Overview of "Why We Don’t Have AGI Yet"
The paper "Why We Don’t Have AGI Yet" by Peter Voss and Mlađan Jovanović offers a comprehensive analysis of the persistent challenges faced in achieving AGI. The authors set the stage by revisiting the original aspirations of AI, focusing on creating systems with human-like cognitive capabilities, contrasting with the prevalent narrow AI solutions.
Historical Context and Conceptual Framework
The authors outline a historical progression from the initial AI concepts, which intended to mimic human-like intelligence, to the development of narrow AI. This transition carried adverse implications by prioritizing pre-defined tasks over inherent intelligence and adaptability. The term AGI was introduced to reinvigorate the original goals of AI, emphasizing systems with general cognitive abilities — such as abstraction, analogy, planning, and problem-solving.
Examination of Existing AI Approaches
The paper critically appraises the current state of AI, particularly the reach and limitations of LLMs like GPT. While these models have garnered significant success, the authors argue that they are ill-suited for evolving into AGI due to their statistical nature and lack of robust reasoning capabilities. In particular, issues such as catastrophic forgetting and the reliance on pre-trained, non-adaptive knowledge starkly differentiate these models from the requirements of AGI, which necessitate real-time learning, high-level reasoning, and metacognition.
Cognitive AI and Essential Abilities
Voss and Jovanović advocate for the development of Cognitive AI, which seeks to parallel human cognitive abilities, including language understanding, commonsense reasoning, and adaptation. They discuss the necessity for AGI to function with human-level cognition within the constraints of incomplete information and limited resources. Cognitive AI represents a strategic path to AGI, integrating the requisite cognitive abilities such as autonomous learning, concept formation, language understanding, and reasoning.
Theoretical Requirements and Methodological Missteps
The paper specifies essential cognitive requirements crucial for AGI, such as autonomous learning of actions and sequences, contextual understanding, and abstract reasoning, among others. These capabilities point towards an integrated approach that departs from the modular frameworks critiqued by the authors. The authors identify a lack of theoretical grounding and the misalignment of current AI practices with AGI aspirations as critical barriers.
Socio-Technical Challenges and Future Directions
A notable aspect of the paper is the exploration of socio-technical factors, which include the misaligned incentives of the AI research community, the dearth of AGI-focused projects, and inadequate funding. The authors emphasize the 'Narrow AI Trap' — the proclivity to prioritize immediate, narrowly defined achievements over long-term foundational progress towards AGI. They propose a reevaluation of research goals and metrics, emphasizing benchmarks that genuinely reflect progress towards AGI.
Conclusion
Voss and Jovanović argue for a paradigm shift toward the Third Wave of AI, seeking to integrate symbolic reasoning capabilities with sub-symbolic learning methods. This shift would prioritize autonomous, real-time, and incremental learning processes, which are aligned with the core objectives of AGI. By advancing Cognitive AI, the authors posit that it is possible to re-center the AI field towards achieving the original vision of AI — the development of true 'Thinking Machines' that enhance human potential and address broader societal challenges.
In summary, this paper offers a critique of current AI trajectories and proposes a strategic redirection towards building AGI with a robust foundation in cognitive learning and reasoning, combined with aligned research priorities and methodologies for achieving human-level intelligence.