Science Fiction Predictions of AI

Ted Chiang’s 2010 novella *The Lifecycle of Software Objects* ranks first in a new analysis scoring science fiction against how AI was actually built, earning 65 points overall and top marks for its depiction of teaching, maintaining and selling AI through human feedback and iteration. Richard Powers’s *Galatea 2.2* placed second with 8.4/10 for portraying language as something a neural network learns imperfectly before improving through experiments and teaching. The review found no pre-2017 work that clearly anticipated the transformer attention mechanism introduced in 2017, nor the rapid sequence from widely used chatbots to agents capable of discovering novel mathematical results. The rankings weigh three categories—creation, capabilities and use—with Chiang strongest on learning paradigm, iteration, human feedback, interface/distribution and economics, but scoring nothing on research discovery. The piece argues science fiction is descriptive rather than predictive, citing Ursula Le Guin, and that its value lies in exploring how people might behave under future circumstances.