Many AI developers and frontier labs are openly pursuing artificial general intelligence (AGI), which scientists describe as human-like intelligence in which a model can reason like humans and learn new capabilities outside its training data. But does that mean that they’re all on the right path?
Chasing the dragon
If we achieved AGI, how would we even know? Many AI benchmarks exist, but François Chollet co-founded the ARC Prize just a few months before speaking with the Dwarkesh Podcast to get to the bottom of this.
This article is part of TechRadar Pro’s QOTD project to provide an insight into the minds of the brightest and most recognized figures in the technology industry today and in years gone by. Read the full series here.
In comments on the podcast, he explained that large language models (LLMs), which are based on neural networks, is a dead end when it comes to achieving AGI. That doesn’t mean there wasn’t room to improve LLMs so they were more powerful, autonomous and useful to businesses. But that isn’t the same thing as AGI.
Intelligent systems
Despite Chollet’s steadfast belief that LLMs will not get us any closer to AGI, many technology executives have spent the last few weeks opining about the possible threat that they pose to humanity.
Ahead of OpenAI and Anthropic’s as-of-yet-undetermined IPOs, both companies have taken turns disclosing increasingly worrying incidents, including cybersecurity breaches. These stories have, in many people’s eyes, simply served to hype up the capabilities of existing technologies for marketing purposes.
As for Chollet, the whole purpose of the ARC-AGI benchmark is to determine true progress toward AGI. While LLMs are performing better on this especially tough benchmark, none have come close to hitting the mark.
