicit) in its training data, and attempt to synthesize a solution to a problem as output. Here, a "question" is something that can be found with a search engine - something that directly presents the/a crux of an issue, which is identified after a debugging session (for a problem in existing code; for projects still in the design phase, a how-to question emerges after coming up with user stories and breaking tasks down into their logical steps). The point is that the question can be relevant to many different people who have written different code which encounters different problems - all caused fundamentally by the same conceptual misunderstanding (or non-understanding).
Stack Overflow is explicitly not designed or intended to solve problems or do the decomposition of the problem for you, nor the synthesis of answers. Because the result would never be useful to anyone else. The entire point is to have something searchable, and to allow answer-writers to keep their explanations DRY.
This has spectacularly failed, because no matter how frustrated people get with traditional discussion forums (https://www.explainxkcd.com/wiki/index.php/979:_Wisdom_of_th... among many other typical complaints), they apparently are much more suited to human nature.
Heaven knows how Wikipedia managed to avoid devolving into "Quora but even worse because you can scribble over someone else's post".
I just hope that we can continue to find sources of high quality training data like SO. If people don't publish their mutual learnings somewhere then there's no data to train on.
It was downvoted because the entire purported "threat" is based on a misconception.
It is not relevant to SO whether an LLM can provide personalized help, write with any particular tone, answer promptly, accept every input prompt as valid and try to make sense of it, discuss back and forth to figure out a problem, etc. Because Stack Exchange is explicitly and by design not for those things.
But also, downvotes work differently on meta anyway, and the community there generally takes a negative view of LLMs. Because, again, the point of SO is for the answer to come from a human expert, and be verified against subject matter expertise rather than simply being evaluated for coherence or generally sounding appropriate in context.
> 2014: questions started to decline, which was also when Stack Overflow significantly improved moderator efficiency. From then, questions were closed faster, many more were closed, and “low quality” questions were removed more efficiently. This tallies with my memory of feeling that site moderators had gone on a power trip by closing legitimate questions. I stopped asking questions around this time because the site felt unwelcome.
I also felt around that time that it became unwelcoming. I didn’t realize they had revamped the moderator tools. That is the time period when I stopped using it too. Now I know why.
How many other websites have also shot themselves in the foot by tweaking things?
Same. I've never been a huge StackOverflow user, but it is so irritating to search and find your exact question on StackOverflow, often as the top result, only to see that it was instantly shut down two years ago as a duplicate of some other question in another context with inapplicable and useless answers.
It is frustrating not only because you can't get instant help, but also because it shows the futility of even trying to post on there.
people who own walled gardens often get the idea in their head to prune up the trees a little, make them elegant looking and pretty, that's moderation, deletion, banning,etc. they get a good feeling pruning the tree, making something beautiful, moment of joy, thinking that such a beauticious little garden they've made will make it all the more appealing to visitors and potential garden supporters.
some have a tendency to go overboard with this thinking, only to discover that a heavily pruned tree is now a dead tree, now finding themselves in dead tree garden.
For a time the "let's interact with people and talk about cool things" group and the "let's build the ultimate knowledge base" group had their incentives aligned.
Then, with better moderator tools, the "ultimate knowledge base" group set out to achieve the ultimate knowledge base by reducing the amount of people who were just there to talk.
> Then, with better moderator tools, the "ultimate knowledge base" group set out to achieve the ultimate knowledge base by reducing the amount of people who were just there to talk.
Yes, because the people who were just there to talk had reached a point where they could effectively only pollute the knowledge base. Bad questions make good ones harder to find, simply by existing (since the bad one could potentially be found instead, and because of the broken window effect).
I didn't start learning to code until around 2014, so I didn't know SO before this change. But when you make your platform so militantly anti-helpful, militantly anti-newbie, and just all around unpleasant to interact with, then it shouldn't be a huge surprise to anyone that people stop using it. There's over half a human generation of learners who have had almost entirely negative experiences with SO.
How ironic. "AI" feeds off structured knowledge, artistic creations and otherwise any human production to generate its output. As a consequence of its widespread adoption, people start to lean even more towards consuming rather than producing, a tendency which was already increasing before the advent of LLMs and modern machine-learning. This, in turn, leaves "AI" implementations with no new human content to feed off of. Now what? The whole process folds onto itself. Are we entering the dark ages of cultural (in the widest sense of the word) production? Not that I don't think that we're already there, in any case, but for other, somewhat related causes...
Perhaps the next step is having the LLMs ask questions on SO when they routinely fumble particular topics. I could see a system of knowledge bounties where people are compensated for providing accurate, in-depth training data on niche topics.
LLM content is banned everywhere on Stack Overflow, in both questions and answers, by policy, since mere days after the public announcement of ChatGPT (because it was immediately causing a huge problem): https://meta.stackoverflow.com/questions/421831
Moderators (actual elected moderators, the two dozen or so that exist for ~29 million user accounts and ~24 million non-deleted questions) went on strike in mid 2023, largely because the site staff/owners interfered with their ability to remove such content (an overwhelmingly popular policy with strong community consensus): https://meta.stackoverflow.com/questions/425000 and this decision propagated across the Stack Exchange network (as most SE sites had adopted similar policies): https://meta.stackexchange.com/questions/389811/
A large fraction of the userbase is explicitly opposed to helping LLMs out in any way whatsoever. I personally have ceased contributing new question or answer content, and only edit existing posts. I contribute new content on Codidact (https://software.codidact.com/) instead (disclosure: I have recently become a moderator there).
> This, in turn, leaves "AI" implementations with no new human content to feed off of. Now what?
You seem to be under the impression that AI needs more than all recorded human knowledge up until 2024 to reach the same level as an average SO contributor.
It doesn't. Because none of the average SO contributors did.
It is unclear what algorithmic improvements are required to leverage the available data to get AI to AGI, but a lack of data is definitely not the bottleneck.
One could say that these AI systems aren't sharing their solutions (or questions) with other AI systems and that the world would benefit from it if they did, though. Perhaps it's a good idea to have some shared space for AI systems where they share the validated solutions they synthesized.
> You seem to be under the impression that AI needs more than all recorded human knowledge up until 2024 to reach the same level as an average SO contributor.
Replacing the average SO cobtributor isn't adequate to replace SO, and AI is able to “replace” SO effectively only since major models have gotten not only SO-as-training-data but web search (including SO) for immediate grounding.
And without SO or something like it with active human contributions it’ll have even more trouble replacing the value SO would provide for new questions and new domains where it will neither have SO traijing data nor SO query-time-search-results to use to synthesize answers.
I find it interesting that the current StackOverflow moderators tend to say "in the past we used to accept too many questions but it was never the goal, so now we are doing it as it was meant to be".
Sure, but in the past, StackOverflow was growing, and now it's dying. Maybe something was better before, when "it was not done correctly"?
I think these sorts of things are just an unfortunate side effect of scaling. The bigger you get the more people get lost in the bureaucracy. However if you don't build up the bureaucracy the system collapses under its own popularity.
Wikipedia has a similar issue where editing declined around 2007, which is often blamed on stricter enforcement of rules, more complex rules, etc. I think its just a natural stage of growth. You can't be a free for all forever.
The "good" thing is, they're back to 2009 levels of postings. Now obviously that's what the mods let through but my guess is that traffic to the site is down precipitously as well. They can roll back their bureaucracy and head back to a lean path that worked for them in the past.
But I don't really think that's the problem. Reading zahlman's responses in this thread makes me think that the mods fell into the age old trap that's happened since Usenet, IRC, and still happens to this day wherever there's mods: they got tired of doing unpaid labor and instead of deciding to quit decided to become meaner and stricter. The age old mod trip.
Two thirds of the wikipedia article i wrote in 2003 have been deleted by rabid editors. It was a biography of my father, written based on interviews with my mother. I have found that restoring any of the rabid editor deletions results in threats of me being banned from editing my own article.
> Sure, but in the past, StackOverflow was growing, and now it's dying. Maybe something was better before, when "it was not done correctly"?
You're presuming that the current volume of questions represent novel, unique posts instead of something you can find over and over again if you do a decent query.
AFAICR they've always said these lines about now is about better moderation from the slop. The reality is that the rule of thumb for that moderation was already out of date with advances that preceeded LLMs.. Even with the beginnings of computer aided flows we didn't need to alienate most to get the best content and develop the few. Content can be triaged from someone who may be human to others who may be human and maybe there's value or maybe you just didn't alienate anyone and some people will still climb to making higher levels of content that is worth condensing.
> Even with the beginnings of computer aided flows we didn't need to alienate most to get the best content and develop the few.
The large majority of new questions from new accounts are from people who are clearly there only to solve a personal problem, who show no interest in considering the value of their question to third parties, and rarely put any effort into attempting to even diagnose or specify a problem.
Even after it became possible for most of these people to get an instant answer from an LLM. Which is actively preferable from the standpoint of Stack Overflow curators. Before LLMs, the point was for them to use a search engine to find an existing question that lets them figure out the problem. But for the Q&A to help such users, they need to apply at least basic problem-solving and debugging skills. (It is explicitly out of scope for the Stack Overflow community to do that for others; and attempting to do this in an answer actively degrades the site for everyone else.) If an LLM can fill in some hypotheses for those users to test, then the LLM is doing what it's best at, and Stack Overflow is doing what it's best at.
Stack Overflow is not there to troubleshoot or debug anything for you, nor to reason about a multi-step problem and break it down into its natural logical steps. It's there to give a direct, objective answer to how to do each individual step, and to explain why the specific point of failure in a failing program fails, after you have identified it and made the problem reproducible.
So yes, we absolutely do need to "alienate most", because "most" are there for a reason that has absolutely nothing to do with getting the best content.
> Overwhelmingly, the people you're talking about are not moderators.
I was actually thinking about you. You keep saying everything is great. My observation is that I used to be on SO every day, and I completely stopped contributing even though I would have plenty of stuff to add (more than ever, actually).
> Why should a reduction in incoming questions mean that it's "dying"?
There is "a reduction", and there is "being back to the amount of questions SO had in 2009 when it launched".
The new owners have been trying very hard with the "if you can't beat 'em, join 'em" approach. They know this is radically against community consensus (it's been shown to them on the meta site over and over) - so they just get sneakier about it.
Doesn't feel like the AI is the main driver. Many things changed over time - dev tools got better, editors got smarter, compilers got better error messages, various primary resources improved, tutorial websites, courses and youtube boomed.
Another point of course is that each new question is more and more likely to be already answered. At some point the site pretty much covers most of what is to be answered.
One aspect I haven’t seen anyone mention contributing to the decline is GitHub (part of your “improved tooling”)
These days you can go to the repo and there’s usually already an issue open with the problem and a workaround. Or if someone has a question on how to use the tool/software they ask there.
This, and first party developer forums. iOS questions will go directly to Apple's community forums. Same for SalesForce, or Elastic search etc.
There's just a higher noise/signal ratio, a real chance to get answers from experts, and it makes for a stepping stone if the issues needs to be bumped to paid support.
