I don't believe it's impossible. In fact, there are two clear ways to do it: either
1. have a database of known algorithms, and map-reduce out the ones that produce the most signal for your data-set (this isn't Hard, but it requires a globally-networked language-neutral ABI-neutral algorithm repository and a free-use cloud compute cluster to run the heterogeneous algorithm-tests on), or
2. expect the computer to invent a novel, efficient (or at least polynomial) algorithm in response to your data-set on the fly. This is a Hard problem--since solving it basically means that computers can now take the jobs of Mathematicians in proving novel theorems. I don't think that's "impossible" either--obviously, Mathematicians are performing some describable algorithm in their heads to come up with novel proofs--but it's likely a Big Data problem in the same way most AI problems have turned out to be; not something you can ask your workstation to do.
Do you really believe that it _should_ be impossible for an average person who desires a sudoku solver to get one without any engineering knowledge?
The spirit of Bret's talk, and even this response, is of thinking broadly. The question is not "is this possible", but "should this be possible".