Ramez Naam has some wonderful posts at Charles Stross’s blog about why The Singularity Is Further Than It Appears, including some graphs on Why AIs Won’t Ascend in the Blink of an Eye. In short, a technological Singularity is plausible only if designing smarter AIs is a problem of linear complexity. If it is a more challenging problem, a hard takeoff becomes more unlikely or impossible. William Hertling makes some thoughtful points in rebuttal, which probably depend too heavily on the lack of constraints to Moore’s “Law”.
The techno-utopianism (or even just techno-optimist) of the Singularity too often relies on really naive ideas of the self, even if clever writers subvert them. People can’t achieve immortality by uploading their souls to a computer because souls are an illusion. The human mind is fundamentally embodied; hours in a sensory deprivation tank cause minds to create powerful illusions of sensation. And the tricks computers use to simulate human behavior also demonstrate their profound lack of the human experience that drives human behavior. Machine translation can improve with better representation of linguistic structure, but it’s never going to approach real translation until the machines understand semantics, pragmatics, and stylistics in the way that humans do.