JPEG -> JPEG optimization certainly makes sense, since it doesn't require any decoder to be updated. JPEG -> [something else] compression (like Brunsli or Lept

JPEG -> JPEG optimization certainly makes sense, since it doesn't require any decoder to be updated. JPEG -> [something else] compression (like Brunsli or Lepton) can compress better though, since they can use (better) predictors, entropy coding, context modeling, etc. But they do require either a new decoder (saving both bandwidth and storage), or an on-the-fly conversion back to JPEG (saving only storage; bandwidth wouldn't change).