The AGI Benchmark: Not the Turing Test, but Valorant
Henry Kim ·
Yesterday, I talked with my kids about Valorant, a first-person shooter game. To my kids, who are deeply engrossed in the game, I made a half-serious, half-joking boast with a feigned solemn expression. I told them that after this year, an era would come where AI would move the mouse for them while they slept, earning top ranks like Radiant or Immortal.
Contrary to my expectations, my children's reaction wasn't one of futility, lamenting the loss of the game's essential fun. Instead, their eyes lit up. It was a desire to be the first to use it, before anyone else. They instinctively understood the power of technological disparity: if everyone could use it, it would be worthless, but if only they could use it, they would be treated like heroes at school.
While I said such a world would come soon, from a cold, technical perspective, there's still a long way to go. Generating text and writing code is a fundamentally different problem from recognizing an enemy in a fleeting 0.01 second, precisely moving the mouse cursor, and landing a headshot. Current AI has an oversized brain but sluggish limbs. This is because there's a barrier of reaction speed, or Latency, where the output of thought translates into physical action.
Therefore, I dare to propose a new criterion for judging the advent of AGI (Artificial General Intelligence). There's no need for complex Turing Tests or philosophical debates. Just throw any game at an AI. It doesn't matter if it's Valorant or an indie game it's never seen before. The day AI can perceive the screen, grasp the rules on its own, and demonstrate physical prowess that overwhelms human pro gamers in just a few rounds – that very day will be the day true AGI arrives.
The reason this is an important indicator is clear. It's not just about learning from data and producing correct answers, but about the ability to grasp principles independently when thrown into an unfamiliar environment (a game), learn through mistakes, and translate that learning into real-time actions. This is akin to how we learn to ride a bicycle not by reading instructions, but by falling and learning through physical experience. 'Embodied AI' is truly the final gateway to human intelligence.
Current AI Agent models, which are just beginning to mimic holding a mouse, are still at an infant stage. They are crawling around, bumping into walls. But the moment this infant stands up and starts running, its speed will be unimaginable to humans. Curious when AGI will arrive? Then, instead of searching the news, look at your monitor screen. Throw any game at that AI Agent.