Command Palette
Search for a command to run...
인공지능은 초인적 적응 능력을 바탕으로 특화되어 받아들여야 한다
인공지능은 초인적 적응 능력을 바탕으로 특화되어 받아들여야 한다
Judah Goldfeder Philippe Wyder Yann LeCun Ravid Shwartz-Ziv
초록
인공지능(AI) 임원과 연구자들부터 최후주의자들, 정치인, 활동가들까지 모두 인공지능 일반화(Arificial General Intelligence, AGI)에 대해 이야기하고 있습니다. 하지만 그들은 종종 AGI의 정확한 정의에 대해 합의하지 못하는 것처럼 보입니다. AGI에 대한 일반적인 정의는 인간이 할 수 있는 모든 것을 수행할 수 있는 AI를 의미합니다. 그러나 인간은 정말로 ‘일반적(general)’일까요? 본 논문에서는 우리의 AGI 개념이 지닌 문제점과, 가장 일관성 있는 형태로 정의되더라도 AGI가 AI의 미래를 설명하는 데 있어 결함이 있는 개념임을 지적합니다. 우리는 가장 널리 수용된 정의들이 타당성, 유용성, 그리고 진정한 일반성을 갖추었는지를 탐구합니다. 우리는 AI가 일반화보다는 전문성을 수용해야 하며, 그 전문성 속에서 초인적 성능(superhuman performance)을 지향해야 한다고 주장합니다. 이에 따라 초인적 적응 지능(Superhuman Adaptable Intelligence, SAI)을 도입합니다. SAI는 우리가 중요하게 여기는 어떤 분야에서도 인간을 능가하도록 학습하고, 인간이 감당할 수 없는 기술적 격차를 메울 수 있는 지능으로 정의됩니다. 이어 우리는 SAI가 과잉 정의된 AGI 개념으로 인해 흐려졌던 AI 논의를 어떻게 정교하게 다듬을 수 있는지, 그리고 이를 미래의 지침으로 사용할 때 도출될 함의들을 추적합니다.
One-sentence Summary
Researchers from New York University and Meta AI challenge the prevailing notion of Artificial General Intelligence, arguing that AI should prioritize specialization and superhuman performance, and propose Superhuman Adaptable Intelligence (SAI), a framework for intelligence that learns to surpass humans in critical tasks and fill human skill gaps, to guide future AI development.
Key Contributions
- Proposes Superhuman Adaptable Intelligence (SAI), defined as the capacity to rapidly adapt to important tasks inside and outside the human domain, learn to exceed human performance, and fill skill gaps where humans are incapable, replacing the human-centric notion of AGI.
- Reframes progress measurement around the speed and efficiency of skill acquisition under realistic resource constraints, shifting evaluation from static human benchmarks to measurable adaptation dynamics.
- Shows that embracing SAI counters the homogenization of autoregressive models by promoting architectural diversity and specialization, and identifies self-supervised learning, predictive world models, and modular composition as promising routes to fast, reliable competence.
Introduction
The authors argue that the widely used notion of Artificial General Intelligence (AGI) is deeply ambiguous, with conflicting definitions that fuel polarized debate and conflate generality with a human-centric skill set. This ambiguity misdirects research by treating human intelligence as a universal benchmark, even though human cognition is itself a collection of specialized adaptations shaped by evolutionary constraints. As a result, progress is often measured against static, human-level task checklists rather than an agent’s ability to acquire new competence rapidly under realistic resource limits. To address this, the authors propose Superhuman Adaptable Intelligence (SAI), a guiding concept that shifts focus from an ill-defined "generality" to measurable adaptation speed and efficiency for tasks both inside and outside the human domain, explicitly embracing specialization, self-supervised learning, world models, and architectural diversity.
Experiment
The evaluation examines why existing AGI definitions fall short against three core criteria: feasibility, internal consistency, and assessability. Definitions claiming true generality violate feasibility due to the No Free Lunch theorem, while those centered on human-like generality are internally inconsistent because human intelligence is only a narrow subset of possible intelligence. Performance-oriented definitions further lack clear progress metrics, whereas learning- or adaptation-focused definitions naturally provide measurable evaluation through speed of adaptation. Overall, the analysis argues that imprecise semantics around "generality" risk misleading the field toward overly narrow goals and obscure practical paths to realization.