> each new question is more and more likely to be already answered
Yeah, except for when there should be current answers. Most of computing is in constant flux. There's a mountain of 10+ year old answers that simply don't apply any more.
> except for when there should be current answers.
Yes.
They belong on the existing question - unless the existing question is poorly asked and the new one is better asked.
New answers can, by default, be added at any time - and should, if existing answers are actually out of date rather than simply being old. (Many 10+ year old answers really do still apply.)
Asking the question again is not how the site is designed or intended to work.
I answered one question 13 years ago where I still gets points for. Computing as a whole isn't so much in constant flux, it is only JavaScript that changes so much. Hell, I learned to use Apache 30 years ago, and I didn't need to learn anything new the last 25 years of that.
There was a noticeable inflection in the question-rate-vs-time curve around the time that ChatGPT was released.
Which is fine. If your question is not answered by `site:stackoverflow.com how to do the thing` but it is answered by an LLM taking `how do I do the thing?` as a prompt and synthezising existing Stack Overflow content, that is inherently not a suitable Stack Overflow question. Because anyone else could put `how do I do the thing?` into the same LLM. It's not any different from using a traditional search engine.
(And when the LLM fails by producing a wrong synthesis, then blessing that result by putting it on Stack Overflow is actively harmful - which is why it's banned by policy.)
> Another point of course is that each new question is more and more likely to be already answered. At some point the site pretty much covers most of what is to be answered.
AI has the ability combine answers from multiple sources and tailor-make to your exact prompt details. Now that is something we call glove fits the hand. Plus it can explain its answers.
Yes, it can combine multiple sources and make up an answer that makes no sense. Even if it explains how "it works", it does not help when the API or a function has never existed.
Woof. Looking at a single metric and extrapolating "LLMs killed the radio star"
Stack Exchange sites are designed to nuke duplicates, help people before they post a new question. It seems a natural conclusion that the number of original questions decreases over time.
I won't pretend that some people live their lives inside and LLM but many of us still use search engines and SO.
It is a bit sad. And obviously the reason why it sees such a decline is because AI (ChatGPT and similar) took the job of answering the basic questions about programming that StackOverflow used to help with.
Looking at my profile since 14 years ago, the most upvoted answer that I solved was about a basic question of how to specify fields properly when you serialize JSON into a C# class.
I do believe the value of StackOverflow was only about people who were lazy enough to read the documentation of the language/framework they were trying to use. I used to be active on StackOverflow back in the days, but in the last 10 years the only value I saw in it was if I needed to get back to some language to just find an answer on how to write a for loop in that specific language (swift vs go vs ...).
I personally do not believe there is much knowledge base on StackOverflow. In most of my questions to "google" for the last 10 years, very rarely would I be directed to StackOverflow for the right answer.
There are a lot of complicated questions on StackOverflow, but the site was flooded by people asking and answering basic questions about programming. And people who are there just to get some karma.
>>And obviously the reason why it sees such a decline is because AI (ChatGPT and similar) took the job of answering the basic questions about programming that StackOverflow used to help with.
A big reason why AI is replacing these things. A big part is the experience itself. There are quite a few people who have have been repulsed due to the smugness, or other wise having their questions marked duplicate/irrelevant/stupid etc.
AI is also pretty much instant. You can also talk to it like you are talking to a person.
The killer AI feature!---> AI listens, without judging you.
I used to answer questions a lot, by around 2013 I had answered maybe ~12% of all HTML canvas questions ever asked. To me it declined a lot sooner, 2014 really does feel like the right inflection point.
There was a belief, sometimes unstated but often explicit, that no more (serious) discussion is really to be had, and further wondering how can one stop people from asking. It became difficult to discuss anything if there was even something vaguely related asked before. It was not possible to discuss something you knew the answer to, but did not know why, or wanted to hear arguments for which of 5 ways might be best. All (to me) very worthwhile technical discussions. Totally shut down.
> There was a belief, sometimes unstated but often explicit, that no more (serious) discussion is really to be had, and further wondering how can one stop people from asking. It became difficult to discuss anything if there was even something vaguely related asked before.
There is nothing to do with unstated belief here.
It is explicit policy that we don't have discussion at all.
We have answers to questions.
Which is why there's a question (explicitly labelled as such, and not just "help me") at the top of the page, and every post below it is labelled as an answer (and is explicitly not a response to anything else but is simply there to answer the question).
StackOverflow should have focused on linking duplicates rather than forbidding duplicates.
No Boilerplate recently said "writing is thinking"[0], and suggested links are the ultimate knowledge graph organizational tool--not tags, not folders--links[1].
StackOverflow tried to prevent all duplicate questions. This was stifling and reduced writing, reduced thought, and most importantly, reduced user engagement.
The people who wanted to write their problems and ask their questions stopped going to StackOverflow. The people who wanted to write and give answers stopped going to StackOverflow.
Look at Discord or IRC and you'll see that people have their own questions to ask, and the people who answer such questions enjoy answering the same questions over and over. Let the people write their questions, and write their answers and give advice. Instead of preventing duplicates, link duplicate questions together.
> StackOverflow should have focused on linking duplicates rather than forbidding duplicates.... StackOverflow tried to prevent all duplicate questions.
Duplicates are not forbidden on Stack Overflow. Duplicate questions are linked together - that's what "duplicate" means in the system. Beyond that, logged-out users who land on a duplicate question that doesn't have its own answers, will be automatically redirected to the target.
When duplicate question are linked on Stack Overflow (and everywhere else on Stack Exchange), they are automatically closed, which merely prevents new answers. The purpose is to allow high-quality answers to be gathered in one place - on the duplicate target, which in turn is ideally a high-quality version of the question (and a focus for curators to improve further, when they notice that it becomes a common duplicate target).
Deletion of duplicates (and posts in general) is not very well understood and people are not all on the same page - see e.g. https://meta.stackoverflow.com/questions/426214 . But normally duplicates should not be deleted unless they are actively harming the search results (i.e.: it causes people to find the wrong target, because it was written in a way that fools keyword search).
Nobody gets formally sanctioned for posting a duplicate. You just get your question closed, and maybe asked to search a bit more carefully in the future (or shown how to do so).
> The people who wanted to write and give answers stopped going to StackOverflow. Look at Discord or IRC and you'll see that people have their own questions to ask, and the people who answer such questions enjoy answering the same questions over and over.
People who want to give the same easy answers over and over to the same easy questions, so as to get imaginary internet points that move them up a leaderboard, should not feel welcome on Stack Overflow. That action is actively counterproductive to what Stack Overflow is trying to accomplish. There are countless discussion forums (and as you say, Discord and IRC channels) already where that behaviour is valued. The Internet should be allowed to have one place where it is not valued - especially when it's a place that was specifically created to accommodate people who want search engines to be useful; who want to write high quality answers once and get many people to read them; etc.
Maybe "answered by duplicate" would have been a more friendly way to say it--but I hear you, yeah, the closed questions were linked, which is what I was asking for. Whatever the case, a lot of people decided to never use the site because their questions were closed.
> People who want to give the same easy answers over and over to the same easy questions, so as to get imaginary internet points that move them up a leaderboard, should not feel welcome on Stack Overflow.
I think it's worth asking, why do people give answers on the internet at all? Maybe it's because of internet points, but more often people just like interacting with other humans and teaching.
In the beginning, StackOverflow was a place for people who wanted to interact with other humans, and also a place for people who wanted to build the ultimate knowledge base--for a time their incentives were aligned.
But then over time the space for interacting with other humans got smaller and smaller, and now StackOverflow is almost entirely about maintaining the knowledge base that has been built.
And yeah, like you say, it's okay if StackOverflow isn't the place for human interaction. StackOverflow has built its knowledge base, and some still maintain it, and the long term success of that knowledge base is becoming ever more apparent--which is to say, not very successful--the day may soon come that StackOverflow isn't even hosted anymore.
(Also, I want to ward off the claims that this is because of AI. StackOverflow was in steady decline long before AI was competent at answering questions; even in 2025 the competence of AI is still in question: https://news.ycombinator.com/item?id=44000118)
I think "number of questions asked" is the wrong metric. Because it feels like all the questions have already been asked. Whenever I need to know something, I can google it and find answers on Stack Overflow. I can't remember the last time I actually had to ask something. Or the last time I found a question that didn't already have a good answer. Stack Overflow's library of question is pretty complete, and the only reason for new questions are new tools.
Certainly LLMs are a huge factor, but I feel that LLMs rarely give good (and trustworthy!) answers to the things I would check on Stackoverflow. Just like LLMs are no good replacement for API references because they get the details wrong all the time.
I think this is true if there aren't new questions to be asked. But technologies shift and evolve all of the time.
One of my top StackOverflow questions for years was around the viability of ECMAscript 6. It's now essentially irrelevant because it's found wide adoption in browsers etc. but at the time a lot of people appreciated the question because they wanted to adopt the technology but weren't sure what its maturity was.
It's also true that some technology stacks mature to a point where there isn't much more to be asked but I think there will continue to be a place for forums of discussion where you can ask and get answers around newer, bleeding edge technologies, use cases etc.
I find that most of the time, when doing research on anything non-trivial, I find a question on SO about this exact problem that has no answers because it was closed by the mods as a duplicate of something that doesn't actually answer that question (but rather something very vaguely related to it).
This was my question. There's a weird sort of self-cannibalism that this hints at. The LLM is only as good as it is because it's been able to train on existing SO answers. But if over time, SO content production declines, then the LLM results will be less reliable. It seems that a new equilibrium could be one in which -- for newer questions/concerns -- both SO and LLMs will be worse than they are now.
To add a bit more nuance, SO has a question-answer type format, which leads very well into prompt-rely format to train these chat applications. Most of the other sources do not, except for Github issues maybe. Without this question-answer format, there'll be a need for a bigger data labeling effort to train LLMs on new stuff, no?
My approved to shut down question rate was about 40/60. And within the approved questions, I only got the answers I looked for like 50-60% of the time. Got to 350 points myself. I hated every second of it.
Really aggressive moderation, people trying to score points for a worthless achievement system by spamming comments like "You should narrow the scope of this question"
Having to grind achievements to be able to comment, like or dislike.
I used it for a year or so back in 2013?, went back to posting in forums like XDA developers, Codeguru and Reddit.
Just how is 25K new questions a month dead? Even if it gradually asymptomates to just 1K, the answers to them are enough to continue serving as a critical base layer of high-quality training data for LLMs.
Let's say for the sake of argument that 95% of humanity perished. Is humanity then dead? It isn't.
It just won't survive with 1k questions per month from the financial point of view. The hosting, the team, etc. Also, there is GitHub, where you can get high-quality answers directly from the authors of the component or library you are using and high-quality training data.
> June 2021: Stack Overflow sold for $1.8B to private equity investor, Prosus. In hindsight, the founders – Jeff Atwood and Joel Spolsky – sold with near-perfect timing, before terminal decline.
Hey, if it all crashes and burns, at least it’s the so-called smartest guys in the room going down with the ship. Just a bunch of VCs learning the hard way that they had no idea how to actually run or grow the company they bought. “Look at how well we optimized it!” Yeah — right into the ground.
