For decades, artificial general intelligence — a machine that can think, reason, and learn across any domain as well as a human — existed only in science fiction. That may be about to change.
OpenAI, Google DeepMind, Anthropic, and a growing list of well-funded startups are all racing toward the same goal: a system that doesn't just excel at one task, but can transfer knowledge fluidly across disciplines, reason through novel problems, and adapt to new situations without explicit retraining.
The Benchmark Problem
Traditional AI benchmarks — standardized tests that measure performance on specific tasks — are increasingly inadequate for measuring progress toward AGI. Systems routinely achieve superhuman performance on these tests while failing at basic reasoning tasks that any child could handle.
In response, researchers have developed more challenging evaluations. The ARC-AGI benchmark, designed to test fluid intelligence rather than memorized knowledge, has proven remarkably resistant to AI systems. Until recently, the best models scored below 50% — roughly where an average human sits.
The Architecture Wars
Behind closed doors, the major labs are pursuing fundamentally different architectural approaches. Some are betting on scaling existing transformer architectures to unprecedented sizes. Others are experimenting with hybrid systems that combine neural networks with symbolic reasoning engines. A third camp believes the answer lies in better training paradigms — teaching systems to learn how to learn, rather than simply memorizing vast amounts of data.
The stakes couldn't be higher. Whoever achieves AGI first will possess a technology that could compress decades of scientific progress into years — accelerating drug discovery, materials science, climate modeling, and virtually every other field of human endeavor.
The Safety Question
Not everyone is racing forward with equal enthusiasm. A growing contingent of AI safety researchers argues that the rush to AGI is dangerously reckless. Without robust methods for aligning a superintelligent system's goals with human values, they warn, the consequences could be catastrophic and irreversible.
The debate has split the AI research community. Some labs have implemented voluntary pause agreements on the most powerful training runs. Others argue that safety concerns are best addressed by having safety-focused organizations reach AGI first, rather than ceding the field to less scrupulous actors.
What's clear is that we are entering uncharted territory. The decisions made in the next few years — about who builds AGI, how they build it, and what safeguards they put in place — may be the most consequential choices in human history.