Artificial General Intelligence (AGI)

Artificial General Intelligence (AGI) refers to AI systems capable of performing any intellectual task that a human can — reasoning across domains, learning from limited examples, transferring knowledge between contexts, and operating with the kind of flexible, adaptive cognition that current narrow AI systems approximate but do not fully achieve. The term has become one of the most contested in technology discourse: simultaneously the holy grail of AI research, a marketing term, a safety concern, and an increasingly slippery definitional problem as the capabilities of large language models and AI agents continue to accelerate.

The definitional problem is real and consequential. Google DeepMind published a framework in 2023 proposing six levels of AGI, from "Emerging" (equal to or somewhat better than unskilled humans) through "Competent," "Expert," "Virtuoso," to "Superhuman" (outperforms all humans). By this taxonomy, current frontier models like Claude, GPT-4, and Gemini arguably sit at Level 1 or 2 for many tasks — performing at or above average human level across a broad range of domains. OpenAI's internal definition reportedly tied AGI to systems that can do "the work of a senior software engineer," while other researchers insist on embodiment, autonomous goal-setting, or the ability to conduct novel scientific research as prerequisites. The lack of consensus means that AGI announcements are as much about definition as about capability.

The ARC-AGI benchmark, developed by François Chollet and the ARC Prize Foundation, represents one of the most rigorous attempts to measure general reasoning — testing whether models can solve novel visual puzzles they've never encountered during training. By February 2026, Google's Gemini 3.1 Pro achieved 77.1% on ARC-AGI-2, a significant leap from baselines of 50–60% just a year earlier, demonstrating measurable progress on tasks designed to resist pattern memorization.

An increasingly common view — particularly among practitioners building with agentic AI tools — is that AGI has effectively arrived, not as a single omniscient model but as the compositional architecture of agents, tools, and human direction working together. Jon Radoff has argued that agentic engineering in Claude Code 4.5+ is functionally AGI: the tight loop of human intent, AI execution, feedback, and iteration produces general-purpose capability that exceeds what any individual — human or AI — can achieve alone. This is a structural rather than model-centric view: AGI emerges from the system, not the weights. In his State of AI Agents and Agentic Engineering 2026, Radoff describes the evolution from "vibe coding" to structured agentic engineering as the practical realization of general artificial capability — where multi-agent systems with experiential feedback loops create convergent intelligence that neither component achieves independently.

This view remains controversial. Skeptics argue that current systems lack genuine understanding, autonomous goal formation, and the ability to operate without human scaffolding. Andrej Karpathy's 2025 pivot — acknowledging that unconstrained AI generation isn't production-ready and needs structured oversight — cuts both ways: it confirms that raw model capability isn't sufficient, but also that human-AI composition produces outcomes neither achieves alone. The academic consensus, to the extent one exists, tends to emphasize that today's systems are remarkably capable pattern matchers and reasoners that still lack certain properties associated with general intelligence — though the list of missing properties shrinks with each model generation.

The distinction between AGI and Artificial Superintelligence (ASI) matters for policy and safety. AGI implies human-level capability; ASI implies capability that exceeds all human performance. Much of the existential risk discourse focuses on the transition from AGI to ASI — the concern that once a system reaches human-level general intelligence, recursive self-improvement could rapidly push it beyond human comprehension and control. Whether you believe AGI has already arrived in agentic form or remains years away, the governance challenge is the same: how do we build institutions and safeguards adequate to systems whose capabilities are advancing faster than our frameworks for understanding them?