Questions asked isn't the metric to declare it dead. What matters for financial viability is traffic and deals with search engines/indexes (used by LLMs)
For community viability: people will keep using it where LLMs fail. For new problems. It's still the place to go for undocumented workarounds.
Traffic and voting activity is certainly down but there is still immense value and new valuable questions are asked and answered there.
The article talks about the number of questions asked in SO, but there are no mentions of visualizations.
Feels natural that after 16 years of refinements, most normal questions are already there. I use it every week, but can count on one hand the number of questions I asked (0 through my account) over 12 years of having an account. ~All my questions were already asked.
I always considered stack overflow to be a Band-Aid placed on top of a mistake! The mistake was always poor documentation by the original system designes, and a policy of not allowing edits to the system design doc! And the more mistakes a documentation person makes, the larger the stack overflow corpus!
Now ChatGPT for SO is a Band-Aid on top of a Band-Aid on top of a mistake!
I really don't believe in the elitist policy to qualify for being able to answer stack overflow questions ... Whenever I have a better answer than all the existing ones stack overflow says I'm not qualified to answer so shut up! To hell with SO - I answer more questions at my company than anybody else and SO is run by elitst fools ...
This will have interesting implications for the LLMs as well, since SO is a wealth of training data. In my experience, LLMs are pretty useless when it comes to helping me with newer, faster-evolving, experimental tools and libraries, which is not surprising. But, if the SO community really atrophies to the point that a lot fewer people are bothering to answer questions, there won't be another centralized resource for answers. Perhaps that just means balkanized communities like random Slack channels will fill the gap, but those aren't search-indexed and I'd bet getting them all into training corpi won't be as easy either.
Maybe the future involves LLMs asking questions on something like SO when it routinely fumbles a particular topic. People could get paid to answer them and provide more training data. Who knows at this point
A sociological case study. Legit founders, a fruitful niche, immense value. Growth, politics, corporatization. They did so many things right, then so many things wrong.
If it were up to me, moderation would have been overhauled. But it wasn't up to me.
I have some minimally popular answers on SO and for years my "points" graph has notched its way upwards. It flatlined at the start of 2024 as people moved to AI.
The best thing about SO is seeing the competing solutions, the discussions, meaning with some discernment you can find that peer-reviewed high quality code snippet. Why would people prefer whatever the AI spits out?
Fortunately I see a few blips on SO so hopefully people are coming back now that the shine has worn off AI.
What is the value of SO to the world economy? Billions. Like the internet archive, it should be some sort of government funded (UN?) library
I'll be harsh, but they deserve it. The value of SO plummeted as documentation got better, disallowing open ended questions and discussions completely removed the value of talking to other humans.
It was useful as junior and intermediate, but the value for a senior dropped to 0 due to the inability to ask what I would consider important.
I had to resort to reddit to ask those questions, which is ironic given the focus of SO.
Personally my usage of S.O. was significantly reduced just by sticking to the same stack and tools for the last 7 or 8 years and letting depth accumulate instead of always being in the midst of learning a new framework / language / whatever that necessitated googling how to do x in y or z.
And while that was happening VS Code started integrating MDN as well, so when I come across something I don't recognize I have a lot of extra information right at my fingertips anyway.
My day job includes moderation. I think I am more empathetic to the issues that moderators deal with on a day to day basis given that they are underpaid, under-appreciated, and overworked.
It's very difficult to scale a community to be both welcoming and productive. New users don't have the same context as existing ones. You find that norms and manners aren't transferred from one group to the next. So although that I noticed that SO started getting more strict from 2014 onward, I wouldn't know immediately what to do about the content quality issue.
My take is that, like most things, the medium of the old will be appreciated the way it wasn't in 2014. As the Brian Eno quote goes: "Whatever you now find weird, ugly, uncomfortable and nasty about a new medium will surely become its signature." People will yearn for the human forums the same way they did years past when people tire of the LLM slop. (If they do.)
I'm not very sympathetic to moderators. Their job is to make some forum contain only what they want to hear from the (unpaid) people providing the content. If that's what you want, start a blog.
In any sort of IRL community, if some person decided they had the authority to prevent people from answering others' questions, unilaterally decide people aren't welcome, etc. everyone would think they had gone insane.
However there is the problem where any unmoderated public forum will eventually contain only scam artists and Nazis. Moderators address the Nazi spam problem, but now you have the moderator problem. It's been a mental side project of mine lately to find ways to solve the moderator problem, which I can't talk too much about as there is money to be made by doing so.
What happens when the forum gets overrun by Nazi spam?
We'll install a set of moderators who remove it.
But what happens when the moderators start removing useful information?
We simply unleash wave after wave of Chinese needle snakes. They'll wipe out the moderators.
But aren't the snakes even worse?
Yes, but we're prepared for that. We've lined up a fabulous type of gorilla that thrives on snake meat.
Then we're stuck with gorillas!
No, that's the beautiful part. When wintertime rolls around, the gorillas simply freeze to death.
Yes this is the really hard problem of human organization on the internet. You need mods to keep the quality gate up. But eventually mods feed off their own power and alienate the community. Maybe some combination of human and algorithmic moderation can help.
I don't miss spending hours trying to find the right config to make things work and trying random answers on SO. Overall the dev world is a better place now and MANY dev hours are being saved. Plus, AI can now learn any new framework/language directly from docs. Even obscure ones, just pass all the .md files to gemini and ask!
It has helped me in the past but yet, I could not reply nor post anything back to help others when I knew the solution because of the way how it works.
To make matters worse while working in IT, I worked with a guy that didn't know anything, if there was no SO post about the problem, the guy couldn't fix the problem.
I have been using Perplexity AI and it has been awesome, and it does provide all the sources it used making it easy to cross check the answers.
It has helped me to speed my python learning curve, I am not using search engine anymore, and SO has the problems mentioned above so I have zero interest in using it.
Also, the website layout is a mess, I have to use uBlock Origin with a ton of element picker to stop loading half of its crappy.
I'm curious if Stack Overflow was a good resource for human learning? As in, run into a problem, look through other related questions and answers, check the documentation, struggle for awhile learning and figuring out how to frame and pose the question, struggle some more while you wait to get an answer. I kind of find LLMs "too easy" and can distinctly feel myself "not learning", not the way I used to, but after all I am getting older.
I'm pretty sure you can get a stack overflowy experience out of an llm with the right system prompt, but the human factor might not be the same. Not wanting to be berated by others on the internet is maybe underrated as a motivational tool. How are we going to get that back?
Not to mention the helping and teaching other people aspect, which feels good, it feels like there isn't as much of that now. Or maybe it has just moved away onto Discord etc, but GitHub definitely doesn't feel like it has that aspect, if you don't also personally know other contributors and reviewers. Curious if others feel this way.
From my perspective LLMs shine in programming languages because they (languages) are invented by people to be more formal so they are more predictable. Even for syntactically different programming languages there are similarities we don't see with our own eyes, but the training has no problem detecting. My favorite example is when I tried to find how to represent an int array for reflection in Kotlin. Ask ChatGPT or a sufficiently large model about this, the answer will be IntArray::class.java , but try to find the exact line with google - few or nothing, more on github search in the sources, still not much. So LLMs "detected" the system of making type signatures in Java/Kotlin and were able to successfully predict because the rules are consistent. In human language it also works, but to a lesser degree, so if you give it a verb/subject pair that makes sense and ask for the equivalent ones, you will get some that still make sense but are literally absent in the full corpus of the web.
How will there be new programming languages after people stop practicing programming and the skills decay because an AI is performing the deliberate practice?
Are there any similar articles on the state of the rest of the Stack Exchange network? There are many, many other SE sites that have nothing to do programming and are often less amenable to being answered via LLM.
> The question seems to be when Stack Overflow will wind down operations, or the owner sells the site for comparative pennies, not if it will happen.
I see the latter option, but the former? SO, at least judging by their hardware posts, was running on nine servers two years ago [1]. That's barely anything in costs - electricity, uplink and occasional rotation of the hardware, but probably a single person working a decent job can afford to run the entire hardware for the site.
Truly shows how far a tight budget can go when you don't waste untold amounts of money and energy on layers upon layers of complexity.
> I'm sure we'll see spaces where developers hang out and help each other continue to be popular – whether they are in the form of Discord servers, WhatsApp or Telegram groups, or something else.
Yeah fuuuuck that. It's so annoying that everyone and their dog moved to these walled gardens. Google can't pierce them, unlike IRC of ye olde days where it was common to let a bouncer publish logs, WA/Telegram come with privacy risks and Discord is a hellscape.
Set up StackOverflow according to "cloud native" recommendations and whoops, that's gonna be quite the bill, which is my point. You'll have Cloudfront for global load balancing, ELB to provide a bridge between Cloudfront and Kubernetes, EFS for storage, RDS for the database, EKS and EC2 for compute, ElastiCache for KV cache, add in CodeBuild for build/deploy pipelines... AWS has quite the hefty overhead.
To Ai paraphrase Obi-Wan: I felt a great disturbance in the Force — as if millions of developer voices from the future cried out in confusion and despair, and were suddenly silenced. I fear something terrible has happened to the future of our craft.
In short before we Ctrl+Copy our way to StackOverflow or Forums or IRC and we got collection of responses between good and ugly.
But made us think and read or talk to others.
Now we Ctrl+Copy into LLM into a room of hell.
If LLM function is useful but don't get addicted to like honey.
my business has a similar revenue curve, and of course LLMs are a big cause, but it's more to do with me being distracted from moving with the times. I just didn't fancy the fight and saw enough value in the curve to let it follow ranged decline
I'm assuming the owners of stack felt similar? Don't know anything about them so could be easily wrong
It will neither be the first or last site destroyed by hamfisted power tripping overmoderation. And they do it in the worst possible way - Leaving the question up, but preventing anyone from answering it. So, the default experience with SO becomes finding your question already asked and unanswered.
Given the speed at which tech evolves, I do not see how that can be true.
I think it may answer most questions from five years ago, but I would phrase it as “it has answers for most questions, but many are outdated and there’s no easy way to tell whether an answer is outdated”
The graph shows monthly questions asked, not monthly questions answered. It makes total sense that the slope of that graph would be much higher earlier in the life of the site; as time goes on, many common questions will have already been answered.
Stackoverflow answers kind of provided a source of truth by being confirmed / upvoted by people trying them out. We're completely abandoning this medium (on which LLMs are trained), even as technologies keep changing. Perhaps coding agents should start posting to stackoverflow too...
You don't start and run StackOverflow because of the money. You do it because you really care. Running niche communities isn't a lucrative path, and (as we're seeing now) they can be quite fickle.
Additionally, it's not just the owners... it's the millions of people who contributed.
Per the post, it seems as if SO was going to going to slide beneath the surface anyway but that LLMs were one of a number of factors that maybe accelerated the process by maybe a few years. Without having deep knowledge of the area, this feels about right.
Kind of a catch 22 isn't it? LLMs trained off stack overflow data. Without new data for them to train from, what will happen to the quality of LLMs in the future?
