A Definition of AGI

 

A Definition

For years, debating Artificial General Intelligence (AGI) has been like trying to nail fog to a wall—a frustrating exercise in hype and shifting definitions. It's been nearly impossible to pin down what AGI truly is, let alone how close we are to achieving it.

Now, a new academic paper, "A Definition of AGI," aims to change that. Authored by a large team of prominent researchers (including AI pioneers like Yoshua Bengio and notable critics like Gary Marcus, representing a broad consensus), the paper introduces a concrete, measurable framework for AGI. This article breaks down the most impactful and surprising takeaways from this groundbreaking research.

1. The Fuzzy Idea of AGI Finally Gets a Concrete Definition

The paper proposes a refreshingly clear definition for AGI: matching the cognitive versatility and proficiency of a well-educated adult.

Having a clear, human-centric benchmark is a game-changer. By grounding the goal in familiar human terms, this definition strips away the sensationalized sci-fi notions of god-like, superhuman entities. It moves the concept of AGI out of the realm of philosophical debate and transforms it into a tangible engineering goal. And like any engineering goal, the next logical step is to measure it.

2. We Can Now Put a Score on AGI—And the Numbers Are In

The framework isn't just a definition; it's a rigorous scoring system. To create it, the researchers grounded their methodology in the Cattell-Horn-Carroll theory, the most empirically validated model of human cognition. The system dissects general intelligence into ten core cognitive domains—like reasoning, memory, and perception—to create a composite AGI score.

Using this new yardstick, the paper assigned AGI scores to leading AI models. The results are eye-opening:

• GPT-4: 27%
• GPT-5: 58%

From an analyst's perspective, these numbers reveal two critical trends: the blistering pace of progress, with the score more than doubling in a single model generation, and the stark reality of the substantial gap that remains before any system can claim to reach the 100% human-level benchmark.

3. Modern AI Has a "Jagged" and Unbalanced Mind

One of the most important findings from the paper is that current AI systems have a highly "jagged" cognitive profile. Imagine a student who can write a perfect doctoral thesis but cannot remember what they had for breakfast. This is the "jagged" and unbalanced mind of today's AI.

This is highlighted in the paper's abstract:

While proficient in knowledge-intensive domains, current AI systems have critical deficits in foundational cognitive machinery...

This insight is crucial because it counters the popular image of AI as a uniformly powerful intelligence. Instead of a smooth, upward curve of capability, the reality is a jagged landscape of specific strengths and profound weaknesses. This unevenness begs the question: where are the most profound weaknesses?

4. The Biggest Weakness Isn't Reasoning, It's Memory

The paper directly answers that question, and its finding is deeply counter-intuitive. Most of the public discourse on AGI's final hurdles focuses on esoteric concepts like consciousness or advanced creativity. The research, however, points to a much more fundamental bottleneck: one of the most critical deficits is in long-term memory storage.

This suggests that before AI can achieve true general intelligence, it must first master a foundational cognitive function that humans often take for granted. This specific, almost mundane weakness in memory is a major obstacle holding back today's most advanced models from reaching the next level of intelligence.

Conclusion

For the first time, the AI community has a clear, quantifiable yardstick to measure its progress toward AGI. This research cuts through the noise, revealing the core tension of the current moment: astonishingly rapid progress colliding with surprisingly basic cognitive roadblocks. We can now see both how far we've come and the precise territory we still need to cover.

The score for AGI more than doubled in a single year. But with the remaining gap defined by a fundamental challenge like memory, will the next 42% be the hardest part of the journey?