AGI: CEOs Announce It, Researchers Denounce It as a Marketing Term

The term AGI is increasingly prominent in the speeches of leaders like Greg Brockman, Jensen Huang, and Elon Musk. However, researchers and engineers contest that such a milestone has been reached, deeming the term poorly defined and viewing it primarily as a marketing tool. The debate contrasts slogans, current technical limitations, and warnings about the trajectory of models.

Technical voices predict a slowdown and limits by 2025

Several figures in research are calling for caution regarding the trajectory of current systems. Yann LeCun stated last year that one cannot assume that a mere increase in data and computing power will lead to more intelligent AI. Ilya Sutskever predicts that by 2025, the industry will need to return to a research phase, emphasizing that models generalize far less effectively than humans, which he considers fundamental. Emily Bender goes further, arguing that a general-purpose thinking machine is not achievable through human engineering and that such a system would require an approach that does not work.

Proclamations of "AGI" multiply on X and in tech

Greg Brockman declared in September the "era of AGI" while presenting OpenAI's latest model. The term, once distant and theoretical, is now widely circulated among leaders and users on X. Elon Musk, Sam Altman, and Jensen Huang assert that this milestone has been reached. Following the release of Astra in September, Jensen Huang wrote on X that "AGI has arrived," mentioning the transition from ChatGPT to o1 and then Astra in four years, and congratulating the OpenAI team. This week, Elon Musk indicated that he "felt AGI" in response to a short film generated by Claude. Matan Grinberg, CEO of Factory, also claims that AGI is already here and speaks of a "post-AGI world."

Industry voices denounce an imprecise definition of AGI

Industry voices argue that AGI is too poorly defined to serve as a useful reference, and that proclaiming it to be already here amounts to misleading marketing. Peter Voss describes these statements as absurd, asserting that if AGI truly existed, there would be no doubt about it. Dario Amodei stated last year that he has never considered AGI to be a clear term and has always viewed it as marketing. Emily Bender believes that the use of futuristic nomenclature fuels the illusion of systems being more advanced than they are, referring to a technology that is imagined rather than existing, and presents AGI as a marketing term. For Alan Chan, AGI is not a good indicator of progress, as definitions confuse the ability to innovate with the capacity to perform tasks in the real world, illustrated by expert systems in coding that are slow and costly to adapt to new jobs. Meanwhile, companies like Anthropic and OpenAI are under significant financial pressure to demonstrate increasingly powerful technologies in anticipation of potential multi-trillion dollar IPOs.

The origins of AGI and what a true system would promise

The term "Artificial General Intelligence" was introduced in 2002 by Ben Goertzel in a book dedicated to speculative versatile AI, to which Peter Voss contributed a chapter presenting it as a still unattained goal. AGI frequently refers to an AI capable of performing many tasks as efficiently, if not more so, than a human, and, more generally, an intelligence that could match or surpass human intelligence.

Current systems struggle with rapid learning of new tasks

Peter Voss observes that leading chatbots remain confined to limited capabilities and do not acquire new skills with human-like plasticity. He recalls that a human can learn a call center job in one or two days, while a system deserving the AGI label should be able to do the same. In contrast, even for customer support, it would currently require billions of dollars in engineering and hundreds of thousands of example conversations to achieve what could be deemed mediocre results. By his definition, a system learning to drive in 24 hours is not expected anytime soon. Voss also believes that the enthusiasm of investors and the public for LLMs has "sucked all the oxygen out," making alternative paths more difficult. While some leaders claim a critical threshold named AGI, other engineers consider this term exaggerated. In the background, figures like Jensen Huang and Greg Brockman are discussing AI at the highest political levels, as seen in a recent meeting with Donald Trump.