Summarizer

There was definitely a bit of Pournelle's Iron Law of Bureaucracy [0] at work. I worked there during a lot of the drama you allude to, and... It sucked, for everyone. But also... For most of SO's history, the vast majority of visitors (and this questions, answers) came in via Google. Not "search engines"; Google. This was pretty much baked in right at the start, and it effectively served as the site's primary user interface for years. And it worked. It worked pretty well! Until it didn't. At some point, Google started surfacing fewer "tried and true" Q&A examples and more unanswered, poorly-answered or moderated examples. This broke the fundamental assumption that sat behind SO's moderation - that curating a smaller set of posts was preferable to encouraging more, and newer. Suddenly, Google wasn't a very good UI for SO anymore. ...and SO didn't really have a fallback. Heck, for a while during this period they actually stopped showing questions on their homepage unless you were already logged in; the core assumption remained that there was a never-ending spring of new people feeding the thing. LLMs weren't the start of the problem, they were the end - the final wake-up call. I don't know that a site like SO can exist without the old Google, the old Internet; it is a product of all that, in the same way that mass-market TV shows were a product of 20th-century broadcast technology, or trade paperbacks of a particular intersection of printing tech and reading habits. [0]: https://www.jerrypournelle.com/reports/jerryp/iron.html

> the core assumption remained that there was a never-ending spring of new people feeding the thing. Hi Shog, hope you're doing well! Just thought this bit was insightful; I can fully believe this was the idea and the motivating factor for a lot of the decisions made seemingly in a vacuum (from the outside). How much do you think Area51 and the push for the SE network rather than sticking with the Big Three affected things? I always got the impression that they tried to scale into places that ultimately attracted too much noise and overestimated the willingness of (community) moderators to effectively work for free for them to take on the wave of less technical/principled users.

Thinking they didn't keep up with the times or that they should've made changes is perfectly fine. It's the vitriol in some of the comments here I really can't stand. As for me, I also don't answer much anymore. But not sure if it's due to the community or frankly because most low hanging fruits are gone. Still sometimes visit, though. Even for thing's an LLM can answer, because finding it on SO takes me 2 seconds but waiting for the LLM to write a novella about the wrong thing often takes longer.

> They were primarily trying to make google searches more effective for the broader internet This is mostly how I engaged with SO for a long, long time. I think it’s a testament to SO’s curation of answers that I didn’t ask almost any questions for like 5+ years after starting programming

> So instead, it's considered preferable that the best possible answer never be allowed to emerge, unless by sheer coincidence the best answer just happened to be the one that was accepted the first time the question was asked, several years ago. What? No. The canonical target isn't closed. So go write the new answer there . The answer acceptance mark is basically irrelevant, and the feature ill-conceived. Except usually there are dozens of answers already; the best possible answer has emerged; and people keep writing redundant nonsense for the street cred of having an answer on a popular Stack Overflow question. > The verdict seems to have more-or-less delivered itself. We do not care that people don't want to come and ask new questions. There are already way, way too many questions for the site's purpose. The policy is aimed at something that you don't care about. The result is a "verdict" we don't care about.

This doesn't mean that it's over for SO. It just means we'll probably trend towards more quality over quantity. Measuring SO's success by measuring number of questions asked is like measuring code quality by lines of code. Eventually SO would trend down simply by advancements of search technology helping users find existing answers rather than asking new ones. It just so happened that AI advanced made it even better (in terms of not having to need to ask redundant questions).

"I suspect that the gradual decline, beginning around 2016, is due to growth in a number of other sources of answers." I think at least one other reason is that a lot of the questions were already posted. There are only so many questions of interest, until a popular new technology comes along. And if you look at mathoverflow (which wouldnt have the constant shocks from new technologies) the trend is pretty stable...until right around 2022. And even since then, the dropoff isn't nearly so dramatic. https://data.stackexchange.com/mathoverflow/query/edit/19272...

> The fundamental value proposition of SO is getting an answer to a question; if you can the same answer faster, you don't need SO. Plus they might find the answer on SO without asking a new question - You probably would expect the # of new questions to peak or plateau even if the site wasn't dying, due to the accumulation of already-answered questions.

Also: the bigger the corpus of already answered questions, it’s more likely that you can just look up an answer instead of asking. Eventually SO becomes a site exclusively for lurkers instead of a platform for active participation

The JavaScript ecosystem has mostly stabilised. React is 12 years old for example.

I don't know what your experience has been, but I do feel that at some point you will find yourself on or beyond SO's "knowledge frontier". The questions you land on will be unanswered or have equally confused replies; or you might be the one who's asking a question instead. I've "paid back" by leaving a high quality response on unanswered SO questions that I've had to figure out myself, but it felt quite thankless since even the original poster would disappear, and anyone who found my answer from Google wouldn't be able to give me an upvote either.

