Some interesting talks I heard recently:
Ron Rivest talked about the MD6 hash function, which is a candidate in the upcoming SHA-3 hash contest, sponsored by NIST. It is based on a tree structure instead of a chain structure so that it's parallelizable, is resistant against certain cut-and-paste attacks, and is provably resistant to differential attacks, which have been worrying people. Extensive computer simulations were used to verify that MD6 "scrambles" its inputs enough to put a lower bound on the complexity of this particular kind of attack which is higher than the birthday bound.
Kevin Knight gave a talk on machine translation without parallel texts. Which sounds, on the face of it, absurd. Except that translation can be thought of as just applying a secret code, and statistical methods for cracking codes (without having much/any parallel text, of course!) have been around for a long time. Nowadays, computers can solve the cryptogram you find in the newspaper, and some analagous problems, but we don't know a whole lot about stuff beyond that. This kind of research could be useful for helping to translate texts in specific domains (e.g. computer manuals), where lots of example text can be found but little parallel text is available. Methods trained on general corpora often do quite poorly in this kind of niche.
By the way, I love this comic by Kevin Knight.