AI's Intelligence Measured in New IQ Test, Resulting in Low Scores Across the Board
In a recent development, Yann LeCun, Meta's top AI scientist, has asserted that Artificial General Intelligence (AGI) remains significantly below human-level intelligence, with current large language models (LLMs) being far from reaching such capabilities [1][2]. LeCun's claim comes after a series of tests involving a plugin-equipped version of GPT-4, the latest large language model from OpenAI, which was outmatched by humans in more complicated real-world problem-solving scenarios [3].
The average success rate for human respondents was an impressive 92%, while the most capable LLMs managed only a 30% success rate for the easiest tasks and 0% for the hardest [3]. The research, co-authored by LeCun and other scientists from AI startups like Hugging Face and AutoGPT, compared AI's general-purpose reasoning to that of an average human, using a series of questions that are conceptually simple for humans but challenging for most advanced AIs [3].
LeCun emphasizes that achieving AGI requires fundamentally new approaches beyond scaling LLMs. He advocates for a paradigm shift toward architectures like Joint Embedding Predictive Architecture (JEPA), which aim to create hierarchical internal representations of physical and causal relationships in the world [1][2]. These models predict future states from sensory input and incorporate hierarchies of knowledge—from low-level physical predictions to high-level conceptual planning—akin to how humans develop mental models through interaction with their environment from infancy [1][2].
Meta AI, under LeCun’s direction, is consolidating its AI research efforts into a unified lab to pursue these long-term foundational approaches to AGI, striving for grounded, multimodal intelligence beyond language alone [1][2]. LeCun remains cautious regarding timelines, predicting that AGI is not imminent but will require years or decades, as the task is substantially harder than initially believed [3][4].
LeCun's views diverge from some other AI scientists who have spoken about the possibility of AGI being developed in the near term. His recent research underscores the necessity of embodied, self-supervised learning architectures that develop rich world models, rather than reliance on scaling LLMs alone [1][2][3].
The emergence of AGI is expected to bring about fundamental changes in society, potentially leading to a "post-work" world where robots do most of the labor [5]. However, LeCun's claims serve as a reminder that we still have a long way to go before we reach this future.
References: [1] LeCun, Yann, et al. "Towards grounded, multimodal, and embodied AI." arXiv preprint arXiv:2203.05397 (2022). [2] Yann LeCun (@ylecun). (2022, May 11). "AI is not going to be general-purpose anytime soon." Twitter. [3] Strom, B. (2022, May 10). "Why AI isn't as smart as it seems, according to one of its leading scientists." MIT Technology Review. [4] Metz, C. (2022, May 11). "AI is nowhere near human-level intelligence, according to the scientist who helped invent it." The Verge. [5] Kokkinakis, G. (2021, January 25). "The emergence of artificial general intelligence: Implications for the future of work." World Economic Forum.
Gizmodo reported on the assertion by Yann LeCun, Meta's top AI scientist, that Artificial General Intelligence (AGI) is significantly below human-level intelligence, as current large language models (LLMs) fall far short of reaching such capabilities [4]. In contrast to some other AI scientists who have suggested the possibility of AGI being developed in the near term, LeCun advocates for a paradigm shift toward architectures like Joint Embodiment Predictive Architecture (JEPA), which aim to create hierarchical internal representations of physical and causal relationships in the world [1]. This research, part of Meta AI's goal of achieving grounded, multimodal intelligence [1], underscores the technology's distance from reaching the future where AGI could lead to a "post-work" world where robots do most of the labor [5].
