So people are not asking new questions.

They are asking new questions. They’re just asking them directly to the AI interface.

The bigger problem is that they’re not getting new answers.

The vast majority of traffic to stack was always people searching for a solution to a problem, finding the question already asked, finding an accepted solution, using that solution, and leaving without contributing anything. (1)

A significantly smaller proportion of the traffic would be users who logged in, then voted on questions or answers they found useful. (2)

Smaller still, people who failed to do any real research prior, and asked a question that would inevitably be marked a duplicate. (3)

A tiny fraction would be people asking genuinely novel questions. (4)

And finally, the hardcore power users who would do all the actual work of sorting, moderating, and answering questions. (5)

We can imagine this traffic probably followed a distribution something like this:

x
x
x
x
x
x
x
x
x x
x x
x x
x x
x x x
x x x
x x x x
x x x x x
1 2 3 4 5

Now that we have LLMs,

(1) LLMs are “better google”. This traffic will be significantly reduced, almost to nothing. That is what the OOP graph is really showing.

(2) This traffic will also be reduced almost to zero, though there will be some holdouts who keep showing up to vote.

(3) LLMs are pretty decent at figuring out how to rephrase things, so people asking obviously duplicate questions will, again, be greatly reduced. Every once in a while, someone’s LLM will fail to translate their question into the question someone else asked on stack, and they will show up and ask their duplicate question, feeding into the main usefulness of duplicate questions - identifying different ways people phrase the same question so that these can all point to the same solution. These new phrasings will be picked up by LLMs, reducing the need for similar rephrasings in the future.

(4) Many of the non-duplicate questions asked on Stack are basic things that someone new to a language, library, or framework might ask. “Hello world” level syntax questions which can be answered via rtfm. These can now easily be answered via LLM. The remaining questions will be about weird bugs specific to particular software versions, or how to integrate arcane parts of different frameworks, etc, which will continue to feed LLMs.

(5) This traffic will probably be largely unimpacted, and these users user experience will probably improve as they now have less junk to deal with. Fewer idiots making vaguely racist remarks in the comments. Fewer duplicate or rtfm questions. A higher signal to noise ratio of interesting novel puzzles that their particular brand of autism thrives on.

So the new graph of traffic post-LLM might look something like this:

x x x x x
1 2 3 4 5

Admittedly, this might be hard on Stack’s business model of deriving ad revenue from web traffic… but as far as its actual purpose of providing answers to questions and providing a good experience to actual experts, it should continue to function just fine - if not better.