The quest for building machines that think and act like us has propelled significant advances in Artificial Intelligence (AI). While we now have systems that do restricted tasks very well, the vision of Artificial General Intelligence (AGI), where machines can seamlessly learn and do anything that a human does, continues to elude us. But ChatGPT is around and it seems magical, right? Does it have AGI? Not quite.
For the uninitiated, ChatGPT is powered by technology that belongs to the family of Large Language Models (LLMs). A language model can tell us that a “cat is sitting on the mat” is better English than “a cat is sitting in the mat”. If given the last few words in a sentence, a language model can also predict which word is most likely to come next. You can think of a language model as a black box with a lot of numbers, called “parameters”. These parameters implicitly capture diverse aspects of language such as grammar, word usage and even world knowledge (“I like noodles with sauce” is more likely than “I like noodles with pizza”, for instance).
Any sentence that you type in at the ChatGPT prompt is converted into a set of numbers that interact with the parameters of the language model to finally yield another set of numbers, which are rendered as output text. “Large” Language Models have on the order billions of parameters that are “learnt” from really large volumes of data. An example is all of the textual content that can be scraped from the entire web.
As part of training ChatGPT, it was also ensured that the system learns from human feedback. As a consequence, it got rewards for doing its job very well and was punished otherwise. Consequently, the end result is impressive: this new tribe of AI technologies is undoubtedly disruptive in more ways than we could have imagined. ChatGPT excels at many jobs humans traditionally take pride in – writing poems, code, online web content and so on. Do many of us then end up losing our jobs? Are such fears well founded?
Courtesy : The Indian Express