"I think we've achieved AGI." That's Nvidia CEO Jensen Huang, speaking on the Lex Fridman podcast this Monday, adding his name to the growing list of tech leaders who've decided the goalpost has arrived at their feet.
AGI, or artificial general intelligence, is supposed to describe an AI system that can match or exceed human-level reasoning across any intellectual task. It's the big milestone the industry has been chasing for decades. The problem is that nobody agrees on what it actually means, which makes it very easy to claim you've reached it.
The Definition Game
Huang's declaration lands in a year where the definition of AGI has been stretched, squeezed, and reshaped by nearly every major figure in AI. OpenAI's Sam Altman has hinted at similar conclusions. Demis Hassabis at Google DeepMind has offered his own goalposts. Each version conveniently matches whatever the speaker's company just shipped.
The pattern is clear: as AI models get better at specific benchmarks, the people building them start claiming those benchmarks were the AGI threshold all along. A year ago, passing the bar exam was the marker. Before that, it was coding ability. Now that current models can do both, the claim shifts from "we're approaching AGI" to "we've arrived."
For Huang specifically, there's an obvious business angle. Nvidia sells the GPUs that power virtually all AI training and inference (the process of actually running a model to get answers). If the industry believes AGI is here or imminent, demand for Nvidia hardware stays white-hot. Every AGI declaration is, in part, a sales pitch for more H100s and B200s.
What Users Actually Experience
Here's the disconnect. If you use ChatGPT, Claude, or Gemini daily for real work, you know these tools are genuinely impressive and genuinely unreliable in roughly equal measure. They can draft a solid marketing email, then hallucinate a citation that doesn't exist. They can write functional code, then confidently explain why a bug fix works when it doesn't.
That's not general intelligence. That's a very capable pattern-matching system with no ability to verify its own output. The gap between "passes standardized tests" and "reliably handles novel problems without supervision" remains wide.
None of this means current AI isn't useful. It is, enormously so. But calling it AGI sets expectations that the technology can't meet, which leads to the exact kind of over-reliance and under-verification that causes real problems in production.
The Real Question Nobody's Asking
Instead of debating whether we've hit AGI, a more productive question is: does the label change anything practical? If you're a marketer using AI to draft campaigns, the model's capability is the same today as it was yesterday, regardless of what Jensen Huang calls it. If you're a developer using Copilot, the autocomplete doesn't get smarter because a CEO made a podcast appearance.
The AGI discourse has become a status game among tech executives, not a meaningful technical benchmark. When every major AI company CEO claims AGI on their own terms, the term stops meaning anything at all. What matters is whether your specific tool solves your specific problem reliably. Right now, the answer is "sometimes, with supervision" - and no amount of relabeling changes that.