The reason they failed is fairly simple. People are not that smart, on average, so the usage of their service, besides initial "have some fun" period, is limited to data researchers (of all kinds), scientists and students.
It may be not the smallest audience, but I think their best bet would be working on direct deals with schools and universities instead of trying to attract small-scale developers -- why would developers be interested in a project that probably won't ever pay back but will start charging after 2000 uses?
Their last pricing strategy failed horribly because it was too expensive. I hope they're just searching for a price point rather than hiding their fees. Anyone in contact with them about pricing?
I signed up for it and have played with the Ruby client library. While the limit of 2000 per month free API calls seems limiting it is enough for both play and possibly using it to generate some machine learning data for your own experiments (e.g., learn to produce similar results in a very narrow domain).
It may be not the smallest audience, but I think their best bet would be working on direct deals with schools and universities instead of trying to attract small-scale developers -- why would developers be interested in a project that probably won't ever pay back but will start charging after 2000 uses?