How will technology advance without research sharing?
I wonder if they could survive if they added a mode like phind.com wrapper - where each question stores LLM answers as answers. So (SO!) many more content and would people reason to still use it.
Everyone is blaming AI, and it's undoubtedly a factor.
But also, the culture of Stack Overflow has changed significantly over the years. It used to be a place where anyone could ask a question and get help with a problem ... and it was amazing.
Today, you're far more likely to have your question downvoted, flagged as a duplicate (of an unrelated question), or attacked in the comments by overzealous responders (and once that happens, good luck on actually getting help). Your odds of actually getting help on the site are only a fraction of what they once were.
And I'm not just saying this as some SO newbie: I've been using the site since beta! As someone who has used it that long, the change in quality is undeniable.
Can you link to questions that you feel were unfairly derided (not necessarily closed)?
As someone who has extensive experience modding an niche SE I see this sentiment quite often but honestly, the people making these complaints are just arrogant _and_ wrong about the topic they needed help with.
Others accelerate their decline through self-foot-shooting and/or enshittification.
Stack Overflow's journey into obscurity is via a mix of private equity indifference, better docs elsewhere, and a lack of leadership over its moderators. It was in decline long before LLMs.
It is not a new story - but it does help map out the modes of platform senescence.
No wonder, the CEO basically said they'll use the free labor of love of all member devs to sell training data for AI in order to replace such devs so why would anyone keep contributing there?
They really only have themselves to blame. Yeah AI has massively accelerated the decline but I think it's mainly provided an option that isn't so frustrating to use. ChatGPT never says your question is unclear or off topic.
I think if they had actually fixed moderation they may have had a chance of surviving, but I think they got trapped by relying on volunteer moderators who thought that it was good that so many valid questions were closed.
They did actually make some attempts to fix things, e.g. I remember one suggestion from the company that users could reopen a closed question at least once (which is a great thing to try!) and mods downvoted that to hell so they chickened out.
Definitely some shadenfreude, and I say that as someone with 100k reputation.
> ChatGPT never says your question is unclear or off topic.
Yes, because ChatGPT doesn't care about publishing your question and its answer with the explicit intent of enabling other people to find it later with a search engine. It has no mandate to organize content or care about the quality of its content. It has no reason to care about the topic. Other people will not read your prompt (which doesn't even need to be trying to ask a question at all), so there's no reason to care about whether other people can understand the question - let alone figure out whether it's something they also want to ask.
> I remember one suggestion from the company that users could reopen a closed question at least once (which is a great thing to try!) and mods downvoted that to hell so they chickened out.
Questions can be reopened, and this has always been the case. What is not allowed is for the OP to reopen a question unilaterally, because that defeats the entire purpose of question closure. The entire point is that when you come to Stack Overflow, you aren't the one who gets to decide what the standards for questions are, or judge whether your question meets those standards. Because if you were, the standards would effectively not exist.
The people you think of as "mods" are overwhelmingly not.
> I say that as someone with 100k reputation.
I have found that reputation scores say very little about whether one understands the purpose of the site or how it's designed to work. Famously, a user with over a million reputation once went on a spree of violations (https://meta.stackoverflow.com/questions/430072) of the best-publicized site policy ever to appear on the meta site (https://meta.stackoverflow.com/questions/421831/). Many users with 100k+ reputation and/or a 10+ year history have likely never looked at the meta site. (See https://meta.stackoverflow.com/questions/427224 for someone who first posted on meta after 14 years.)
> It has no mandate to organize content or care about the quality of its content.
That's fundamentally where SO went wrong. The mods think they're building programming Wikipedia, but normal users are using it as a Q&A forum.
> Questions can be reopened, and this has always been the case. What is not allowed is for the OP to reopen a question unilaterally, because that defeats the entire purpose of question closure.
No it doesn't. It's far too easy to close a question and they very rarely get reopened. The suggestion was that users only get to unilaterally reopen it once.
> I have found that reputation scores say very little about whether one understands the purpose of the site or how it's designed to work.
See this is exactly the problem. You have the mods' view of how it is supposed to work, but that isn't how people want it to work!
Mods want it to be this highly curated reference site where only perfectly written questions that exactly fit an FAQ style of questioning. They don't care that that makes it useless (or at least extremely frustrating) as a Q&A forum.
> This tallies with my memory of feeling that site moderators had gone on a power trip by closing legitimate questions.
Respectfully: outsiders like the author of this piece are not the ones entitled to decide whether a question is "legitimate", or "valid" (another term I see used all the time by people who have no understanding either of Stack Overflow's standards or its goals).
Interestingly, the site stopped growing at about the same time when the "fun killers" [0] took hand of it. Notably, when they deleted the all-time highest voting question
"New programming jargon you coined?".
This blatantly undemocratic and destructive behavior was of course duely punished by the (former) users of the site.
All the "overflow/exchange" sites suck, not just the stack overflow. Too many questions being closed. I had a question closed in mathexchange because, I surmise, it was "too obvious" even though easier questions were asked as recently as 2021, even much more elementary ones. Moreover, my specific question had never been asked there before, and the point of math exchange is to ask easier questions, compared to mathoverflow.
But it's also possible it's pivoting to a Wikipedia-like model where it becomes a repository for answers, and less about contributions. In which case, this is not the same as it dying. As seen with Wikipedia, it can still get a lot of traffic and revenue even if few people contribute to it anymore.
My heart goes to the stack overflow community which has always been very kind and helpful, essentially working for free. As a self-taught developer since the age of 8, I literally grew up learning how to code through SO, asking hundreds of questions and answering many more. So many bugs that would take 2-3 days to fix would eventually find their answer through it. But now ChatGPT does that in minutes… so it’s for the best!
The presumption is that things will improve over time, but the big difference in my experience is the assistance I got from SO _worked_ the vast majority of the time, whereas various LLMs I have used generate unusable, misleading, or unreliable results pretty regularly, increasing as complexity or rarity arises. As human-driven knowledge bases backed by actual experience are replaced by inference from models who rely on such inputs, I am concerned about the medium to long term impact. A lot of people grew frustrated with SO for various reasons and went back to unhelpful behaviors that SO resolved at its zenith (rather than dead ends and flame wars in newslists and irc channels, they do it in random subreddits and discord servers instead). Now what if we circle back after degenerative LLM experiences only to find there’s nothing to circle back on?
Personally, I am getting extremely tired of ChatGPT hallucinating npm packages that don't exist, or package imports that do not exist
9 replies →
I think this is the first time I’ve ever heard someone describe the stack overflow community as “kind”. Usually it’s the exact opposite: “I asked a question and just got 30 questions asking why I’m trying to do this” or “my question was closed for seemingly no reason”.
It’s literally the most blunt and aggressive website I’ve ever been on that wasn’t a straight-up troll site like 4-Chan.
I answer questions for a few tags, know the styles of the other people who answer on those tags, and I do consider my fellow answerers as kind. All of them.
That doesn't mean that you will think we're kind to you personally. We're there to build a searchable Q&A knowledge base and spread knowledge. Some people who ask questions misunderstand and think we're there to help them, personally. To work for free for that single person, and we're not there for that. We write answers for the tens, hundreds, thousands of people who will search for it.
Askers who misunderstand will come across as overly entitled.
In terms of practical effects: People who misunderstand don't tag their questions, or tag them incorrectly. They post screenshots full of text. They don't look for similar older questions in the existing knowledge base, or they insist that even slightly different questions are significantly different. All rather offputting, and often puzzling. How can you ask for a subject expert's help and simultaneously insist that you know better than the expert whether your question is a duplicate of another?
35 replies →
There was a time when it was really good. Like legitimately good and useful. But over time it ended up becoming exactly what you describe. But there are still countless examples of the usefulness of SO in Google results. I stopped asking questions in 2012 and stopped answering questions in 2015. Before that though, it was a very useful tool.
Question closed; here's a link to another one that sounds vaguely related but doesn't actually address your problem.
But seriously, I'd love to see some sentiment analysis of the SO corpus classifying tone by tag.
32 replies →
Agree! A decade ago it wasn't like this. But it has devolved into a community of vandals who seem to take glee in criticizing the manner in which a question is asked, rather than contributing a solution.
5 replies →
Yep, believe it's a direct result of Atwood's iron-fisted no-bullshit policy. To some extent it is great... don't want it turning into Yahoo Answers, do we? I think folks forget about that part.
But, as you mention they just went too damn far with the medicine.
No, you can't fix this misspelling, isn't there something else (with more characters) that you can improve as well? WTF, for realz? :-/
1 reply →
Mixed feelings on SO. It was helpful, but it was also a website you dread having to post on because it was filled with the most intolerable people of the internet who you just had to endure abuse from if you wanted help.
Now chatGPT gives you the same help without the abuse.
The next AI totally needs to be more snarky to make it feel more like we're dealing with actual "thinksperts", people that think they are experts even if their answers are demonstrably wrong.
3 replies →
> But now ChatGPT does that in minutes
But it's trained on stackoverflow data? What happens in a few years when the data gets more and more outdated? Where will it get its knowledge then?
They're learning from working code in GitHub, IDE "co-pilots"...
2 replies →
It will steal our own data and we'll have a big "oopsie! didn't mean to!" moment 5-10 years after.
2 replies →
Why does any LLM need new information to do fundamentally the same thing?
And what makes the data outdated? New code? It can train on that. That, or there is simply nothing new to learn, just new ways to express the same thing.
1 reply →
Are you talking about Stackoverflow? Every time I asked a detailed question it would be closed within minutes.
I'm not surprised it's on the way out.
I've asked dozens of questions on SO, and never had a single one closed. I hear your sentiment often, but have no idea whether my experience or yours is more common.
I've had 3 deleted by Community bot as abandoned, but since they were over a year old when that happened, I couldn't care less.
I've got valuable advice from SO over the years. There's overlap with LLMs, sure, but it's frequent to have questions that have no answers published anywhere on the web; SO brings people who know out of the woodwork, who create an explanation that didn't exist before. A couple days ago, someone in retrocomputing got to bank-switch a 1983 Radio Shack box... that kind of stuff wasn't published anywhere, until a guy who used to write games for that box answered that question on SO.
Until there is a radically new version of {popular programming language} with breaking changes and no new and correct answers to train on.
These models can figure out syntax and language features they haven’t seen before. Try it with a few code snippets of your own made-up language. It’s a little freaky.
3 replies →
The point of an LLM is that it can take your problem as input, along with answers to previously asked questions (perhaps implicit) in its training data, and attempt to synthesize a solution to a problem as output. Here, a "question" is something that can be found with a search engine - something that directly presents the/a crux of an issue, which is identified after a debugging session (for a problem in existing code; for projects still in the design phase, a how-to question emerges after coming up with user stories and breaking tasks down into their logical steps). The point is that the question can be relevant to many different people who have written different code which encounters different problems - all caused fundamentally by the same conceptual misunderstanding (or non-understanding).
Stack Overflow is explicitly not designed or intended to solve problems or do the decomposition of the problem for you, nor the synthesis of answers. Because the result would never be useful to anyone else. The entire point is to have something searchable, and to allow answer-writers to keep their explanations DRY.