Right. I often end up on Stack Exchange when researching various engineering-related topics, and I'm always blown away by how incredibly toxic the threads are. We get small glimpses of that on HN, but it was absolutely out of control on Stack Exchange. At the same time, I think there was another factor: at some point, the corpus of answered questions has grown to a point where you no longer needed to ask, because by default, Google would get you to the answer page. LLMs were just a cherry on top.

I agree there was some natural slow down as the corpus grew - the obvious questions were answered. But if the community was healthy, that should not have caused growth to stop. New technologies get created all the time, each starting with zero SO questions. (Or Google releases v2.0 which invalidates all answers written about v1.) SO just stopped being fun for me. I wish more systems would use their point systems though.

People overestimate the impact of toxicity on number of monthly questions. The initial growth was due to missing answers. After some time there is a saturation point where all basic questions are already answered and can be found via Google. If you ask them again they are marked as dups.

That would be true if no new technologies were created every year (even more often).

There are new technologies, but if you look at the most viewed questions, they will be about Python, JS, Java, C, and C++ without libraries.

Use of GPT3 among programmers started 2021 with GitHub Copilot which preceded ChatGPT. I agree the toxic moderation (and tone-deaf ownership!) initiated the slower decline earlier that then turned into the LLM landslide. Tbf SO also suffered from its own success as a knowledgebase where the easy pickings were long gone by then.

Actual analysts here that have looked at this graph like... a lot, so let me contextualize certain themes that tend to crop up from these: - The reduction of questions over time is asymptomatic of SO. When you have a library of every question asked, at some point, you asked most of the easy questions. Have a novel question becomes hard. - This graph is using the Posts table, not PostsWithDeleted. So, it only tells you of the questions that survived at this point in time, this [0] is the actual graph which while describes a curve that shows the same behavior, it's more "accurate" of the actual post creation. - This is actually a Good Thing™. For years most of the questions went unanswered, non-voted, non-commented, just because there was too many questions happening all the time. So the general trend is not something that the SO community needs to do anything about. Almost 20% of every question asked is marked as duplicate. If people searched... better™ they wouldn't ask as many questions, and so everyone else had more bandwidth to deal with the rest. - There has been a shift in help desk style of request, where people starting to prefer discord and such to get answers. This is actually a bad thing because that means that the knowledge isn't public nor indexed by the world. So, information becomes harder to find, and you need to break it free from silos. - The site, or more accurately, the library will never die. All the information is published in complete archives that anyone can replicate and restart if the company goes under or goes evil. So, yeah, such concerns, while appreciated, are easily addressed. At worst, you would be losing a month or two of data. [0]: https://data.stackexchange.com/stackoverflow/query/edit/1926...

> When you have a library of every question asked, at some point, you asked most of the easy questions. Have a novel question becomes hard This would be true if programming were a static field, but given that new programming languages/frameworks/technologies/techniques/etc. are constantly coming out and evolving, that argument doesn't make sense.

Programming is not a static field in the answers side, but it's in the question side. "How to print characters on a terminal with python?" is the same problem today as it was 25 years ago. The answer changed but the problem remained. That's what people saying that programming isn't static is missing: the problem space grows significantly slower than the solution space.

But you’re supposed to replace Python with a new language that hasn’t been asked about.

The various admonitions to publish to a personal blog, while encouraging, don't really get at the 0xfaded's request which I'd summarize as follows: With no one asking questions these technical questions publicly, where, how and on what public platform will technical people find the problems that need solving so they can exercise their creativity for the benefit of all?

It's also was a bit frustrating for me to answer. There was time when I wanted to contribute, but questions that I could answer were very primitive and there were so many people eager to post their answer that it demotivated me and I quickly stopped doing that. Honestly there are too many users and most of them know enough to answer these questions. So participating as "answerer" wasn't fun for me.

I spent a lot of time answering rather primitive questions, but since it was on a narrow topic (Logstash, part of the ELK stack), there wasn't many other people eager to post answers. Though it often ended up with the same type of issues, not necessarily duplicates, but similar enough that I got bored with it.

The graph is scary, but I think it's conflating two things: 1. Newbies asking badly written basic questions, barely allowed to stay, and answered by hungry users trying to farm points, never to be re-read again. This used to be the vast majority of SO questions by number. 2. Experiencied users facing a novel problem, asking questions that will be the primary search result for years to come. It's #1 that's being canibalized by LLM's, and I think that's good for users. But #2 really has nowhere else to go; ChatGPT won't help you when all you have is a confusing error message caused by the confluence of three different bugs between your code, the platform, and an outdated dependency. And LLMs will need training data for the new tools and bugs that are coming out.

