A whole lot of people – including computer scientists who should know better and academics who are usually thoughtful – are caught up in fanciful, magical beliefs about chatbots. Any sufficiently advanced technology and all that. But why chatbots specifically? Not all ‘generative AI’ or even all LLMs spark so much anthropomorphization and attributions of ‘general intelligence.’ It wasn’t StableDiffusion’s image generator or GPT-2 that excited this imagination. We don’t see much anthropomorphization for the latest versions of Copilot (a code generator) or Suno (a song generator), either. It was ChatGPT. Now often it’s character.ai.
Instruction tuning, and later chat tuning, are the methodological innovations that make some machine learning pass the Turing Test. An object that we can use imperative language with (“write me an email…”) seems to have captured imaginations in ways that other, frankly better, objects don’t.
Continue reading “it’s the interface”