Critters tend to parrot exact sentences from the model

[lynn]

The problem is threefold:

• We could discard exact matches, and call the Markov synth function again if we run into one. This is what twitter_ebooks does, but it sounds tricky for us, for a bunch of reasons.
• We could make the text generator jump around more often. I think it'd be cute to sometimes only use 1 word of lookback instead of 2.
• UI-wise, it should be more obvious that you have to feed the critters a heckton of sentences (preferably from different sources!) before they start getting mixed up in really funny ways.