Comment by secretsatan
5 hours ago
It’s even noted they’re looking for text pre 2022 as afterward, it’s tainted by their own shit, they don’t believe in the crap they’re making.
5 hours ago
It’s even noted they’re looking for text pre 2022 as afterward, it’s tainted by their own shit, they don’t believe in the crap they’re making.
> they don’t believe in the crap they’re making
A spurious claim; they simply want to avoid model collapse.
Simply? It’s flawed if you can’t train on new data past 2022 without first checking ai hasn’t tainted it. Dismissing that as simply avoiding model collapse seems to miss the point.