The creation of intelligence: building complex agents.

We stand at the cusp of the next evolutionary step in AI: the creation of complex, generative agents. Compared to simple chatbots, generative agents are capable of autonomous decision-making, actions, learning, and interactions in immersive environments. 

Last month, we launched Suck Up: Love Bites, a game set in the season of love. In the game, we had dozens of AI agents that embodied marquee guests of the season: from Romeo to Elizabeth Bennett. As guests mingled in the festival, players played as the anti-Cupid, spreading chaos and breaking up couples. It was a lot of irreverent fun for our players. 

The premise of “convincing someone to break up” sounds simple, but is actually a lot more difficult than we might think. Consider if someone tried to talk you out of your relationship. You’d likely dismiss them out of hand. But what if that someone was your best friend? What if you were already in a bad mood?  In other words, our decisions are based on a multitude of factors (trust, logic, motivation, emotions, etc.) in a chain of complex decision making. It is what makes us human. But how do we simulate this type of decisioning with AI?

The human brain evolved over billions of years. From the more ancient parts of the brain that deal in emotion, trust and memory, to the more recent parts that deal in logic and higher order reasoning, the brain makes decisions by integrating many layers of complex circuitry. The emergence of ‘humanness’ comes through the balance of these distinct nodes. To create artificial life, we need to recreate this ‘emergence’.

Let’s go back to the Valentines event. We created two layers of cognition in our agents. Two distinct nodes, if you will. The first node assesses trust: do I trust the stranger enough to believe him? This decision is primal and instinctive. The second node assesses credibility and belief: are the arguments convincing enough for me to break up? The two nodes work in tandem for the agent to make her decisions.

This ‘node’ based approach applies not only for decisions, but also for actions. In one of our prototypes, we created an AI pet capable of engaging with the world. To achieve this, we built two distinct nodes: one for ‘intent’ one for ‘action’. The first node (intent) considers inputs like player command, perception, and memory to create an intention. The second node translates this intention into executable actions: execute run from A to B, execute grab, and so on. This is not so distinct from our brain, in which the intent formed from the prefrontal cortex is translated into muscle instructions in the motor cortex. 

As we move from “simple” AI applications like chatbots to complex generative agents, we will likely build in ways that mirror the human brain. 

The creation of life-like intelligence is just getting started. To quote Dr. Jim Fan of NVIDIA: “Games provide the ‘primordial soup’ for generalist AI to emerge. Games (and simulation in general) will provide the next trillion high-quality tokens to train our foundation models. Before we have a million robots in the physical world, we will first see a billion embodied agents in virtual worlds.”

Two things I’m certain: First, the creation of lifelike intelligence will be the most consequential technology advancement of the next 20 years. Second, that creation will begin in games.