What OpenAI's box surfing AI agents tell us about the future of AI.
I’m obsessed with digging back into the YouTube archives.
Having the superpower of hindsight gives you a profound perspective on conversations with founders in the early days of their company, investors in the midst of a financial downturn, and athletes at their highest highs and lowest lows. You get to listen to the conversation with a powerful lens—a lens into the future.
In addition to interviews, the YouTube channels of companies are a gold mine of content. OpenAI’s is one of my favorites.
Before gaining widespread recognition for ChatGPT, OpenAI was primarily (and still, to a large extent, is) an AI research institution.
A quick dive into their YouTube channel archives reveals their humble but forward-thinking roots. From demo presentations from their hackathon in 2018 to the infamous robot hand solving a Rubik’s Cube, there is some incredible content that gives you an unfiltered view into OpenAI’s journey since inception. It's also a reminder of just how fast AI has evolved in such a short span.
Multi-Agent Hide and Seek
My favorite video on OpenAI’s channel is Multi-Agent Hide and Seek. In this experiment, OpenAI researchers wanted to see how well virtual agents could play a game of hide-and-seek. What they discovered was fascinating.
The experiment involves agents in a computer game where some try to hide (the "hiders" in blue) and others try to find them (the "seekers" in red). To make it more interesting, there are objects scattered around the game that the agents can grab and use to their advantage.
The agents were simply given an objective (for hiders, it was to evade the seekers; for seekers, it was to find the hiders) with no further instructions.
Surprisingly, the agents began to figure out creative ways to play—all on their own.
Through trial and error over millions of games of hide-and-seek, they developed six different strategies—some of which were completely unexpected.
A Winning Strategy: Box Surfing
One of those unexpected strategies was “box surfing.” As the hiders became more advanced, they learned to lock the ramps inside the enclosure with them so they could not be used by the seekers to infiltrate the enclosure. To combat this, the seekers learned to jump on the boxes and “surf” them over the barriers of the enclosure to find the hiders.
Despite the agents not being explicitly programmed to surf boxes, they had autonomously discovered this strategy, showcasing what is known as an emergent capability.
Emergent Capabilities
Whether an agent or a plain old LLM, AI has an incredible feature: emergent capabilities. Emergent capabilities are unexpected skills or behaviors that emerge during the training process.
In the case of Multi-Agent Hide and Seek, “box surfing” was an emergent capability. The only instructions provided to the seeking agents were to find and tag the hiding agents. Through trial and error, they learned to use the resources at their disposal as tools to accomplish their objective.
While box surfing may seem like a relatively straightforward instance of an emergent capability, it hints at the large potential and unpredictability of AI's future.
The Future of AI
As AI systems become smarter, what emergent capabilities will we discover? What happens when instead of learning to box surf, agents learn to bypass cybersecurity protocols? Or coerce a group of people online? On a more positive note, maybe these agents will be able to cure diseases or create more sustainable materials with emergent capabilities.
As we continue to push the boundaries of AI, we'll uncover new emergent capabilities that will transform industries and society in ways we can't yet imagine. The next chapter of AI's story is just beginning, and one thing is for sure—it will be anything but boring.
Bonus: Check out the YouTube comments on the video for some Grade-A entertainment.
See you in the future,
Bennie 3
What is 𝐖𝐀𝐈𝐓, 𝐎𝐍𝐄 𝐌𝐎𝐑𝐄 𝐓𝐇𝐈𝐍𝐆?
1x per week, I send out one interesting thing I came across in the world of tech.
That’s right, just one. The message is short and sweet—a 30-second read. I share products, demos, Tweets, thoughts, announcements, articles, and more.
Subscribe to get WAIT, ONE MORE THING straight to your inbox.