I think the point GP was making was that he got literally 4 words into the "simple to explain" explanation before realizing this explanation is going to be like tons of other things he reads here and elsewhere that require so much prior and specialized knowledge as to be nearly impossible for someone lacking a high level of experience in the area to understand.
I often feel this way when security topics come up on HN. I've been writing code a long time and I think I'm of moderate intelligence and I often have no idea what anyone is talking about. I try not to lose sleep over it because I understand that security can be really, really complicated (even if the issue is "easy to explain") and so I'm just not going to get it.
So it's not very important that you understand exactly how to go from "errors/exceptions generated by the target" to "recovery of plaintext". Even if you understand the specifics of those kinds of attacks, they're even more complicated (and fun) here.
The interesting nut of the attack is: there's a well-known error oracle in both TLS and SSLv2, but it's long since been mitigated in both protocols. However, details of the crappy crypto in SSLv2 conspire to make the mitigation for padding oracles in SSLv2 ineffective, and you can use the SSLv2 oracle to attack TLS ciphertexts if TLS and SSLv2 ever share keys.
I think everyone can understand why "Error: The fourth character of your password is wrong" is both more "usable" and insecure.
Had to deal with a program recently that gave "corrupted MAC" error on the server. All searching etc hinted at it being a network issue. Only after having fiddled with the network for ages and getting nowhere did i set up a loopback test and found the same error.
Turns out the program i was dealing with was using an old lib.
Padding oracles are a subset of "error oracles", which are oracles that come from exception processing. The "useful computation" you're being forced into doing is "decrypting, checking for errors, and then somehow signaling the error."
Think of error oracles this way: imagine a situation in which an adversary can hand you a ciphertext which, when decrypted, will trigger an error (in a padding oracle, that error is "the padding is wrong"). Now imagine that the error, revealed to the adversary, allows them to infer one bit of the plaintext. Not very useful by itself, but, as a final step, imagine that the attacker can permute the same ciphertext, generating (or not generating) an error and allowing them to reveal a different bit of the plaintext.
As long as they can keep doing that, they'll eventually recover the whole plaintext.
Here's a good starting point:
http://cryptopals.com/sets/6/challenges/46/