When an older technology fades from online discussion, does that leave future AI models with less to learn from? And what happens with new tools if fewer people post troubleshooting examples publicly?
Will AI run out of useful information as technologies stop being discussed online?
Answers
Old discussions don't automatically vanish from a model's training data just because people stop adding new ones. A model also doesn't need a constant stream of new material just to keep working.
The bigger issue is access over time. Old pages can disappear, become harder to find, or be unavailable for a future training run. Even when examples survive, they may describe an old version of a tool and miss later changes or obscure bugs.
Reply to this answer
For newer tools, lack of public examples could be a real gap. Documentation tells you what a feature is supposed to do, while troubleshooting discussions often show the weird inputs, wrong turns and fixes that don't make it into a reference page.
I'd use AI suggestions as leads, then check the relevant version of the docs and test changes somewhere safe. If you solve a tricky problem, publishing a small reproducible example and the fix helps other people as well as future datasets.
Reply to this answer
I wouldn't assume the discussion has stopped. Code repositories, tutorials and public technical discussions still contain plenty of examples. Some conversations have also moved into private or poorly indexed spaces, so a search engine may not turn them up.
That creates a problem for humans too. An answer buried in a private chat isn't much help to the next person with the same error.
Reply to this answer
There is a possible feedback loop if future models are trained heavily on low quality AI generated material. Repeatedly training on synthetic output can degrade a model under some conditions, but it isn't an automatic fate for every model that uses synthetic data.
What gets selected and checked matters. Training data can include documentation, code and other sources alongside generated examples, rather than indiscriminately feeding a model its own output.
Reply to this answer
Add your answer
Answers are reviewed before publication.
b.624~tramspool:
So there's a difference between what an existing model has learned and what a future model can actually collect and use.