We have AGI - Artificial General Intelligence

Artificial General Intelligence: The Goalposts Have Already Been Crossed

AI leaders like Andrej Karpathy and Sam Altman keep suggesting that AGI (Artificial General Intelligence) is 5 to 10 years away. But if you look closely at the capabilities of today’s large language models (LLMs), you have to ask a fundamental question: Is our definition of AGI broken?

For decades, we’ve used traditional metrics to define "intelligence," and current LLMs are smashing them.

The Broken Intelligence Benchmarks

For most of the last 50 years, the pursuit of General Intelligence centered on two key, easily testable standards. Today, both are functionally obsolete as measures of AI readiness.

1. The IQ Test

Current models like GPT-5, Claude 4.5, and Grok are consistently scoring in the 120–130 IQ range on standardized tests. This isn’t just general adult human intelligence; this is the threshold of genius.

If the accepted metric for analytical intelligence says an AI is superior to 90% of the human population, why is AGI still considered a decade away? Furthermore, modern models are excelling in emotional intelligence and creative writing.

2. The Turing Test

Five years ago, fooling a human judge in a sustained text conversation was the ultimate goal. Today, the Turing Test is functionally crushed. Virtually every state of the art LLM can pass it with ease. If the benchmark for "human-level conversation" is this easy to beat, the test is now irrelevant to the AGI debate.

AI is already demonstrably "smarter" than an animal or a human child. If we’ve beaten the old tests and surpassed simpler forms of general intelligence, what, exactly, are the experts waiting for?

The Real Bar: Superintelligence, Not AGI

The experts pushing the 5–10 year timeline; they are looking at a goal far beyond average adult general intelligence. The capabilities they demand are those that even the vast majority of humans struggle with:

  • Long-term, Proactive Planning: Identifying and executing complex, multi-stage goals over weeks or months, without constant prompting or external correction.

  • Novel Problem Solving: Developing a working, unrehearsed solution to an entirely new conceptual or physical problem (e.g., inventing a new material or solving a complex political crisis).

  • Deep Contextual Reasoning: Consistently understanding the subtle social, emotional, and infrastructural context of a situation, and acting with perfect foresight.

These are not barriers for Artificial General Intelligence; they are barriers for Artificial Superintelligence.

We mistake AGI for something that is perfect, but "general intelligence" means mirroring the average human, including our flaws, super intelligence is about exceeding us, which I agree is 5-10 years out.

Humans are notoriously bad at long-term planning. We constantly struggle with delayed gratification, budgeting for retirement, and sticking to multi-month projects. The fact that an AI sometimes needs a nudge to stay on task simply means it shares a failing common to human general intelligence.

Humans are rarely novel problem solvers. Novelty is so rare and valuable that we created the patent system to provide decades of financial reward just to incentivize it. If novelty was a requirement for general intelligence, few humans would qualify.

We all struggle with contextual reasoning. Social awkwardness, poor reading of the room, and emotional missteps are hallmarks of human interaction. If AGI requires perfect contextual reasoning, then almost no adult human possesses general intelligence.

Conclusion: We Have Achieved AGI

If an entity can master the world’s knowledge, achieve a genius-level IQ, and hold a flawless conversation, it has achieved General Intelligence. The fact that it occasionally needs a goal reminder or assistance with a truly unprecedented task does not negate its status as "general."

The experts haven't failed to create AGI; they have simply shifted the target to ASI (Artificial Superintelligence)

AGI is here. It’s just not perfect, and that is precisely what makes it general.

LARG*net