How Far Are We From AGI: Are LLMs All We Need?

TL;DR

The paper surveys progress toward Artificial General Intelligence (AGI) by decomposing AGI into internal mind-like capabilities (perception, memory, reasoning, metacognition), external interfaces (digital, physical, intelligent), and supporting systems, then addressing alignment and a roadmap. It synthesizes current state-of-the-art in perception, reasoning, memory, and metacognition, explores the challenges and architectures for scalable systems, and discusses evaluation, safety, and governance with multiple workshop perspectives. Case studies across science, generative vision, world models, decentralized AI, coding, robotics, and human-AI collaboration illustrate practical trajectories and risks. The work aims to foster a shared understanding and responsible discussion among researchers and practitioners about how far we are from AGI and how to responsibly steer its development.

Abstract

The evolution of artificial intelligence (AI) has profoundly impacted human society, driving significant advancements in multiple sectors. AGI, distinguished by its ability to execute diverse real-world tasks with efficiency and effectiveness comparable to human intelligence, reflects a paramount milestone in AI evolution. While existing studies have reviewed specific advancements in AI and proposed potential paths to AGI, such as large language models (LLMs), they fall short of providing a thorough exploration of AGI's definitions, objectives, and developmental trajectories. Unlike previous survey papers, this work goes beyond summarizing LLMs by addressing key questions about our progress toward AGI and outlining the strategies essential for its realization through comprehensive analysis, in-depth discussions, and novel insights. We start by articulating the requisite capability frameworks for AGI, integrating the internal, interface, and system dimensions. As the realization of AGI requires more advanced capabilities and adherence to stringent constraints, we further discuss necessary AGI alignment technologies to harmonize these factors. Notably, we emphasize the importance of approaching AGI responsibly by first defining the key levels of AGI progression, followed by the evaluation framework that situates the status quo, and finally giving our roadmap of how to reach the pinnacle of AGI. Moreover, to give tangible insights into the ubiquitous impact of the integration of AI, we outline existing challenges and potential pathways toward AGI in multiple domains. In sum, serving as a pioneering exploration into the current state and future trajectory of AGI, this paper aims to foster a collective comprehension and catalyze broader public discussions among researchers and practitioners on AGI.

Figures (11)

• Figure 1: Overall Structure of This Paper. This paper starts with discussing core AGI components, including AGI Internal (\ref{['sec:agi_internal']}), AGI Interface (\ref{['sec:agi_interface']}), and AGI Systems (\ref{['sec:agi_systems']}); these discussions help us measure the ability of AGI and estimate how far we are from AGI. As we get closer to AGI, we further expect AGI to meet various constraints, which can be realized by AGI Alignment (\ref{['sec:agi_alignment']}) techniques. We further outline an AGI Roadmap (\ref{['sec:approach_agi_responsibly']}) that helps researchers approach AGI responsibly. Finally, some Case Studies (\ref{['sec:case_studies']}) are presented to illustrate the current development of early-stage AGI in various fields.
• Figure 2: Current State and Future Expectation of AGI Internal. We outline four major components for AGI Internal, the mind of AGI: Perception (\ref{['sec:perception']}), Reasoning (\ref{['sec:reasoning']}), Memory (\ref{['sec:memory']}), and Metacognition (\ref{['sec:meta-abilities']}), each of which consists of discussions of its current state and future expectation.
• Figure 3: There are three categories for multimodal models with LLM external connections: projection-based, query-based, and language-based.
• Figure 4: The Interconnected Spheres of AGI Interface. In the left part, we present some key elements in three interfaces: Digital (\ref{['sec:interfaces_digital']}), Physical (\ref{['sec:interfaces_physical']}), and Intelligence Interface (\ref{['sec:interfaces_intelligence']}). On the right side of the figure, we outline several potential future aspects that could be significant.
• Figure 5: Taxonomy of Current AGI Systems. We discuss several advancements in various categories of AGI systems, including scalable model architectures, large-scale training, optimized inference techniques, methods for reducing cost and improving efficiency, as well as next-generation computing platforms for AI.