AGI is an increasingly useless concept
AGI made sense when we were imagining intelligence that's roughly human-shaped. But machine intelligence is jagged in ways that human intelligence isn't, and that jaggedness breaks the definition.
🌶️ Friday take: AGI is an increasingly useless concept. It made sense when we were imagining intelligence that's roughly human-shaped. But machine intelligence is jagged in ways that human intelligence isn't, and that jaggedness breaks the definition.
We're already past the Turing test, so now we're left with definitions of AGI that are more vibes than substance. AGI is effectively asking "are the computers as smart as humans" and there's really only two versions of that question that are useful: a) Is the AI as good as an average person across a wide range of domains b) Is the AI as good as an average human in every domain
The first one we've arguably already hit. Today's models operate across a wide range of domains at a level that would've been called AGI five years ago. And they're not just average, they're performing at the 99th percentile of humans. The second one is a category that will never be occupied. By the time an AI closes all the remaining gaps, it will inherently be superhuman in the dimensions where it's already superhuman. You don't pass through AGI, you skip it.
And ANI doesn't really work anymore either. Calling today's models "narrow" is like calling a Swiss Army knife a specialized tool. We're in a weird middle ground where the old taxonomy has broken down but we haven't replaced it yet. There is no term for a broad, flexible intelligence that is superhuman across a wide range of dimensions and falls far short of your average 12 year old in others.
The underlying question definitions of AGI are actually trying to answer is "can a computer fully substitute a humans without needing the help of other humans?" What seems increasingly clear to me is the answer is already yes to the first part and by the time it's yes to the second part it we'll be so far past human intelligence that AGI is an irrelevant waypoint.