[Position Paper] Beyond the AGI Myth: Transitioning to Superhuman Adaptable Intelligence (SAI)
AI Must Embrace Specialization via Superhuman Adaptable Intelligence
The paper introduces Superhuman Adaptable Intelligence (SAI), a new framework that rejects the ill-defined concept of AGI. It advocates for specialization over universal generality, defining intelligence by the speed of adaptation to new tasks rather than static performance benchmarks, aiming to surpass human capabilities in both human and non-human domains.
TL;DR
Is "AGI" a scientific goal or just a romanticized marketing term? This paper, co-authored by Yann LeCun, argues the latter. It posits that human intelligence is not "general" but specialized for survival, and AI should follow suit. By introducing Superhuman Adaptable Intelligence (SAI), the authors shift the goalpost from doing everything a human can do to learning to exceed humans at any task—especially those humans are evolutionarily incapable of performing.
The Illusion of Human Generality
The AI community is obsessed with AGI, yet we cannot even agree on its definition. Most definitions are "Cognitive Mirrors"—they try to replicate human-level performance. However, the authors point out a fundamental blind spot: Humans are not general-purpose thinkers.
Through the lens of Moravec’s Paradox, we see that what is "easy" for us (walking, spatial reasoning) is incredibly hard for AI, while "hard" tasks (chess, high-dimensional math) are trivial for silicon. We feel general only because we cannot perceive our own limitations. Magnus Carlsen is "good" at chess only compared to humans; compared to a simple specialized engine, he is objectively poor. If our benchmark is a limited, specialized biological brain, our AI progress will be unnecessarily throttled.
The Semantic Trap: Why AGI Definitions Fail
The paper categorizes current AGI definitions along two axes: Capability (Learn vs. Do) and Scope (Universal vs. Human-centric).
The authors argue these definitions fail three critical tests:
- Feasibility: The "No Free Lunch" theorem proves that no single algorithm can be optimal for all possible problems. A truly universal "general" intelligence is a physical impossibility under finite resource constraints.
- Consistency: Claiming an AI is "General" while only testing it on a "Human-important" subset of tasks is a logical contradiction.
- Assessability: Benchmarking against an "ever-growing list of tasks" is a Sisyphean task. We need a metric for the process of intelligence, not the output.
Methodology: Embracing the Specialist
The "Bitter Lesson" taught us that scaling compute beats hand-coded knowledge. But the authors clarify: Scaling does not mean Generality. Even as we scale, the most efficient systems will be those that specialize.
The SAI North Star
SAI defines intelligence as Adaptation Speed. Instead of a checklist of skills, we should measure how quickly an agent acquires a new, superhuman-level skill in a novel environment.
The Path: SSL and World Models
To achieve SAI, the paper advocates for two technical pillars:
- Self-Supervised Learning (SSL): To acquire the "generic" world knowledge required for fast adaptation without the bottleneck of labeled data.
- World Models (JEPA): Moving away from token-level (autoregressive) prediction. Autoregressive models suffer from exponential error divergence (see Figure below). By predicting in Latent Space, models can plan and reason about the world's dynamics without getting bogged down in pixel-level or token-level noise.
Why Specialization Wins
Biology favors specialists. Economics favors specialists. Machine Learning is no different.
- AlphaFold is a specialist that solved protein folding better than any "generalist" model ever could.
- Negative Transfer: Forcing a model to learn unrelated tasks often degrades performance on both.
As the authors memorably state: "The AI that folds our proteins should not be the AI that folds our laundry!"
The real value of AI is not in mimicking us (the "Human Domain"), but in filling the massive gaps in the "Universal Task Space" where humans are blind or weak (high-dimensional statistical inference, complex mechanistic modeling, etc.).
Critical Insight & Conclusion
The paper is a call to arms against the homogenization of AI research. The current dominance of autoregressive LLMs is a local optimum that may be leading us away from true intelligence. By pursuing Superhuman Adaptable Intelligence, we stop trying to build a "digital human" and start building a modular, hierarchical "system of systems" that can learn to master any domain—human or otherwise—with unprecedented speed.
Future Work: The industry must move toward objective-driven architectures that use predictive world models to plan. Only then can we move beyond "stochastic parrots" and toward agents that genuinely understand the causal structure of the universe.