Moving toward AGI: bridging the gap between text and understanding

The quest for understanding

Recent breakthroughs in generative AI have revolutionized machine-generated text, enabling the production of coherent and context-rich narratives. While exceptional fluency, achieved through token prediction, is a key aspect of intelligence, it is not synonymous with understanding. As David Deutsch argued in ‘The Fabric of Reality’ and later in ‘The Beginning of Infinity’, genuine knowledge is not merely a collection of facts but is built on the ability to explain why things happen. In other words, true understanding requires developing deep, explanatory models that capture causal mechanisms rather than simply regurgitating patterns.

At the heart of Deutsch’s thesis is the notion that explanation is the sole means by which knowledge becomes deep and reliable. He asserts that possessing a vast repository of facts or, in the case of LLMs, statistically generated tokens, cannot substitute for an understanding of underlying causes. Applied to AI, this perspective implies that while current models excel at pattern matching, they lack the ability to generate the explanatory theories needed to reason about why something happens.

Enhanced reasoning

Recent developments have introduced reasoning-centric models designed to “think” before responding. Advances in chain-of-thought prompting (first demonstrated by Wei et al. 2022) and later refined through tree-of-thought and self-consistency approaches have markedly improved large language models’ capacity to decompose complex problems into intermediate reasoning steps. Emerging frameworks like the causal chain of prompting aim to integrate causal inference into LLMs, empowering them to identify and reason about cause-effect relationships. These methods allow models to “reflect” on their outputs and correct errors, thereby moving closer to the human-like reasoning that Deutsch identifies as essential to knowledge. While this synthesis enhances accuracy and advances AI systems toward producing outputs with explanatory depth, these improvements remain superficial if they do not lead to genuine explanations that clarify why a phenomenon occurs.

Yann LeCun’s ‘The Shape of AI to Come!’

During the AI Action Summit in Paris (2025), Yann LeCun delivered a speech that resonated with this evolving narrative. LeCun underscored that while current LLMs generate impressively fluent language, they primarily “fill in the blanks” without true reasoning or planning. He notably remarked that, in many respects, these systems are even less intelligent than a cat, emphasizing their lack of common sense and explanatory ability. LeCun argued that for AI to progress toward genuine intelligence, it must build an internal model of the world that incorporates causal reasoning and physical intuition. In his view, echoing Deutsch’s philosophy, true understanding requires AI systems to form explanations by integrating sensory inputs, learning from interactions, and employing causal reasoning to predict outcomes. This vision, which he refers to as “Objective-Driven AI,” calls for architectures that learn, reason, and plan rather than merely mimic human language.

From prediction to explanation

While generative AI and LLMs represent significant technological milestones, they remain, at their core, prediction machines. As Deutsch argues, without the capacity to explain, knowledge remains shallow. Even as we strive to equip AI with explanatory abilities, a critical question arises: can genuine explanation and true intelligence be achieved solely within the realm of text? Today’s LLMs are fundamentally text-bound systems. Their understanding is derived entirely from the vast corpus of human text, a dataset that, although rich, represents only a mediated, abstract view of the world. A key limitation is the absence of direct, embodied interaction with the physical world. Human intelligence is deeply intertwined with sensory experience, physical action, and feedback from real-world consequences. LLMs, lacking this embodied grounding, struggle to develop the common sense, intuition, and robust causal models that characterize human understanding. They can simulate reasoning patterns found in text, but this simulation lacks grounding in lived experience.

Toward AGI: beyond text

In this context, the trend toward Objective-Driven AI represents a critical step in the evolution toward artificial general intelligence (AGI). This evolution is not merely about scaling up models but about transforming them from pattern matchers into systems capable of developing deep, causal, and explanatory models of the world.

Bridging the gap between statistical prediction and genuine understanding is essential for AI’s future advancement. Drawing on David Deutsch’s insights, true knowledge is built on explanation and causal reasoning, not merely on the accumulation of facts. However, to realize AGI, we must move beyond text-based learning integrating reasoning techniques with multimodal sensory input, embodied interaction in real or simulated environments, and learning through action and consequence. Only by addressing the fundamental limitation of text-bound understanding can AGI become attainable, ushering in AI systems that not only perform tasks and explain their actions but also possess a deeply grounded, adaptable intelligence rooted in the complexities of the physical world.