This has spectacularly failed, because no matter how frustrated people get with traditional discussion forums (https://www.explainxkcd.com/wiki/index.php/979:_Wisdom_of_th... among many other typical complaints), they apparently are much more suited to human nature.
Heaven knows how Wikipedia managed to avoid devolving into "Quora but even worse because you can scribble over someone else's post".
Closing this as its a statement not a question
I moved this over to the Stackoverflow Will Be Missed thread. It does not belong in the Let's Dance On Stackoverflow's Demise thread.
I just hope that we can continue to find sources of high quality training data like SO. If people don't publish their mutual learnings somewhere then there's no data to train on.
Ironically, when someone first asked the Meta SO community if ChatGPT could ever become a threat, it was downvoted with prejudice:
https://meta.stackexchange.com/questions/384355/could-chatgp...
It was downvoted because the entire purported "threat" is based on a misconception.
It is not relevant to SO whether an LLM can provide personalized help, write with any particular tone, answer promptly, accept every input prompt as valid and try to make sense of it, discuss back and forth to figure out a problem, etc. Because Stack Exchange is explicitly and by design not for those things.
But also, downvotes work differently on meta anyway, and the community there generally takes a negative view of LLMs. Because, again, the point of SO is for the answer to come from a human expert, and be verified against subject matter expertise rather than simply being evaluated for coherence or generally sounding appropriate in context.
> 2014: questions started to decline, which was also when Stack Overflow significantly improved moderator efficiency. From then, questions were closed faster, many more were closed, and “low quality” questions were removed more efficiently. This tallies with my memory of feeling that site moderators had gone on a power trip by closing legitimate questions. I stopped asking questions around this time because the site felt unwelcome.
I also felt around that time that it became unwelcoming. I didn’t realize they had revamped the moderator tools. That is the time period when I stopped using it too. Now I know why.
How many other websites have also shot themselves in the foot by tweaking things?
Yes I felt the same around then. It seemed like stack overflow was sending the message to not ask questions anymore. It was really weird.
Same. I've never been a huge StackOverflow user, but it is so irritating to search and find your exact question on StackOverflow, often as the top result, only to see that it was instantly shut down two years ago as a duplicate of some other question in another context with inapplicable and useless answers.
It is frustrating not only because you can't get instant help, but also because it shows the futility of even trying to post on there.
people who own walled gardens often get the idea in their head to prune up the trees a little, make them elegant looking and pretty, that's moderation, deletion, banning,etc. they get a good feeling pruning the tree, making something beautiful, moment of joy, thinking that such a beauticious little garden they've made will make it all the more appealing to visitors and potential garden supporters.
some have a tendency to go overboard with this thinking, only to discover that a heavily pruned tree is now a dead tree, now finding themselves in dead tree garden.
For a time the "let's interact with people and talk about cool things" group and the "let's build the ultimate knowledge base" group had their incentives aligned.
Then, with better moderator tools, the "ultimate knowledge base" group set out to achieve the ultimate knowledge base by reducing the amount of people who were just there to talk.
> Then, with better moderator tools, the "ultimate knowledge base" group set out to achieve the ultimate knowledge base by reducing the amount of people who were just there to talk.
Yes, because the people who were just there to talk had reached a point where they could effectively only pollute the knowledge base. Bad questions make good ones harder to find, simply by existing (since the bad one could potentially be found instead, and because of the broken window effect).
I didn't start learning to code until around 2014, so I didn't know SO before this change. But when you make your platform so militantly anti-helpful, militantly anti-newbie, and just all around unpleasant to interact with, then it shouldn't be a huge surprise to anyone that people stop using it. There's over half a human generation of learners who have had almost entirely negative experiences with SO.
How ironic. "AI" feeds off structured knowledge, artistic creations and otherwise any human production to generate its output. As a consequence of its widespread adoption, people start to lean even more towards consuming rather than producing, a tendency which was already increasing before the advent of LLMs and modern machine-learning. This, in turn, leaves "AI" implementations with no new human content to feed off of. Now what? The whole process folds onto itself. Are we entering the dark ages of cultural (in the widest sense of the word) production? Not that I don't think that we're already there, in any case, but for other, somewhat related causes...
Perhaps the next step is having the LLMs ask questions on SO when they routinely fumble particular topics. I could see a system of knowledge bounties where people are compensated for providing accurate, in-depth training data on niche topics.
LLM content is banned everywhere on Stack Overflow, in both questions and answers, by policy, since mere days after the public announcement of ChatGPT (because it was immediately causing a huge problem): https://meta.stackoverflow.com/questions/421831
Moderators (actual elected moderators, the two dozen or so that exist for ~29 million user accounts and ~24 million non-deleted questions) went on strike in mid 2023, largely because the site staff/owners interfered with their ability to remove such content (an overwhelmingly popular policy with strong community consensus): https://meta.stackoverflow.com/questions/425000 and this decision propagated across the Stack Exchange network (as most SE sites had adopted similar policies): https://meta.stackexchange.com/questions/389811/
A large fraction of the userbase is explicitly opposed to helping LLMs out in any way whatsoever. I personally have ceased contributing new question or answer content, and only edit existing posts. I contribute new content on Codidact (https://software.codidact.com/) instead (disclosure: I have recently become a moderator there).
you’re one or two additional sentences away from the plot to The Matrix
> This, in turn, leaves "AI" implementations with no new human content to feed off of. Now what?
You seem to be under the impression that AI needs more than all recorded human knowledge up until 2024 to reach the same level as an average SO contributor. It doesn't. Because none of the average SO contributors did.
It is unclear what algorithmic improvements are required to leverage the available data to get AI to AGI, but a lack of data is definitely not the bottleneck.
One could say that these AI systems aren't sharing their solutions (or questions) with other AI systems and that the world would benefit from it if they did, though. Perhaps it's a good idea to have some shared space for AI systems where they share the validated solutions they synthesized.
> You seem to be under the impression that AI needs more than all recorded human knowledge up until 2024 to reach the same level as an average SO contributor.
Replacing the average SO cobtributor isn't adequate to replace SO, and AI is able to “replace” SO effectively only since major models have gotten not only SO-as-training-data but web search (including SO) for immediate grounding.
And without SO or something like it with active human contributions it’ll have even more trouble replacing the value SO would provide for new questions and new domains where it will neither have SO traijing data nor SO query-time-search-results to use to synthesize answers.
1 reply →
I find it interesting that the current StackOverflow moderators tend to say "in the past we used to accept too many questions but it was never the goal, so now we are doing it as it was meant to be".
Sure, but in the past, StackOverflow was growing, and now it's dying. Maybe something was better before, when "it was not done correctly"?
I think these sorts of things are just an unfortunate side effect of scaling. The bigger you get the more people get lost in the bureaucracy. However if you don't build up the bureaucracy the system collapses under its own popularity.
Wikipedia has a similar issue where editing declined around 2007, which is often blamed on stricter enforcement of rules, more complex rules, etc. I think its just a natural stage of growth. You can't be a free for all forever.
The "good" thing is, they're back to 2009 levels of postings. Now obviously that's what the mods let through but my guess is that traffic to the site is down precipitously as well. They can roll back their bureaucracy and head back to a lean path that worked for them in the past.
But I don't really think that's the problem. Reading zahlman's responses in this thread makes me think that the mods fell into the age old trap that's happened since Usenet, IRC, and still happens to this day wherever there's mods: they got tired of doing unpaid labor and instead of deciding to quit decided to become meaner and stricter. The age old mod trip.
2 replies →
Two thirds of the wikipedia article i wrote in 2003 have been deleted by rabid editors. It was a biography of my father, written based on interviews with my mother. I have found that restoring any of the rabid editor deletions results in threats of me being banned from editing my own article.
1 reply →
>Sure, but in the past, StackOverflow was growing, and now it's dying. Maybe something was better before, when "it was not done correctly"?
"Growing" by an utterly irrelevant metric.
Popular != good.
> Sure, but in the past, StackOverflow was growing, and now it's dying. Maybe something was better before, when "it was not done correctly"?
You're presuming that the current volume of questions represent novel, unique posts instead of something you can find over and over again if you do a decent query.
AFAICR they've always said these lines about now is about better moderation from the slop. The reality is that the rule of thumb for that moderation was already out of date with advances that preceeded LLMs.. Even with the beginnings of computer aided flows we didn't need to alienate most to get the best content and develop the few. Content can be triaged from someone who may be human to others who may be human and maybe there's value or maybe you just didn't alienate anyone and some people will still climb to making higher levels of content that is worth condensing.
> Even with the beginnings of computer aided flows we didn't need to alienate most to get the best content and develop the few.
The large majority of new questions from new accounts are from people who are clearly there only to solve a personal problem, who show no interest in considering the value of their question to third parties, and rarely put any effort into attempting to even diagnose or specify a problem.
Even after it became possible for most of these people to get an instant answer from an LLM. Which is actively preferable from the standpoint of Stack Overflow curators. Before LLMs, the point was for them to use a search engine to find an existing question that lets them figure out the problem. But for the Q&A to help such users, they need to apply at least basic problem-solving and debugging skills. (It is explicitly out of scope for the Stack Overflow community to do that for others; and attempting to do this in an answer actively degrades the site for everyone else.) If an LLM can fill in some hypotheses for those users to test, then the LLM is doing what it's best at, and Stack Overflow is doing what it's best at.
Stack Overflow is not there to troubleshoot or debug anything for you, nor to reason about a multi-step problem and break it down into its natural logical steps. It's there to give a direct, objective answer to how to do each individual step, and to explain why the specific point of failure in a failing program fails, after you have identified it and made the problem reproducible.
So yes, we absolutely do need to "alienate most", because "most" are there for a reason that has absolutely nothing to do with getting the best content.
8 replies →
>the current StackOverflow moderators
Overwhelmingly, the people you're talking about are not moderators. I explained this to someone else a week ago (> Sure, but in the past, StackOverflow was growing
So what? Stack Overflow users get $0.00 for this, whether they're moderators, active curators or just signed up. For users, growing the site isn't the goal. Growing interaction with the site is not the goal. The goal is building a useful artifact (palata
1 year ago
> Overwhelmingly, the people you're talking about are not moderators.
I was actually thinking about you. You keep saying everything is great. My observation is that I used to be on SO every day, and I completely stopped contributing even though I would have plenty of stuff to add (more than ever, actually).
> Why should a reduction in incoming questions mean that it's "dying"?
There is "a reduction", and there is "being back to the amount of questions SO had in 2009 when it launched".
7 replies →
The timing on the sale was genius. Similar to Mark Cuban with Broadcast.com. I guess it's best to sell something before the value plummets to 0.
As far as its demise? AI ate its lunch. I use to use Stack Overflow all the time and haven't even gone to the site for a couple of years now.
The new owners have been trying very hard with the "if you can't beat 'em, join 'em" approach. They know this is radically against community consensus (it's been shown to them on the meta site over and over) - so they just get sneakier about it.
