Abstract
Everyone from AI executives and researchers to doomsayers, politicians, and activists is talking about Artificial General Intelligence (AGI). Yet, they often don't seem to agree on its exact definition. One common definition of AGI is an AI that can do everything a human can do, but are humans truly general? In this paper, we address what's wrong with our conception of AGI, and why, even in its most coherent formulation, it is a flawed concept to describe the future of AI. We explore whether the most widely accepted definitions are plausible, useful, and truly general. We argue that AI must embrace specialization, rather than strive for generality, and in its specialization strive for superhuman performance, and introduce Superhuman Adaptable Intelligence (SAI). SAI is defined as intelligence that can learn to exceed humans at anything important that we can do, and that can fill in the skill gaps where humans are incapable. We then lay out how SAI can help hone a discussion around AI that was blurred by an overloaded definition of AGI, and extrapolate the implications of using it as a guide for the future.
The Problem: The Illusion of Generality
The AI community is currently fractured by debates over AGI, yet there is no consensus on what the term actually means. Most definitions benchmark AI against human intelligence, assuming humans are "general" beings. This paper argues this is a "circular" error: we define generality in human terms and then declare humans the paradigm of generality.
In reality, humans are specialized for a narrow sliver of tasks relevant to biological survival. We suffer from an "illusion of generality" because we are blind to the infinite space of tasks we cannot perform, such as high-dimensional statistical inference or echolocation. By chasing "AGI," the field risks pursuing a goal that is theoretically unfeasible due to the "No Free Lunch" theorem, internally inconsistent, and difficult to measure.
Re-evaluating AGI Definitions
To address this lack of clarity, we provide a theoretical framework for re-evaluating AI goals. We categorize existing AGI definitions using a two-dimensional semantic map based on two axes:
- Capability: Learning/Adaptability (the ability to acquire new skills) vs. Performance (the ability to perform tasks out of the box).
- Scope: Universal (anything) vs. Human-centric (anything humans do).
These definitions are then evaluated against three criteria: Feasibility, Internal Consistency, and Assessability. Our analysis reveals a "Taxonomy of Failure" among major industry definitions.
Key Insights: Specialization and Adaptation
Mathematically, the "No Free Lunch" theorem dictates that no single algorithm can be optimal for every possible problem. An AI that tries to be "good at everything" will inevitably be outperformed by specialized systems that allocate their finite computational energy more efficiently. Therefore, specialization is a feature, not a bug; it allows AI to fill the "blind spots" of human cognition. As the paper notes: "The AI that folds our proteins should not be the same AI that folds our laundry!"
The true measure of intelligence is not a checklist of static skills, but the speed and efficiency with which an agent acquires new skills. Current autoregressive models (like LLMs) suffer from exponential error divergence at the token level. To achieve true adaptability, the field must move toward latent prediction and world models to enable better planning and zero-shot transfer.
Results: The SAI Framework
Our evaluation of prominent AGI definitions shows that most fail to meet the criteria of being consistent, feasible, or assessable. For instance, definitions focused on "economically valuable work" are labeled "Not Assessable" because the benchmark is an ever-growing, ill-defined set of tasks.
In response, we propose Superhuman Adaptable Intelligence (SAI) as a replacement framework. SAI is formally defined as intelligence capable of adapting to exceed humans at any task humans can do, while also adapting to tasks outside the human domain that have high utility. This reframes the "North Star" of the field from mimicking humans to complementing them through superior adaptability in domains where biological intelligence is weak.
Limitations and Future Work
While SAI provides a more coherent goal, several challenges remain. The definition relies on tasks having "utility," but precisely defining how to determine task importance (economic vs. societal value) is ongoing work. Additionally, while Self-Supervised Learning and World Models are promising technical pathways, the specific architecture—the "substrate for fast adaptation"—that allows for superhuman learning speeds remains an open research question. Future work must focus on creating standardized, practical metrics for "adaptation speed" across diverse, non-human domains.
Summary updated