The lab next door to mine uses this system out the wazoo in mammalian tissue culture and someone in my lab is thinking of using a modified version of it in yeast.

Step #1 - hold your horses. Generally good advice in most biological circumstances.

This system when unmodified and vanilla is a great way to make a few germline modifications more efficiently than the usual methods. It is made from elements appropriated from a bacterial antiviral defense system. It consists of two segments, the Cas9 protein and guide RNAs. Cas9 is a nuclease, a DNA cutting enzyme, that requires a GGG triplet to exist directly next to its cut site (nobody's modded that that I've seen so far, I don't know enough about its structure to know if thats possible). The rest of its targeting is mediated by the guide RNA - a small 100ish base pair RNA that consists of a constant hairpin structure that latches into Cas9 and a 20 base-pair segment that you can modify to contain any sequence. With that 20 base pair sequence adjacent to the GGG cut-site, you can target the nuclease activity to anywhere in a large genome with pretty good specificity, about one in trillions of random base pairs (ignoring things like repetitive sequence which is all over the place in actual genomes).

The point for heritable genetic modification being that when you create a DNA lesion like this, the DNA repair mechanisms kick in and try to repair the chromosome off of any similar DNA in the cell (usually the other copy of the chromosome) via homologous recombination. A paper on zebrafish genetic modification I've seen with this system microinjects into eggs RNA coding for the CAS9 protein, guide RNAs, and DNA containing the transgene and chunks of homologous sequence bracketing it so that the genomic DNA just keeps getting cut until it repairs off the introduced DNA. They got mutagenesis efficiency between 75% and 99% with this method, a good deal higher than usual for zebrafish microinjections. I know it's been done in mice but I don't know how efficiently, although these days the main way you transform mouse germline DNA has been doing low-efficiency transformation and selection of embryonic stem cells in culture followed by 2 generations of mucking around making chimeric embryos from those stem cells and breeding the resulting animals so this definitely saves time if it works well. People are also using it along with the usual methods of getting nucleic acids into cultured cells to produce stable changes to such cultured cells at high efficiency and specificity - most introduction of DNA into tissue culture is either temporary with the DNA never making it into chromosomes and going away before too long, or done via retroviruses and thus is kind of random in terms of where and how frequently it inserts.

Course, any time you cut the DNA like this you have a reasonable subset of cells not actually repair off a template but just nonspecifically stitch any recently-cut DNA (not necessarily from the artificial cut site, DNA is breaking and repairing everywhere all the time) together via non-homologous end joining. I don't know how frequently this happens with this protocol and how frequently when it does happen it leads to something problematic. I would imagine that this would become more of an issue any time you are cutting at multiple sites at once with multiple guide RNAs (which you can do with this system) and care about getting the same number of outputs you got out as inputs you put in without lots of selection of successes from the failures. There's also apparently detectable levels of off-target effects, where a few base pairs difference between the guide RNA sequence and the DNA sequence still allows cutting at similar but not identical sites.

In short people are excited because of its sequence specificity and ability to target DNA changes via DNA repair in a single generation as opposed to multiple, it's cool, and works pretty efficiently as opposed to methods that need lots of selection. It's an incremental improvement, even if it is a pretty nice and big increment.

I have a hard time understanding how you could use this for somatic gene therapy with much more success than has already been done with retroviruses. No matter how you slice it, if you want to do gene therapy in a multicellular creature you STILL need to get a package of nucleic acids into a cell which is not easy at all in a living organism as opposed to a monolayer of cells in a controlled environment, or a population of microbes that you can just select the successes from and discard the more numerous failures. Perhaps it's a way to make the cells you do manage to modify less likely to turn all cancerous, but it doesn't solve the problem of actually getting the nucleic acids INTO them.

To me, work going on in which people are mutating Cas9 to kill its DNA-cutting activity and modifying it to use as a scaffold for programmable transcription factors/repressors for artificial regulatory systems is rather more interesting. Using these systems you can potentially arbitrarily target any gene and mess with its expression once you get the required bits into the genome. A major boon for research and artificial genetic systems. Cutting DNA in vivo can be a messy proposition compared to regulating it sometimes even if the artificial not-tuned-by-evolution regulation is at times low-precision.