Disagree. Say I'm looking for a list of countries and their populations.
Wikipedia almost certainly has this in a nice table, which I can sort by any column, and all the countries are hyperlinked to their own articles, and it probably links to the concept of population estimation too.
There will be a primary source - But would a primary source also have articles on every country? That are ad-free, that follow a consistent format? That are editable? Then it's just Wikipedia again. If not, then you have to rely on the LLM to knit together these sources.
I don't see wikis dying yet.
At work, I had rigged one of my internal tools so that when you were looking at a system's health report, it also linked to an internal wiki page where we could track human-edited notes about that system over time. I don't think an AI can do this, because you can't fine-tune it, you can't be sure it's lossless round-tripping, and if it has to do a web search, then it has to search for the wiki you said is obsolete.
OpenStreetMap does the same thing. Their UIs automatically deep-link every key into their wiki. So if you click on a drinking fountain, it will say something like "amenity:drinking_water" and the UI doesn't know what that is, but it links you to the wiki page where someone's certainly put example pictures and explained the most useful ways to tag it.
There has to be a ground truth. Wikipedia and alike are a very strong middle point on the Pareto frontier between primary sources (or oral tradition, for OSM) and LLM summary
Wikipedia almost certainly has this in a nice table, which I can sort by any column, and all the countries are hyperlinked to their own articles, and it probably links to the concept of population estimation too.
There will be a primary source - But would a primary source also have articles on every country? That are ad-free, that follow a consistent format? That are editable? Then it's just Wikipedia again. If not, then you have to rely on the LLM to knit together these sources.
I don't see wikis dying yet.
At work, I had rigged one of my internal tools so that when you were looking at a system's health report, it also linked to an internal wiki page where we could track human-edited notes about that system over time. I don't think an AI can do this, because you can't fine-tune it, you can't be sure it's lossless round-tripping, and if it has to do a web search, then it has to search for the wiki you said is obsolete.
OpenStreetMap does the same thing. Their UIs automatically deep-link every key into their wiki. So if you click on a drinking fountain, it will say something like "amenity:drinking_water" and the UI doesn't know what that is, but it links you to the wiki page where someone's certainly put example pictures and explained the most useful ways to tag it.
There has to be a ground truth. Wikipedia and alike are a very strong middle point on the Pareto frontier between primary sources (or oral tradition, for OSM) and LLM summary