Could you elaborate here? With a self-signed cert, the server is still not sending secret information in the first few packets; it just tells you (without authentication) which public key to use to encrypt the later packets (well, the public key to encrypt the private key for later encryption).
The threat model would be eavesdroppers who can't control the channel, only look. Using the SS cert would be better than an unencrypted connection, though still shouldn't be represented as being as secure as full TLS. As it stands, the server is either forced to wait to get the cert, or serve unencrypted such that all attackers can see.
I think the CA system by itself is inadequate, but unlike unauthenticated TLS, actually does provide some security.
What's especially baffling about self-signed certificate advocacy is the implied threat model. Low- and mid-level network attackers and crime syndicates can't compromise a CA. Every nation state can, of course (so long as the site in question isn't public-key-pinned). But nation states are also uniquely capable of MITMing connections!