Core Concepts
Artificial General Intelligence (AGI) is a continuous spectrum, and it already exists in the form of large language models (LLMs) that exhibit functional general intelligence, despite the common perception of a distinct threshold.
Abstract
The author argues that the current definitions and perceptions around Artificial General Intelligence (AGI) are misguided and unnecessary. They assert that Large Language Models (LLMs) already exhibit functional general intelligence, despite not being formally classified as AGI.
The author agrees with the Microsoft Research team's assessment that GPT-4 has "sparks of AGI", suggesting that AGI is not a binary state but rather a continuous spectrum. They contend that it is "disingenuous" to claim that LLMs are not functionally generally intelligent, as demonstrated by their responses to text injected into the prompt.
The author suggests that the common notion of a distinct threshold for AGI is flawed, and that the reality is a continuous spectrum of increasingly capable AI systems that already possess many attributes of general intelligence. This challenges the traditional conceptualization of AGI as a distinct, well-defined milestone.
Quotes
"AGI is a continuous spectrum. It's here already. It's us that have a 'threshold'."
"As a natural language understanding (NLU) researcher I find the definitions all a little misguided and largely unecessary."
"That's because, i think it's disingenuous to claim that LLMs aren't functionally (as tested by responses to text injected into the prompt) . . already generally intelligent."