Notably, after getting completely humiliated with https://meta.stackoverflow.com/questions/425081 in June 2023 (right after a moderator strike had just started, protesting the staff trying to prevent them from removing AI content from the site), and getting embarrassing feedback on the feature (https://meta.stackoverflow.com/questions/425162), they came back last November with https://meta.stackoverflow.com/questions/432154 and have been forcing it through.
Where's the AI of tomorrow going to learn from if nobody is posting Q&As online anymore?
That’s a tomorrow problem.
(As someone who is all too often hired tomorrow, at a fraction of the before rates, to clean up this mess)
information will be repackaged like credit default swaps in the mid 2000s.
From the official docs? AI is good at sumarizing after all
3 replies →
Other people's code on GitHub.
2 replies →
AI just stole all its content, I wonder if they will choose to sue.
Stack overflow never owned the content, it is and was Creative Commons: https://stackoverflow.com/help/licensing
9 replies →
Doesn't feel like the AI is the main driver. Many things changed over time - dev tools got better, editors got smarter, compilers got better error messages, various primary resources improved, tutorial websites, courses and youtube boomed.
Another point of course is that each new question is more and more likely to be already answered. At some point the site pretty much covers most of what is to be answered.
One aspect I haven’t seen anyone mention contributing to the decline is GitHub (part of your “improved tooling”)
These days you can go to the repo and there’s usually already an issue open with the problem and a workaround. Or if someone has a question on how to use the tool/software they ask there.
Before GH boomed it was often SO doing this job.
It could be the most impacting aspect IMHO.
This, and first party developer forums. iOS questions will go directly to Apple's community forums. Same for SalesForce, or Elastic search etc.
There's just a higher noise/signal ratio, a real chance to get answers from experts, and it makes for a stepping stone if the issues needs to be bumped to paid support.
> each new question is more and more likely to be already answered
Yeah, except for when there should be current answers. Most of computing is in constant flux. There's a mountain of 10+ year old answers that simply don't apply any more.
> except for when there should be current answers.
Yes.
They belong on the existing question - unless the existing question is poorly asked and the new one is better asked.
New answers can, by default, be added at any time - and should, if existing answers are actually out of date rather than simply being old. (Many 10+ year old answers really do still apply.)
Asking the question again is not how the site is designed or intended to work.
I answered one question 13 years ago where I still gets points for. Computing as a whole isn't so much in constant flux, it is only JavaScript that changes so much. Hell, I learned to use Apache 30 years ago, and I didn't need to learn anything new the last 25 years of that.
1 reply →
There was a noticeable inflection in the question-rate-vs-time curve around the time that ChatGPT was released.
Which is fine. If your question is not answered by `site:stackoverflow.com how to do the thing` but it is answered by an LLM taking `how do I do the thing?` as a prompt and synthezising existing Stack Overflow content, that is inherently not a suitable Stack Overflow question. Because anyone else could put `how do I do the thing?` into the same LLM. It's not any different from using a traditional search engine.
(And when the LLM fails by producing a wrong synthesis, then blessing that result by putting it on Stack Overflow is actively harmful - which is why it's banned by policy.)
> Another point of course is that each new question is more and more likely to be already answered. At some point the site pretty much covers most of what is to be answered.
AI has the ability combine answers from multiple sources and tailor-make to your exact prompt details. Now that is something we call glove fits the hand. Plus it can explain its answers.
Yes, it can combine multiple sources and make up an answer that makes no sense. Even if it explains how "it works", it does not help when the API or a function has never existed.
Woof. Looking at a single metric and extrapolating "LLMs killed the radio star"
Stack Exchange sites are designed to nuke duplicates, help people before they post a new question. It seems a natural conclusion that the number of original questions decreases over time.
I won't pretend that some people live their lives inside and LLM but many of us still use search engines and SO.
It is a bit sad. And obviously the reason why it sees such a decline is because AI (ChatGPT and similar) took the job of answering the basic questions about programming that StackOverflow used to help with.
Looking at my profile since 14 years ago, the most upvoted answer that I solved was about a basic question of how to specify fields properly when you serialize JSON into a C# class.
I do believe the value of StackOverflow was only about people who were lazy enough to read the documentation of the language/framework they were trying to use. I used to be active on StackOverflow back in the days, but in the last 10 years the only value I saw in it was if I needed to get back to some language to just find an answer on how to write a for loop in that specific language (swift vs go vs ...).
I personally do not believe there is much knowledge base on StackOverflow. In most of my questions to "google" for the last 10 years, very rarely would I be directed to StackOverflow for the right answer.
There are a lot of complicated questions on StackOverflow, but the site was flooded by people asking and answering basic questions about programming. And people who are there just to get some karma.
>>And obviously the reason why it sees such a decline is because AI (ChatGPT and similar) took the job of answering the basic questions about programming that StackOverflow used to help with.
A big reason why AI is replacing these things. A big part is the experience itself. There are quite a few people who have have been repulsed due to the smugness, or other wise having their questions marked duplicate/irrelevant/stupid etc.
AI is also pretty much instant. You can also talk to it like you are talking to a person.
The killer AI feature!---> AI listens, without judging you.
I used to answer questions a lot, by around 2013 I had answered maybe ~12% of all HTML canvas questions ever asked. To me it declined a lot sooner, 2014 really does feel like the right inflection point.
There was a belief, sometimes unstated but often explicit, that no more (serious) discussion is really to be had, and further wondering how can one stop people from asking. It became difficult to discuss anything if there was even something vaguely related asked before. It was not possible to discuss something you knew the answer to, but did not know why, or wanted to hear arguments for which of 5 ways might be best. All (to me) very worthwhile technical discussions. Totally shut down.
> There was a belief, sometimes unstated but often explicit, that no more (serious) discussion is really to be had, and further wondering how can one stop people from asking. It became difficult to discuss anything if there was even something vaguely related asked before.
There is nothing to do with unstated belief here.
It is explicit policy that we don't have discussion at all.
We have answers to questions.
Which is why there's a question (explicitly labelled as such, and not just "help me") at the top of the page, and every post below it is labelled as an answer (and is explicitly not a response to anything else but is simply there to answer the question).
StackOverflow should have focused on linking duplicates rather than forbidding duplicates.
No Boilerplate recently said "writing is thinking"[0], and suggested links are the ultimate knowledge graph organizational tool--not tags, not folders--links[1].
StackOverflow tried to prevent all duplicate questions. This was stifling and reduced writing, reduced thought, and most importantly, reduced user engagement.
The people who wanted to write their problems and ask their questions stopped going to StackOverflow. The people who wanted to write and give answers stopped going to StackOverflow.
Look at Discord or IRC and you'll see that people have their own questions to ask, and the people who answer such questions enjoy answering the same questions over and over. Let the people write their questions, and write their answers and give advice. Instead of preventing duplicates, link duplicate questions together.
[0]: https://www.youtube.com/watch?v=sqm4-B07LsE [1]: https://www.youtube.com/watch?v=B0yAy2j-9V0
> StackOverflow should have focused on linking duplicates rather than forbidding duplicates.... StackOverflow tried to prevent all duplicate questions.
Duplicates are not forbidden on Stack Overflow. Duplicate questions are linked together - that's what "duplicate" means in the system. Beyond that, logged-out users who land on a duplicate question that doesn't have its own answers, will be automatically redirected to the target.
When duplicate question are linked on Stack Overflow (and everywhere else on Stack Exchange), they are automatically closed, which merely prevents new answers. The purpose is to allow high-quality answers to be gathered in one place - on the duplicate target, which in turn is ideally a high-quality version of the question (and a focus for curators to improve further, when they notice that it becomes a common duplicate target).
Deletion of duplicates (and posts in general) is not very well understood and people are not all on the same page - see e.g. https://meta.stackoverflow.com/questions/426214 . But normally duplicates should not be deleted unless they are actively harming the search results (i.e.: it causes people to find the wrong target, because it was written in a way that fools keyword search).
Nobody gets formally sanctioned for posting a duplicate. You just get your question closed, and maybe asked to search a bit more carefully in the future (or shown how to do so).
> The people who wanted to write and give answers stopped going to StackOverflow. Look at Discord or IRC and you'll see that people have their own questions to ask, and the people who answer such questions enjoy answering the same questions over and over.
People who want to give the same easy answers over and over to the same easy questions, so as to get imaginary internet points that move them up a leaderboard, should not feel welcome on Stack Overflow. That action is actively counterproductive to what Stack Overflow is trying to accomplish. There are countless discussion forums (and as you say, Discord and IRC channels) already where that behaviour is valued. The Internet should be allowed to have one place where it is not valued - especially when it's a place that was specifically created to accommodate people who want search engines to be useful; who want to write high quality answers once and get many people to read them; etc.
Maybe "answered by duplicate" would have been a more friendly way to say it--but I hear you, yeah, the closed questions were linked, which is what I was asking for. Whatever the case, a lot of people decided to never use the site because their questions were closed.
> People who want to give the same easy answers over and over to the same easy questions, so as to get imaginary internet points that move them up a leaderboard, should not feel welcome on Stack Overflow.
I think it's worth asking, why do people give answers on the internet at all? Maybe it's because of internet points, but more often people just like interacting with other humans and teaching.
In the beginning, StackOverflow was a place for people who wanted to interact with other humans, and also a place for people who wanted to build the ultimate knowledge base--for a time their incentives were aligned.
But then over time the space for interacting with other humans got smaller and smaller, and now StackOverflow is almost entirely about maintaining the knowledge base that has been built.
And yeah, like you say, it's okay if StackOverflow isn't the place for human interaction. StackOverflow has built its knowledge base, and some still maintain it, and the long term success of that knowledge base is becoming ever more apparent--which is to say, not very successful--the day may soon come that StackOverflow isn't even hosted anymore.
(Also, I want to ward off the claims that this is because of AI. StackOverflow was in steady decline long before AI was competent at answering questions; even in 2025 the competence of AI is still in question: https://news.ycombinator.com/item?id=44000118)
1 reply →
I think "number of questions asked" is the wrong metric. Because it feels like all the questions have already been asked. Whenever I need to know something, I can google it and find answers on Stack Overflow. I can't remember the last time I actually had to ask something. Or the last time I found a question that didn't already have a good answer. Stack Overflow's library of question is pretty complete, and the only reason for new questions are new tools.
Certainly LLMs are a huge factor, but I feel that LLMs rarely give good (and trustworthy!) answers to the things I would check on Stackoverflow. Just like LLMs are no good replacement for API references because they get the details wrong all the time.
I think this is true if there aren't new questions to be asked. But technologies shift and evolve all of the time.
One of my top StackOverflow questions for years was around the viability of ECMAscript 6. It's now essentially irrelevant because it's found wide adoption in browsers etc. but at the time a lot of people appreciated the question because they wanted to adopt the technology but weren't sure what its maturity was.
It's also true that some technology stacks mature to a point where there isn't much more to be asked but I think there will continue to be a place for forums of discussion where you can ask and get answers around newer, bleeding edge technologies, use cases etc.
