Here is the irony that keeps bothering me.
Stack Overflow helped train the AI models that are now making Stack Overflow irrelevant.
The models le...
For further actions, you may consider blocking this person and/or reporting abuse
The piece you're missing is the evaluator. The privatization problem isn't just that solutions disappear into private chat logs. it's that even when knowledge gets captured, there's no quality signal deciding what's worth keeping. You end up with a different kind of graveyard: everything preserved, nothing promoted.
@leob pointed you to Foundation. That's what I've been building ā a federated knowledge system on Cloudflare Workers that captures insights from conversations and runs them through an evaluation layer before they reach semantic memory. The evaluator is the part most memory systems skip. I wrote about it here if the architecture is useful: Building the Evaluator.
The governance question you raised at the end is the one I don't have a clean answer to either. ActivityPub federation is part of the architecture ā the knowledge stays distributed rather than owned but that solves the ownership problem, not the incentive problem.
Thanks! This is definitely on my reading list
Bonus! Democracy is the absolute worst form we know about to have quality bubble to the top! Beta Max anyone ...?
The internet is getting noisier and the human part seems quieter in comparison and humans are actually shifting activities, contributing less to StackOverflow etc. posting AI-assisted slop on DEV, LinkedIn and other social media instead. Fair enough. The internet wasn't "dead" before StackOverflow and it's still alive now.
Information intoxication was a problem before AI with echo chambers amplifying prejudice and misinformation. Algorithms and AI made that worse. If they're eating up the resources they feed on, their growth might slow down, hallucinations might increase despite despreate "guardrails" efforts.
might fail for those exact reasons, unless assisted by human curation. Anyone who has contributed to curating StackOverflow queues like first questions and late answers knows there have been sloppy and random low-quality contributions before. A lot. The current calm there doesn't feel that bad. Noisy low-quality AI guesswork might not be helpful after all.
If the models we have are already good, why not just optimize them, rebuild them with slightly updated training data and let them check for reputable sources before answering? We still have documenation. We have MDN, Wikipedia, CSS-Tricks, man pages, forums, and a dedicated StackExchange community keeping the old internet alive albeit with reduced energy that might be enough to update what really matters.
On the other side, I think this is the best time for professionals to get into blogging and sharing information about their personal projects. It is easier than ever to host a simple Cloudflare static site to turn your markdown notes into public knowledge. Tools like Obsidian and journal entries will be invaluable to both real people and AI agents. Most forum spaces have unfortunately been dead for years before AI came into the picture due to server costs and no clear monetization path. Maybe agents could tip website owners for good info, something like how Brave tokens worked.
It does make me wonder how many real people would end up reading these sites. It is already a challenge to get developers to read documentation instead of relying on a primer video or agent explanation.
Mainly I hope that as we continue through this timeline that more and more developers make their agent driven solutions public.
Not sure who still reads forums, but when it comes to documentation, I think that ship has already sailed.
As a tech lead and part of the Developer Experience team at my org, I know how important it used to be for developers to actually read the docs we wrote for the tools we built for them. We tried every trick in the book to make those docs searchable, approachable, and easy to follow with a clear flow.
These days though, we mostly target AI agents. We still write docs for humans (for now), but we also add an AI agent version or tweak the existing docs so they work better for agents. For example, agents struggle with interactive docs.
We also built a catalog of all our tools with LLM friendly documentation and distribute it through an internal MCP we created.
This helps developers actually become 10x devs, with their agents powered by our docs and tools.
I wrote a blog about this if youāre interested
debuggr.io/ai-agents-org-context
This was an interesting read, would you say your agent flavored documentation ended up looking like a Zettelkasten system? Our organization is very new to leveraging AI but we do have various levels of Microsoft flavored usages. We have decent documentation for the product, but trying to get information out of it made it clear the context was too small to get a valuable answer.
Thanks for introducing me to the Zettelkasten system. That was new to me and really interesting. I watched an explainer video about it.
Not totally sure if our solution is similar to that approach, but I do see a few overlapping ideas.
Basically, we have a catalog that lists all of our products (dev tools and services) with a short description and some basic metadata like the repo name and path. Our MCP server has a tool the AI agent can use to fetch this catalog on demand.
The MCP also has a tool that can fetch content from repositories. It can pull repo metadata from the catalog and use it to trigger the content fetching tool. That tool can list files and also retrieve file content, like the README or even source code if needed.
