70 years ago Ross Ashby designed an 'intelligence-amplifier'. Have we finally built it?

In the summer of 1956, John McCarthy convened the Dartmouth workshop to explore how a machine might be made to simulate intelligence; it was the meeting that named the field artificial intelligence. Among those who presented was the Cybernetician W. Ross Ashby, passing through on his way home to England after a year's fellowship at Stanford. What he demonstrated was something different: his homeostat, a prototype "intelligence-amplifier" built on his own exacting definition of intelligence as appropriate selection. On that much McCarthy agreed: the notes record him holding, with Ashby, that selection was the whole problem. Seventy years later, we argue that today's large language models have realised Ashby's ambition rather than McCarthy's — they are less machines that are themselves intelligent than amplifiers of a human's selection, coupled into a loop. Locating the intelligence in the model rather than the coupling is a category error, and it lies beneath a series of misunderstandings about how these systems can be used and what their performance really is. Ashby had seen the deeper point: that forgetting is not a defect of such a machine but its method — "any device that can lose information can generalise," he wrote; to generalise at all is to throw the particular away. The machine we have built runs on exactly that, and the field that runs it has forgotten it was his.