The supply of vision data is nearly unbounded, but the supply of quality human text data is much more like a finite resource—and one for which we have probably passed the point of peak supply. The nominal supply of text may continue to expand, but it’s increasingly auto-generated.
IL’s comment has a BOTEC arguing that video data isn’t that unbounded either (I think the 1% usefulness assumption is way too low but even bumping it up to 100% doesn’t really change the conclusion that much).
The supply of vision data is nearly unbounded, but the supply of quality human text data is much more like a finite resource—and one for which we have probably passed the point of peak supply. The nominal supply of text may continue to expand, but it’s increasingly auto-generated.
IL’s comment has a BOTEC arguing that video data isn’t that unbounded either (I think the 1% usefulness assumption is way too low but even bumping it up to 100% doesn’t really change the conclusion that much).