Blue sky ideas: forget the Booker, Turing test is goal of authorbots
“The Lex writer had always loved busy Wall Street.” So begins “The Pink Ballpoint”, an auto-generated short story about a fictional Lex writer.
Programmers hope computers will eventually write fiction and play music as skilfully as people, the second blue sky idea we examine in today’s thematic column.
If they succeed, Artificial intelligence would have aced a variant of the Turing test, which determines whether a computer can pass as human. It could also be a way to make money. Consumers might pay for tales and tunes tailored to them.
The programme that wrote “The Pink Ballpoint” shows its limitations via unintentional surrealism. As our heroine gazes out at “the snow flurrying like a typing rhinoceros”, an FT subscriber approaches her with the curious greeting: “I love you and I want information.”
In contrast, the GPT-3 AI system developed by OpenAI, a San Francisco tech laboratory, can produce sophisticated text in response to cues from human operators. It can sometimes fool readers into believing it might be a human.
The programme produced an ingenious Christmas tale of its own about this column, marred only by a Scrooge-like characterisation of the section editor and an irrelevant grudge against climate campaigner Al Gore.
GPT-3’s secret – and sometimes its downfall – is exhaustive statistical sampling of text. OpenAI, which has received $1bn of investment from Microsoft, hopes one day to commercialise its products. Another of these is MuseNet, a neural network that synthesises music. Its sound files convincingly resemble famous artists, albeit “performing at a festival in a strong wind”, as one listener put it.
OpenAI has set itself a tough task: faking human singing with analogue instrumental accompaniments. Electronic music is easier to auto-generate. Start-ups such as AiMi of the US and AI Music of the UK have programs that produce never-ending streams of electro.
What may be missing in this brave new world is originality. The more closely auto-generated content corresponds to prior art, the better programmers deem it to be. However, cruder efforts may have a clunky robotic charm absent from next-generation alternatives. They are authentically inauthentic.
This is the second of five articles on blue sky thinking published by Lex today. Look out for the others in Lex online.