Send us a text

The contemporary landscape of artificial intelligence is defined not by a linear ascent toward omniscience, but by a perplexing asymmetry. We stand at a juncture where foundational models—systems capable of passing the Uniform Bar Exam with 90th-percentile proficiency—simultaneously struggle to reliably stack physical blocks, maintain causal consistency over long conversational horizons, or perform simple arithmetic without error. This phenomenon, characterised by brilliance in abstract, evolutionary novel domains and incompetence in ancient, sensorimotor domains, challenges our deepest assumptions about the nature of intelligence itself.

This podcast is motivated by the recent discourse from Shane Legg, co-founder of DeepMind, regarding the "arrival of AGI". In his analysis, Legg highlights a critical measurement challenge: how do we define and quantify "general intelligence" when the capability profile of our most advanced agents is profoundly "jagged"? These systems do not fail in the predictable, brittle manner of traditional software; they fail probabilistically, often exhibiting what researchers describe as a "jagged technological frontier". Within this frontier, a system may act as a virtuoso creative partner one moment and a hallucinating fabulist the next, blurring the line between tool and agent.

The central thesis of this investigation is that these limitations—the "jaggedness" of current systems—are not merely engineering bugs to be patched by scale, but profound signals about the architecture of cognition. They serve as a mirror, reflecting the distinctions between crystallized intelligence (static knowledge access, where AI excels) and fluid intelligence (adaptive, embodied reasoning, where AI lags). By dissecting these capabilities through the frameworks of DeepMind’s "Levels of AGI" ontology and cognitive science theories such as Moravec’s Paradox and Dual-Process Theory, we can operationalize the path to Artificial General Intelligence (AGI).

Furthermore, this analysis addresses the reflexive inquiry posed by the user: What does the machine’s struggle tell us about the human mind? The fact that high-level reasoning (chess, mathematics) has proven computationally cheaper to replicate than low-level sensorimotor perception (walking, folding laundry) inverts the traditional hierarchy of intellectual value. It suggests that what humans perceive as "difficult" tasks are often evolutionarily recent and computationally shallow, while "easy" tasks are deep, ancient, and immensely complex adaptations.

In the following chapters, we will explore the transition from binary Turing Tests to nuanced, multi-dimensional ontologies. We will examine the empirical reality of the "jagged frontier" as revealed by recent Harvard Business School studies, the architectural gap between "System 1" generation and "System 2" reasoning, and the shift from static benchmarks to "living" evaluations necessary to track an intelligence that is universal in aspiration but alien in construction.