The newbies vastly outnumber the experienced people (in every discipline), and have more to ask per-capita, and are worse at asking it. Category 2 is much smaller. The volume of Stack Overflow was never going to be sustainable and was not reasonably reflective of its goals. We are talking about a site that has accumulated more than three times as many questions as there are articles on Wikipedia. Even though the scope is "programming languages" as compared to "literally anything that is notable". But there are other places people can go, such as https://software.codidact.com (fd: I am a moderator there).

While AI might have amplified the end, the drop-off preceded significant AI usage for coding. So some possible reasons: - Success: all the basic questions were answered, and the complex questions are hard to ask. - Ownership: In its heyday, projects used SoF for their support channel because it meant they don't have to answer twice. Now projects prefer to isolate dependencies to github and not lose control over messaging to over-eager users. - Incentives: Good SoF karma was a distinguishing feature in employment searches. Now it wouldn't make a difference, and is viewed as being too easy to scam - Demand: Fewer new projects. We're past the days of Javascript and devops churn. - Community: tight job markets make people less community-oriented Some non-reasons: - Competition (aside from AI at the end): SoF pretty much killed the competition in that niche (kind of like craigslist).

> - Success: all the basic questions were answered, and the complex questions are hard to ask. I think this is one major factor that is not getting enough consideration in this comment thread. By 2018-2020, it felt like the number of times that someone else had already asked the question had increased to the point that there was no reason to bother asking it. Google also continued to do a better and better job of surfacing the right StackOverflow thread, even if the SO search didn't. In 2012 you might search Google, not find what you needed, go to StackOverflow, search and have no better luck, then make a post (and get flamed for it being a frequently-asked question but you were phrasing yours in a different / incorrect way and didn't find the "real" answer). In 2017, you would search Google and the relevant StackOverflow thread would be in the top few results, so you wouldn't need to post and ask. In 2020, Google's "rich snippets" were showing you the quick answers in the screen real estate that is now used by the AI Overview answers, and those often times had surfaced some info taken from StackOverflow. And then, at the very end of 2022, ChatGPT came along and effectively acted as the StackOverflow search that you always wanted - you could phrase your question as poorly as you want, no one would flame you, and you'd get some semblance of the correct answer (at least for simple questions). I think StackOverflow was ultimately a victim of it's own success. Most of the questions that would be asked by your normal "question asker" type of user were eventually "solved" and it was just a matter of how easy it was to find them. Google, ChatGPT, "AI Overviews", Claude Code, etc have simply made finding those long-answered questions much easier, as well as answering all of the "new" questions that could be posed - and without all of the drama and hassle of dealing with a human-moderated site.

The volume of basic questions is unlimited. There are new technologies every year.

Not sure. As software becomes a commodity I can see the "old school" like tech slowing down (e.g. programming languages, frameworks frontend and backend, etc). The need for a better programming language is less now since LLM's are the ones writing code anyway more so these days - the pain isn't felt necessarily by the writer of the code to be more concise/expressive. The ones that do come out will probably have more specific communities for them (e.g. AI)

Web has been solved for a decade imo.

I used to joke that when SO goes under, I will move professions. The joke came from my experience of how many common issues in technology could not be solved with knowledge found via a search engine. I don’t see that niche as gone, so I wonder what is satisfying that requirement such that new questions do not show up at SO?

A decline in number of questions asked can also be because most people's questions are already answered in the database. How would you query this for post views over time?

I think one of the phenomenon that people haven't mentioned is that the question space was heavily colonized by 2016. I was one of the top 30 or 50 answerers for the SVG tag on SO, and I found that the question flow started to degrade around 2016, because so many of the questions asked had been answered (and answered well) already.

Everyone is saying LLMs did this site in, but what if we just asked all the questions already? We should be celebrating how we solved programming!

Interesting timing. I just analyzed TabNews (Brazilian dev community) and ~50% of 2025 posts mention AI/LLMs. The shift is real. The 2014 peak is telling. That's before LLMs, before the worst toxicity complaints. Feels like natural saturation, most common questions were already answered. My bet, LLMs accelerated the decline but didn't cause it. They just made finding those existing answers frictionless.

So it seems all the questions have now been answered– Great!

Surprising to see it bottom out so hard. I imagine at least some of the leveling off could be due to question saturation. If duplicates are culled (earnestly or overzealously) then there will be a point where most of the low hanging fruit is picked.

Maybe the average question will be more "high level" now that all simple questions are answered by LLMs ?

The brilliance of StackOverflow was in being the place to find out how to do tricky workarounds for functionality that either wasn't documented or was buggy such that workarounds were needed to make it actually work. Software quality is now generally a bit better than it was in 2010, but that need is ultimately still there.