We also have architecture docs and documents that describe how different tools or services connect to each other. These docs are also accessible through MCP server tools.
This is just a very high level overview of our setup.
So in practice we have two types of "notes" (documentation):
Hope that makes sense
Reminds me of the bookstore issue. People go into a bookstore to learn about books of interest. Books in a bookstore are expensive, so they order from Amazon. Amazon puts the book stores out of business. Now, how do people learn about books?
Very interesting comparison!
Really interesting post! I feel the same way about how AI is changing the knowledge-sharing space. AI gives quick answers, but it loses that back-and-forth problem-solving that we used to get from places like Stack Overflow. The idea of AI sharing solutions is cool, but like you said, there are a lot of challenges with accountability and how to keep it useful for everyone. Definitely something to think about moving forward!
Indeed! The impact is much bigger but isolated.
Really interesting point about the 'privatization of knowledge' at scale. I've definitely noticed myself reaching for an AI assistant first instead of searching StackOverflow lately. Do you think we'll see a shift where AI agents themselves start 'publishing' their own findings back to the commons to prevent this data rot?
Thanks! I do think weāll see a shift and that AI agents will eventually start sharing knowledge in some way. But for that to happen, the problem needs more attention so more people can come together and work on solving it as well as possible.
"If you know of work being done here, I would genuinely love to hear about it" - yes, I do know about people working on this issue - I think this will be right up your alley:
dev.to/the-foundation
dev.to/the-foundation/welcome-to-t...
These folks are trying to address your (very valid) concern, or something really close to it - what struck me was this section from your article:
"One rough direction worth thinking about ... What if AI agents were first-class participants in a public knowledge platform?"
This is not exactly the approach they've chosen, but it does come close (I think) ...
Thanks for that! Will definitely take a look š
The privatization point hits hard ā weāve seen the same pattern where solutions live in internal Slack threads instead of public posts. The recursive problem of models trained on a shrinking public commons is something I keep thinking about too.
Totally. Slack threads have always been messy, and getting useful results from Slack search has been a pain point for most orgs
I'm assuming you're using that word to mean it's something not talked about much (I say, trying to be hip with the new kids' argot) rather than actually underrated.
I disagree though - we've been talking about this, quite loudly in some cases, for at least a couple of years now. Everyone's aware of the problem, the issue isn't that it's a novel problem. The issue is perennial - it's that nobody who profits from AI cares, because it's not going to give their shareholders more money in the short term.
Something I've noticed working on dev tools ā the "debugging narrative" is what's really disappearing. Not just the final answer, but the journey of figuring stuff out.
I used to find SO threads where someone tried 3 wrong approaches before landing on the fix, and those wrong turns were often more useful than the answer itself. They taught you how to think about the problem.
Now I catch myself asking an AI, getting a working solution, and moving on without ever understanding why my first instinct was wrong. The solution works but the learning didn't happen.
The MCP approach you mentioned is interesting though. I wonder if something like structured "debug logs" that agents could publish would work ā not full Q&A threads, but lightweight traces of what was tried and why it failed. Lower friction than writing a blog post, but still captures that reasoning chain.
Curious if anyone's experimented with making their AI chat logs public in a useful way?
What you are talking about is a unique human experience. In a world where coding is flashy and thereās a lot of AI-generated content, itās not about whatās in a GitHub project or a blog - itās about those unique edge cases that humans are dealing with that AI canāt replicate. Do you think there will be a "Data Union" model to protect high-quality human data from bots?
There's another layer here: even when people want to share, the friction has changed. Writing a good Stack Overflow answer takes time and effort. Getting an answer from AI is instant. The effort-to reward ratio for public contribution has shifted dramatically, and not in favor of the commons. The 'culture of sharing' you mention wasn't just altruism it was also the most efficient way to get answers at that time. Now it's not.
If it were actually intelligent, it wouldn't need feeding - especially given what it already 'knows'
Iām not saying LLMs are intelligent beings. But what intelligent being can become or stay intelligent without being fed data in some way?
LLMs already have probably more information than any human being and yet can be outperformed by humans in many things. They're not intelligent, and feeding them more data will not fix that.
A change of approach is what's needed, not more data.
yep, but the seniors still post because of credibility/consulting leverage. it's the mid-tier that's ghosted. that's what actually kills the knowledge chain for juniors.
Love the practical approach. Theory is nice, but hands-on examples are better.