I find that most of the time, when doing research on anything non-trivial, I find a question on SO about this exact problem that has no answers because it was closed by the mods as a duplicate of something that doesn't actually answer that question (but rather something very vaguely related to it).
Where will training data come from for new tech & programming languages if SO dies?
This was my question. There's a weird sort of self-cannibalism that this hints at. The LLM is only as good as it is because it's been able to train on existing SO answers. But if over time, SO content production declines, then the LLM results will be less reliable. It seems that a new equilibrium could be one in which -- for newer questions/concerns -- both SO and LLMs will be worse than they are now.
All the same places it comes from for human programmers before a language has many answers on SO.
- Documentation - Open source projects using it - Github issues - Source code - Blogs - Youtube videos
The list goes on
To add a bit more nuance, SO has a question-answer type format, which leads very well into prompt-rely format to train these chat applications. Most of the other sources do not, except for Github issues maybe. Without this question-answer format, there'll be a need for a bigger data labeling effort to train LLMs on new stuff, no?
Stack Overflow was a question and answer site that discouraged people from asking or answering questions.
LLMs probably sped things up, but it seems like it was inevitable that it would fall into disuse and eventually be overtaken one way or another.
Most of what people "ask" on Stack Overflow is simply not a question per the intent and design of the site.
https://meta.stackoverflow.com/questions/284236/why-is-can-s...
Your comment looks like a good one but you've posted it as an answer so I've no choice but to delete it.
When you earn enough reputation please come back and re-post your comment.
My approved to shut down question rate was about 40/60. And within the approved questions, I only got the answers I looked for like 50-60% of the time. Got to 350 points myself. I hated every second of it.
Really aggressive moderation, people trying to score points for a worthless achievement system by spamming comments like "You should narrow the scope of this question"
Having to grind achievements to be able to comment, like or dislike.
I used it for a year or so back in 2013?, went back to posting in forums like XDA developers, Codeguru and Reddit.
Just how is 25K new questions a month dead? Even if it gradually asymptomates to just 1K, the answers to them are enough to continue serving as a critical base layer of high-quality training data for LLMs.
Let's say for the sake of argument that 95% of humanity perished. Is humanity then dead? It isn't.
It just won't survive with 1k questions per month from the financial point of view. The hosting, the team, etc. Also, there is GitHub, where you can get high-quality answers directly from the authors of the component or library you are using and high-quality training data.
> June 2021: Stack Overflow sold for $1.8B to private equity investor, Prosus. In hindsight, the founders – Jeff Atwood and Joel Spolsky – sold with near-perfect timing, before terminal decline.
Hey, if it all crashes and burns, at least it’s the so-called smartest guys in the room going down with the ship. Just a bunch of VCs learning the hard way that they had no idea how to actually run or grow the company they bought. “Look at how well we optimized it!” Yeah — right into the ground.
Questions asked isn't the metric to declare it dead. What matters for financial viability is traffic and deals with search engines/indexes (used by LLMs)
For community viability: people will keep using it where LLMs fail. For new problems. It's still the place to go for undocumented workarounds.
Traffic and voting activity is certainly down but there is still immense value and new valuable questions are asked and answered there.
The article talks about the number of questions asked in SO, but there are no mentions of visualizations.
Feels natural that after 16 years of refinements, most normal questions are already there. I use it every week, but can count on one hand the number of questions I asked (0 through my account) over 12 years of having an account. ~All my questions were already asked.
I always considered stack overflow to be a Band-Aid placed on top of a mistake! The mistake was always poor documentation by the original system designes, and a policy of not allowing edits to the system design doc! And the more mistakes a documentation person makes, the larger the stack overflow corpus!
Now ChatGPT for SO is a Band-Aid on top of a Band-Aid on top of a mistake!
I really don't believe in the elitist policy to qualify for being able to answer stack overflow questions ... Whenever I have a better answer than all the existing ones stack overflow says I'm not qualified to answer so shut up! To hell with SO - I answer more questions at my company than anybody else and SO is run by elitst fools ...
This will have interesting implications for the LLMs as well, since SO is a wealth of training data. In my experience, LLMs are pretty useless when it comes to helping me with newer, faster-evolving, experimental tools and libraries, which is not surprising. But, if the SO community really atrophies to the point that a lot fewer people are bothering to answer questions, there won't be another centralized resource for answers. Perhaps that just means balkanized communities like random Slack channels will fill the gap, but those aren't search-indexed and I'd bet getting them all into training corpi won't be as easy either.
Maybe the future involves LLMs asking questions on something like SO when it routinely fumbles a particular topic. People could get paid to answer them and provide more training data. Who knows at this point
A sociological case study. Legit founders, a fruitful niche, immense value. Growth, politics, corporatization. They did so many things right, then so many things wrong.
If it were up to me, moderation would have been overhauled. But it wasn't up to me.
I have some minimally popular answers on SO and for years my "points" graph has notched its way upwards. It flatlined at the start of 2024 as people moved to AI.
The best thing about SO is seeing the competing solutions, the discussions, meaning with some discernment you can find that peer-reviewed high quality code snippet. Why would people prefer whatever the AI spits out?
Fortunately I see a few blips on SO so hopefully people are coming back now that the shine has worn off AI.
What is the value of SO to the world economy? Billions. Like the internet archive, it should be some sort of government funded (UN?) library
I'll be harsh, but they deserve it. The value of SO plummeted as documentation got better, disallowing open ended questions and discussions completely removed the value of talking to other humans. It was useful as junior and intermediate, but the value for a senior dropped to 0 due to the inability to ask what I would consider important.
I had to resort to reddit to ask those questions, which is ironic given the focus of SO.
Personally my usage of S.O. was significantly reduced just by sticking to the same stack and tools for the last 7 or 8 years and letting depth accumulate instead of always being in the midst of learning a new framework / language / whatever that necessitated googling how to do x in y or z.
And while that was happening VS Code started integrating MDN as well, so when I come across something I don't recognize I have a lot of extra information right at my fingertips anyway.
My day job includes moderation. I think I am more empathetic to the issues that moderators deal with on a day to day basis given that they are underpaid, under-appreciated, and overworked.
It's very difficult to scale a community to be both welcoming and productive. New users don't have the same context as existing ones. You find that norms and manners aren't transferred from one group to the next. So although that I noticed that SO started getting more strict from 2014 onward, I wouldn't know immediately what to do about the content quality issue.
My take is that, like most things, the medium of the old will be appreciated the way it wasn't in 2014. As the Brian Eno quote goes: "Whatever you now find weird, ugly, uncomfortable and nasty about a new medium will surely become its signature." People will yearn for the human forums the same way they did years past when people tire of the LLM slop. (If they do.)
I'm not very sympathetic to moderators. Their job is to make some forum contain only what they want to hear from the (unpaid) people providing the content. If that's what you want, start a blog.
In any sort of IRL community, if some person decided they had the authority to prevent people from answering others' questions, unilaterally decide people aren't welcome, etc. everyone would think they had gone insane.
However there is the problem where any unmoderated public forum will eventually contain only scam artists and Nazis. Moderators address the Nazi spam problem, but now you have the moderator problem. It's been a mental side project of mine lately to find ways to solve the moderator problem, which I can't talk too much about as there is money to be made by doing so.
What happens when the forum gets overrun by Nazi spam? We'll install a set of moderators who remove it. But what happens when the moderators start removing useful information? We simply unleash wave after wave of Chinese needle snakes. They'll wipe out the moderators. But aren't the snakes even worse? Yes, but we're prepared for that. We've lined up a fabulous type of gorilla that thrives on snake meat. Then we're stuck with gorillas! No, that's the beautiful part. When wintertime rolls around, the gorillas simply freeze to death.
Yes this is the really hard problem of human organization on the internet. You need mods to keep the quality gate up. But eventually mods feed off their own power and alienate the community. Maybe some combination of human and algorithmic moderation can help.
I don't miss spending hours trying to find the right config to make things work and trying random answers on SO. Overall the dev world is a better place now and MANY dev hours are being saved. Plus, AI can now learn any new framework/language directly from docs. Even obscure ones, just pass all the .md files to gemini and ask!
I have mixed feeling about this.
It has helped me in the past but yet, I could not reply nor post anything back to help others when I knew the solution because of the way how it works.
To make matters worse while working in IT, I worked with a guy that didn't know anything, if there was no SO post about the problem, the guy couldn't fix the problem.
I have been using Perplexity AI and it has been awesome, and it does provide all the sources it used making it easy to cross check the answers. It has helped me to speed my python learning curve, I am not using search engine anymore, and SO has the problems mentioned above so I have zero interest in using it.
Also, the website layout is a mess, I have to use uBlock Origin with a ton of element picker to stop loading half of its crappy.
I'm curious if Stack Overflow was a good resource for human learning? As in, run into a problem, look through other related questions and answers, check the documentation, struggle for awhile learning and figuring out how to frame and pose the question, struggle some more while you wait to get an answer. I kind of find LLMs "too easy" and can distinctly feel myself "not learning", not the way I used to, but after all I am getting older.
I'm pretty sure you can get a stack overflowy experience out of an llm with the right system prompt, but the human factor might not be the same. Not wanting to be berated by others on the internet is maybe underrated as a motivational tool. How are we going to get that back?
Not to mention the helping and teaching other people aspect, which feels good, it feels like there isn't as much of that now. Or maybe it has just moved away onto Discord etc, but GitHub definitely doesn't feel like it has that aspect, if you don't also personally know other contributors and reviewers. Curious if others feel this way.
How will ChatGPT learn the next computer language now?
From my perspective LLMs shine in programming languages because they (languages) are invented by people to be more formal so they are more predictable. Even for syntactically different programming languages there are similarities we don't see with our own eyes, but the training has no problem detecting. My favorite example is when I tried to find how to represent an int array for reflection in Kotlin. Ask ChatGPT or a sufficiently large model about this, the answer will be IntArray::class.java , but try to find the exact line with google - few or nothing, more on github search in the sources, still not much. So LLMs "detected" the system of making type signatures in Java/Kotlin and were able to successfully predict because the rules are consistent. In human language it also works, but to a lesser degree, so if you give it a verb/subject pair that makes sense and ask for the equivalent ones, you will get some that still make sense but are literally absent in the full corpus of the web.
How will there be new programming languages after people stop practicing programming and the skills decay because an AI is performing the deliberate practice?
We are in the future, nothing new needed here.
Of course it's dead, that's what you get for harassing users who try to ask questions.
Are there any similar articles on the state of the rest of the Stack Exchange network? There are many, many other SE sites that have nothing to do programming and are often less amenable to being answered via LLM.
> The question seems to be when Stack Overflow will wind down operations, or the owner sells the site for comparative pennies, not if it will happen.
I see the latter option, but the former? SO, at least judging by their hardware posts, was running on nine servers two years ago [1]. That's barely anything in costs - electricity, uplink and occasional rotation of the hardware, but probably a single person working a decent job can afford to run the entire hardware for the site.
Truly shows how far a tight budget can go when you don't waste untold amounts of money and energy on layers upon layers of complexity.
> I'm sure we'll see spaces where developers hang out and help each other continue to be popular – whether they are in the form of Discord servers, WhatsApp or Telegram groups, or something else.
