Hacker News Reader

Comment by hirundo | Hacker News Reader

This is a large decline. It does not correlate well with the rise of LLMs, which have shifted my own habits away from Stack Overflow recently. Looking through the graphs, the most dramatic shift was the large drop in new visits in the first half of last year, by two thirds or more. That seems late to be a Covid effect. There were some political controversies, but it would surprise me if they had such a large effect.

What accounts for this?

View on HN

> What accounts for this?

IMHO people have tired of the smug dismissive assholes that dominate SO and its its insipid gamification and they have, gradually, found alternatives.

For some, it's more welcoming and knowledgeable communities in github issues+discussion.

For others it's special-purpose, interactive forums that provide more guidance and non-hostile support for users of certain platforms/tools (eg Posit Community).

For many, copilot has been fulfilling that need, Does it always give "the correct" answer? No. Does it provide a sketch of a solution that gets you half-way there when dealing with tedious humdrum stuff that you just forgot because it's so boring? YES.

For yet others it's just plain old reddit where you can ask a question and whether it gets smacked down or not depends on how cool the community happens to be.

I am speculating, an impact this size must be related to a change in Google search ranking.

Although there appears to be a Monica effect (2020) in the data too. I stopped contributing myself around that time:

https://meta.stackexchange.com/questions/340906/update-an-ag...

You trust LLM responses? I feel they are not worth the paper they are written on

You don't necessarily need to trust them for them to be useful.

I can often evaluate a Copilot autocompletion (check that code looks right at first glance, check that it compiles, hover over method+type signatures to see their docs, run the code) in less time than it would take me to find+read a Stack Overflow answer.

No, no one sane would. Just like no one sane would trust random SO authors with any authority either. You still have to verify the information somehow, but at least in the LLM case I instantly get replies 24/7, and it's been reading more papers than any of the authors on SO.

ChatGPT has gotten quite a bit more accurate over the last few months. It does seem to have lost some creativity in the process, but it's much more true to the training data now.

Code is easy to verify anyway.

I trust AI as much as I trust random answer on SO. It is a direction for further research and evaluation.

"ask LLM, then look up the language/library docs to see if the idea makes sense" has worked relatively well for me.

I wonder if Google started de-ranking them in search results, or changed something similar that caused people to click through to Stack Overflow less frequently.

It is also possible there's a methodology issue. There's no source for where this data comes from.

I just googled a Microsoft tech question and first was a 3rd party provider plugin, then official MS Learn, then "people also asked...", and 4th was the SO block.

I scrolled down a lot and didn't see any scraper site links.

Copilot.

Reply on news.ycombinator.com