Boffin rediscovers 1960s attempt to write fiction with computers
IBM 650 could cobble together fiction long before Watson marketing kicked into gear
Next time IBM tries to convince you that Watson is the latest and greatest innovation that couldn't possibly have been done any time other than now, know that Big Blue tried to get a computer writing short stories in the 1960s.
The existence of IBM's old work has been re-discovered by James Ryan of the University of California Santa Cruz, who found old documentation of work by one Joseph E Grimes to make an IBM Model 650 spin yarns. Here's a sample of its output:
A LION HAS BEEN IN TROUBLE FOR A LONG TIME. A DOG STEALS SOMETHING THAT BELONGS TO THE LION. THE HERO, LION, KILLS THE VILLAIN, DOG, WITHOUT A FIGHT. THE HERO, LION, THUS IS ABLE TO GET HIS POSSESSION BACK.
The story of the story-generator is more interesting than the story the generator tells. In his paper, Ryan explains that Grimes' work dates back to the early 1960s.
Given that the IBM 650 (as IBM's archives explain was a punch-card machine that used magnetic drum memory able to hold just 20,000 digits at 2,000 addresses (later getting upgraded to tape storage), getting it to do anything at all was quite a feat. So how did Grimes manage to program it to generate stories?
Working on a machine built by the Universidad Nacional Autonoma de Mexico (which invited social scientists to use its machine), Grimes' aim was to generalise “rules of folk tales” published by Russian Vladimir Propp in a monograph called “Morphology of the Folktale”.
In essence, Propp's work broke folk stories down into classifiable elements; Grimes took those elements and told the computer how to branch between them.
Grimes told Ryan he started work in 1960 or 1961, and this came to the attention of IBM Mexico, who sent Life Magazine photographer Cornell Capa and a journalist along for a piece that ran in an IBM house journal in 1963.
That long-ignored story told us that “Dr. Grimes uses a basic fairy tale pattern and then programs the computer using a Monte Carlo approach so it will print out stories in a completely random manner. This involves selecting at random the path from episode to episode within the story, the action by which each episode is accomplished, and the characters which will assume the various roles. The computer can produce about 1020 plausible variations on a single fairy tale theme, according to Dr. Grimes.”
Interestingly, Grimes himself treated the computer-generated stories as “field instruments” for studying linguistics. By getting people to read the stories, he could observe when they got bogged down “trying to follow the plot”. That, he explained to Ryan, let him formulate “a hypothesis about the underlying linguistic system.”
Grimes later reworked his program, “probably in FORTRAN”, to run on an IBM 1401, but eventually he abandoned the project because the stories were “all boring”.
Ryan notes that nearly all academic sources refer to a much later project as the first “story generator” – the 1971 “automatic novel writer” demonstrated by Sheldon Klein, a history that has to be rewritten to include Grimes' work.
MIT started a movie plot-based story generator called SAGA II at more-or-less the same time, but Ryan notes that even if that work was demonstrated first, Grimes's was “the earliest known Proppian system,” and the first to use what's called “story grammars”. ®