OpenAI ties the arrival of AGI to Astra's September 3 release

Two different axes have been folded into one word. Generality is one. Level of ability is the other, and Scientific American treats it as independent: a pocket calculator is superhuman and narrow, while the first large language models were general and barely articulate ^[10]. On that reading, "AGI" can name a system that does many things poorly or one that does a few things better than any person. The article does not say what criteria OpenAI used to make the call for Astra ^[21].

The substantive dispute is whether language is the whole of it. Blaise Agueera y Arcas, vice president of Technology and Society at Google ^[8], argues that next-word prediction forces a model to acquire grammar and a large amount of world knowledge, which is what makes language an "AGI-complete" task ^[13]. "We got AGI when we started to have language," he said ^[11]. He and Peter Norvig of Stanford made that case in a 2023 essay, resting it on language models doing things they had not been directly trained to do ^[12]. "The set of things that you could ask the model to do becomes open-ended," he said ^[15].

Human neurology points the other way: people with brain damage can lose the ability to understand or speak words and still do math, play chess and compose music ^[14]. His answer is that language models have no such clean separation between linguistic and reasoning ability, and that they can call on other computational modules to cover their weak spots ^[16].

Joscha Bach of the California Institute for Machine Consciousness puts the line somewhere else: a genuinely intelligent system should not have to devour the entire Internet, or spend vast computing power on the simplest query ^[17]. "We would have a system that is much more compact than us and has read far less than us," he said ^[18]. Training data volume, parameter count and compute per query can be measured and published, and one model's numbers compared with another's.

Fiction was precise about dates: the magazine notes that HAL 9000 went into operation on January 12, 1992, and that Skynet became "a new order of intelligence" on August 29, 1997 ^[19].

Ben Goertzel of the SingularityNET Foundation kept the term alive through the AI winter of the 1990s, when, he said, "If you even talked about building real thinking machines, people thought you were a nutcase" ^[6]. He does not think the human level is the interesting mark: "The human level is an arbitrary point," he said ^[7]. For decades each advance pushed the boundary between machine and human further out; the frontier labs are now trying to collapse the distinction ^[5].

So three researchers in one article place the boundary in three spots: Agueera y Arcas behind us, Bach ahead of us, Goertzel nowhere in particular ^[20]. Scientific American's own answer is that the only reason anyone would draw a precise line is that they have decided enough is enough ^[4].