If you take even a passing interest in artificial intelligence, you will inevitably have come across the notion of artificial general intelligence. AGI, as it is often known, has ascended to buzzword status over the past few years as AI has exploded into the public consciousness on the back of the success of large language models (LLMs), a form of AI that powers chatbots such as ChatGPT.
That is largely because AGI has become a lodestar for the companies at the vanguard of this type of technology. ChatGPT creator OpenAI, for example, states that its mission is 鈥渢o ensure that artificial general intelligence benefits all of humanity鈥�. Governments, too, have become obsessed with the opportunities AGI might present, as well as possible existential threats, while the media (including this magazine, naturally) report on claims that we have already seen 鈥渟parks of AGI鈥� in LLM systems.
Advertisement
Despite all this, it isn鈥檛 always clear what AGI really means. Indeed, that is the subject of heated debate in the AI community, with some insisting it is a useful goal and others that it is a meaningless figment that betrays a misunderstanding of the nature of intelligence 鈥� and our prospects for replicating it in machines. 鈥淚t鈥檚 not really a scientific concept,鈥� says at the Santa Fe Institute in New Mexico.
Artificial human-like intelligence and superintelligent AI have been staples of science fiction for centuries. But the term AGI took off around 20 years ago when it was used by the computer scientist Ben Goertzel and Shane Legg, cofounder of the AI firm DeepMind. The phrase neatly encapsulated the growing sense that the field should move beyond narrow applications to build systems that can do everything a human can do.
Since then, DeepMind in particular has sought to redefine AGI such that it pertains only to 鈥渃ognitive tasks鈥�. Last year, Legg, together with DeepMind cofounder Demis Hassabis and their colleagues, elaborated on what constitutes an AGI. They proposed a , where the top level is a system that can 鈥渙utperform 100 per cent of humans鈥� across a 鈥渨ide range of non-physical tasks, including metacognitive abilities like learning new skills鈥�.
The problem with AGI
鈥淭he levels idea is really pointing out that there鈥檚 this continuum,鈥� says team member at DeepMind, now part of Google. 鈥淭here鈥檚 this progression as technology evolves.鈥� Morris hopes their work will draw more attention to the idea, and ultimately to some form of consensus on what AGI actually is: 鈥淲e would love to have folks from those other fields that study intelligence and learning working together with our researchers on developing these benchmarks.鈥�
But Mitchell points out that intelligence is itself a multidimensional concept, with a lot of crossovers with other equally murky concepts, such as sentience and understanding. As such, it isn鈥檛 readily measurable with a test in the same way as other, more concrete tasks, like the ability to translate language.
Applying more scrutiny to when an AI could be considered an AGI might yield progress, but Mitchell is still sceptical that the sort of machine that AGI proponents envisage will be achieved, because it is unclear whether the faculties of human intelligence can ever be abstracted into standalone concepts 鈥� never mind replicated in AI. 鈥淭here鈥檚 a kind of faith that the field has had for a long time, that we can develop human-level intelligence in these disembodied substrates,鈥� she says. 鈥淲hether that鈥檚 possible or not, I think it鈥檚 a big open question.鈥�
For at Oregon State University, the problem with AGI is a more practical one 鈥� namely that it is a mistake to define artificial intelligence with respect to humans. 鈥淲e have this focus on replicating our capabilities, and this leads to the rampant anthropomorphisation of our systems, giving them names like Siri and Alexa.鈥�
Instead, he says, we should think of AI as 鈥渁n intelligence prosthetic that can do certain things for us鈥� 鈥� which sounds a lot like what the AI community had in mind before the concept of AGI came along.
