AIサービスのシュリンクフレーション
https://search.bilibili.com/all?vt=44084965&keyword=Seedance降智&from_source=webtop_search&spm_id_from=333.1007&search_source=5bilibili 検索 "Seedance降智"
@HclHno3: AI video models have a shrinkflation problem.
Launch day: full model, stunning quality, one-shot results.
1 months later: silently swapped to a distilled model, worse outputs, you need 5 draws to get what used to take 1.
The price stays the same. The burger got smaller.
A 10-sec video costs $1.30–$5.00 in real GPU compute. Companies charge $0.05–$0.20. The math doesn't work — so they distill the model and pray you don't notice.
Sora did it. Seedance 2.0 is doing it. #Seedance2
https://pbs.twimg.com/media/HBiNvY5WEAAFTbr.jpg
最初は高性能なモデルでサービスを行い、その後こっそり蒸留モデルへ切り替える
これに関しては蒸留モデルだけでなくNVFP4と言った量子化モデルを使用している可能性もあるかもしれない(蒸留より作るコストが遥かに安いし…)