Yeah fuuuuck that. It's so annoying that everyone and their dog moved to these walled gardens. Google can't pierce them, unlike IRC of ye olde days where it was common to let a bouncer publish logs, WA/Telegram come with privacy risks and Discord is a hellscape.
[1] https://news.ycombinator.com/item?id=34950843
Hardware is usually just a small part of running a website.
Set up StackOverflow according to "cloud native" recommendations and whoops, that's gonna be quite the bill, which is my point. You'll have Cloudfront for global load balancing, ELB to provide a bridge between Cloudfront and Kubernetes, EFS for storage, RDS for the database, EKS and EC2 for compute, ElastiCache for KV cache, add in CodeBuild for build/deploy pipelines... AWS has quite the hefty overhead.
1 reply →
To Ai paraphrase Obi-Wan: I felt a great disturbance in the Force — as if millions of developer voices from the future cried out in confusion and despair, and were suddenly silenced. I fear something terrible has happened to the future of our craft.
In short before we Ctrl+Copy our way to StackOverflow or Forums or IRC and we got collection of responses between good and ugly. But made us think and read or talk to others.
Now we Ctrl+Copy into LLM into a room of hell.
If LLM function is useful but don't get addicted to like honey.
my business has a similar revenue curve, and of course LLMs are a big cause, but it's more to do with me being distracted from moving with the times. I just didn't fancy the fight and saw enough value in the curve to let it follow ranged decline
I'm assuming the owners of stack felt similar? Don't know anything about them so could be easily wrong
It will neither be the first or last site destroyed by hamfisted power tripping overmoderation. And they do it in the worst possible way - Leaving the question up, but preventing anyone from answering it. So, the default experience with SO becomes finding your question already asked and unanswered.
Still, it may find its place as a last resort.
It looks like Stackoverflow was already in a decline even before Covid.
I wonder what developers started using during that time.
They still used it, they just didn't contribute. It's not that StackOverflow is useless, but rather that it already answers most questions.
That, and genuinely new questions get aggressively shut down.
Given the speed at which tech evolves, I do not see how that can be true.
I think it may answer most questions from five years ago, but I would phrase it as “it has answers for most questions, but many are outdated and there’s no easy way to tell whether an answer is outdated”
1 reply →
The graph shows monthly questions asked, not monthly questions answered. It makes total sense that the slope of that graph would be much higher earlier in the life of the site; as time goes on, many common questions will have already been answered.
Discussion on Meta SO: https://meta.stackoverflow.com/questions/433864/do-you-agree...
Stackoverflow answers kind of provided a source of truth by being confirmed / upvoted by people trying them out. We're completely abandoning this medium (on which LLMs are trained), even as technologies keep changing. Perhaps coding agents should start posting to stackoverflow too...
Didn’t SamA paid out StackOverflow already?[0] So as far as the owners go they are doing fine.
[0] https://openai.com/index/api-partnership-with-stack-overflow...
You don't start and run StackOverflow because of the money. You do it because you really care. Running niche communities isn't a lucrative path, and (as we're seeing now) they can be quite fickle.
Additionally, it's not just the owners... it's the millions of people who contributed.
Per the post, it seems as if SO was going to going to slide beneath the surface anyway but that LLMs were one of a number of factors that maybe accelerated the process by maybe a few years. Without having deep knowledge of the area, this feels about right.
Kind of a catch 22 isn't it? LLMs trained off stack overflow data. Without new data for them to train from, what will happen to the quality of LLMs in the future?
How will technology advance without research sharing?
I wonder if they could survive if they added a mode like phind.com wrapper - where each question stores LLM answers as answers. So (SO!) many more content and would people reason to still use it.
An old fact about StackOverflow is that there’s a disproportionate amount of views compared to questions asked.
Something like 1000x more views than posts/comments… I wonder if that statistic has changed over the years?…
Stack Overflow simulator is a functional museum of Stack Overflow... using AI
https://sosimulator.xyz
Everyone is blaming AI, and it's undoubtedly a factor.
But also, the culture of Stack Overflow has changed significantly over the years. It used to be a place where anyone could ask a question and get help with a problem ... and it was amazing.
Today, you're far more likely to have your question downvoted, flagged as a duplicate (of an unrelated question), or attacked in the comments by overzealous responders (and once that happens, good luck on actually getting help). Your odds of actually getting help on the site are only a fraction of what they once were.
And I'm not just saying this as some SO newbie: I've been using the site since beta! As someone who has used it that long, the change in quality is undeniable.
Can you link to questions that you feel were unfairly derided (not necessarily closed)?
As someone who has extensive experience modding an niche SE I see this sentiment quite often but honestly, the people making these complaints are just arrogant _and_ wrong about the topic they needed help with.
stackoverflow started declining long before AI. Their ban-happy moderators were to blame.
AI is great but when it gets stuck I go to stack overflow. Happened to me yesterday
the stock market looked like that chart in 2008 and look where we are now
unless LLMs can be instantly trained on all new software frameworks and languages that come out, im not worried stackoverflow will still have a place
All things have their time.
Some get superseded.
Others accelerate their decline through self-foot-shooting and/or enshittification.
Stack Overflow's journey into obscurity is via a mix of private equity indifference, better docs elsewhere, and a lack of leadership over its moderators. It was in decline long before LLMs.
It is not a new story - but it does help map out the modes of platform senescence.
Wasn't the first; won't be the last.
Or: all programming problems have been solved! ;-)
No wonder, the CEO basically said they'll use the free labor of love of all member devs to sell training data for AI in order to replace such devs so why would anyone keep contributing there?
They really only have themselves to blame. Yeah AI has massively accelerated the decline but I think it's mainly provided an option that isn't so frustrating to use. ChatGPT never says your question is unclear or off topic.
I think if they had actually fixed moderation they may have had a chance of surviving, but I think they got trapped by relying on volunteer moderators who thought that it was good that so many valid questions were closed.
They did actually make some attempts to fix things, e.g. I remember one suggestion from the company that users could reopen a closed question at least once (which is a great thing to try!) and mods downvoted that to hell so they chickened out.
Definitely some shadenfreude, and I say that as someone with 100k reputation.
> ChatGPT never says your question is unclear or off topic.
Yes, because ChatGPT doesn't care about publishing your question and its answer with the explicit intent of enabling other people to find it later with a search engine. It has no mandate to organize content or care about the quality of its content. It has no reason to care about the topic. Other people will not read your prompt (which doesn't even need to be trying to ask a question at all), so there's no reason to care about whether other people can understand the question - let alone figure out whether it's something they also want to ask.
> I remember one suggestion from the company that users could reopen a closed question at least once (which is a great thing to try!) and mods downvoted that to hell so they chickened out.
Questions can be reopened, and this has always been the case. What is not allowed is for the OP to reopen a question unilaterally, because that defeats the entire purpose of question closure. The entire point is that when you come to Stack Overflow, you aren't the one who gets to decide what the standards for questions are, or judge whether your question meets those standards. Because if you were, the standards would effectively not exist.
The people you think of as "mods" are overwhelmingly not.
> I say that as someone with 100k reputation.
I have found that reputation scores say very little about whether one understands the purpose of the site or how it's designed to work. Famously, a user with over a million reputation once went on a spree of violations (https://meta.stackoverflow.com/questions/430072) of the best-publicized site policy ever to appear on the meta site (https://meta.stackoverflow.com/questions/421831/). Many users with 100k+ reputation and/or a 10+ year history have likely never looked at the meta site. (See https://meta.stackoverflow.com/questions/427224 for someone who first posted on meta after 14 years.)
> It has no mandate to organize content or care about the quality of its content.
That's fundamentally where SO went wrong. The mods think they're building programming Wikipedia, but normal users are using it as a Q&A forum.
> Questions can be reopened, and this has always been the case. What is not allowed is for the OP to reopen a question unilaterally, because that defeats the entire purpose of question closure.
No it doesn't. It's far too easy to close a question and they very rarely get reopened. The suggestion was that users only get to unilaterally reopen it once.
> I have found that reputation scores say very little about whether one understands the purpose of the site or how it's designed to work.
See this is exactly the problem. You have the mods' view of how it is supposed to work, but that isn't how people want it to work!
Mods want it to be this highly curated reference site where only perfectly written questions that exactly fit an FAQ style of questioning. They don't care that that makes it useless (or at least extremely frustrating) as a Q&A forum.
3 replies →
I wonder how the other -overflows are doing?
I suspect there will always be a place for SO-esque sites, but it will shift to be primarily for creating high quality data for ingestion by LLMs.
This is the absolute worst possible outcome! Humans toiling just to produce content to train AI on
What a miserable future to look forward to
Even if you were paid to make high quality answers?
> This tallies with my memory of feeling that site moderators had gone on a power trip by closing legitimate questions.
Respectfully: outsiders like the author of this piece are not the ones entitled to decide whether a question is "legitimate", or "valid" (another term I see used all the time by people who have no understanding either of Stack Overflow's standards or its goals).
Reference reading:
What is Stack Overflow’s goal? (https://meta.stackoverflow.com/questions/254770)
How much research effort is expected of Stack Overflow users? (https://meta.stackoverflow.com/questions/261592)
Question Close Reasons - Definitions and Guidance (https://meta.stackoverflow.com/questions/417476)
How long should we wait for a poster to clarify a question before closing? (https://meta.stackoverflow.com/questions/260263)
A satirical answer to "The rudeness on Stack Overflow is too damn high" (https://meta.stackoverflow.com/questions/262791/_/309018#309...)
What is the point of closing questions for details and clarity, debugging details, needs more focus, or very low quality? (https://meta.stackoverflow.com/questions/405519)
Why should I help close "bad" questions that I think are valid, instead of helping the OP with an answer? (https://meta.stackoverflow.com/questions/429808)
Why is the rate of positively scoring questions and answers steadily declining? (https://meta.stackoverflow.com/questions/393032)
When is Stack Overflow going to stop demonizing the quality-concerned users who have made the site a success? (https://meta.stackoverflow.com/questions/366858)
Is ChatGPT and LLM killing Stack Overflow (https://meta.stackoverflow.com/questions/430994)
[flagged]
Interestingly, the site stopped growing at about the same time when the "fun killers" [0] took hand of it. Notably, when they deleted the all-time highest voting question "New programming jargon you coined?".
This blatantly undemocratic and destructive behavior was of course duely punished by the (former) users of the site.
[0] https://blog.codinghorror.com/new-programming-jargon/
All the "overflow/exchange" sites suck, not just the stack overflow. Too many questions being closed. I had a question closed in mathexchange because, I surmise, it was "too obvious" even though easier questions were asked as recently as 2021, even much more elementary ones. Moreover, my specific question had never been asked there before, and the point of math exchange is to ask easier questions, compared to mathoverflow.
But it's also possible it's pivoting to a Wikipedia-like model where it becomes a repository for answers, and less about contributions. In which case, this is not the same as it dying. As seen with Wikipedia, it can still get a lot of traffic and revenue even if few people contribute to it anymore.