Artificial Intelligence and the Search for Creativity

What Emergence, Multiplayer Games and AI show us about Creativity

Intelligence and creativity are challenging terms because they’re so wrapped-up in our notions of human exceptionalism. In this article, I hope to connect the two with an idea that both are more about finding solutions within a “search space” of infinite possibility.¹

To get there, I need to weave together several topics: emergence, multiplayer games and virtual worlds, composability, complexity theory and the recent innovations in generative agents.

Let’s start with games.

Emergent Gameplay

Emergence is a quality well-known to game makers: from a set of simple underlying rules, complex systems may emerge. This is evident in Conway’s “Game of Life” where a grid of cells express complex behaviors.

You can configure the grid however you like (i.e., you bring your own creativity to the game). The grid then undergoes a series of generations. Each is subject to the following rules: any cell with 2-3 neighbors survives, dead cells with 3 neighbors become alive, and anything else dies (try a playable version here).

Here’s the really interesting thing about the Game of Life: it is a Turing-complete computational system. That means it has enough capabilities to solve any computational problem. That doesn’t mean it is easy to configure it into a computer program—but it is possible.

Searching for computational answers from inside the Game of Life is cumbersome because of the user interface; the AND and OR gates are tricky to set up and difficult to “program.” But it is possible! That’s what emergent complexity is all about.

A few things can accelerate the emergence of useful properties within a system: social structures that allow the integration of outside knowledge and creativity, good user interfaces, and structures that facilitate composability.

Social Inputs Accelerate Emergence

As more inputs are available to the system, it is possible for the game to become far more complex. This is what made roleplaying games like Dungeons & Dragons so compelling—the emergent complexity came from the ability for players to add their own creativity and storytelling to the experience.

Side note: good games are usually those that don’t overwhelm the player with this complexity within the basic rules—otherwise the game becomes too hard to learn. But when the learning curve is balanced with complexity that’s more emergent in nature, it often makes for long-term fun as players continuously learn new forms of interaction with the environment.

Multiuser dungeons, virtual worlds and then massively multiplayer online roleplaying games (MMORPGs) added even more emergent complexity: they scaled-up the number of players and their network of social interactions.

The Corrupted Blood” incident in World of Warcraft is a well-document example of emergent gameplay: not only due to the propagation of the in-game disease—but for the social interactions is generate (helpful vs. griefing emergent behaviors). Image Source: Corrupted Blood and coronavirus: what World of Warcraft can tell us about the spread of disease

Creativity and Emergence

Whereas virtual worlds like World of Warcraft emphasized adventure stories—and emergence was a consequence of the humans interacting with each other around these adventures—other worlds emphasized creative expression. Much of this was in user interface innovation and tools that enhanced composability.

Minecraft gave players the ability to shape the structure of the world, build custom servers and invent mods that affected the experience of other players. One of the most emergent features of Minecraft is Redstone: a material that may be used to create machines, transfer power and implement circuits.

Like the Game of Life, Minecraft is Turing-Complete. Using Redstone, you can create logic gates, and using those it is possible to construct a complete computer including those that can play computer games like Tetris

And since you could build a computer in Minecraft, it is also possible to implement a neural network inside Minecraft. Here’s a Minecraft map that implements a convolution neural network that can perform digit recognition tasks:

Emergence in Human Behavior

All the emergent complexity we observe in life on Earth is due to a small family of replicating molecules. The number of genes in the human genome is in the tens of thousands. This extends to the sphere of human behavior: about a third of these genes are believed to be related to our brains.

Overall, the difference between humans and our closest relatives—chimpanzees—is around 1.2-3% depending on how you calculate. Small building-blocks can result in incredible improvements.

David Holz (the founder of Midjourney) opined that we’ve already had a Singularity:

The number of uniquely-human genes that result in our neurobiological phenotype isn’t huge. But they enable us to:

• Form complex social organizations (cities, corporations, institutions, online gaming leagues, DAOs)

• Communicate with language (speech, writing, theater, games, websites)

• Create incredible technologies (airplanes, microprocessors, space probes)

• Create cultural artifacts like Shakespeares plays, Hollywood movies, videogames and AI-generated Balenciaga videos.

“Real” artificial intelligence will likely not require massive amounts of code to implement—although the science and understanding of these systems is obviously very complicated.

Emergence and Artificial Intelligence

Much of the recent excitement in artificial language is that the natural-language interfaces “just work.” And while these systems hallucinate and makes mistakes (itself a quality we attribute to humans more than machines) it is a universal interface that allows us to interact with them efficiently.

Much of this qualitative improvement in our experience of using products like ChatGPT is due to quantitative improvements that resulted from scaling-up the number of parameters present in language models.

Jason Wei tracks 137 emergent capabilities of large language models

We don’t completely understand why this is yet. The human brain is the most complex naturally-occurring object that we’re aware of—and the vast number of mysteries within the brain remain elusive.

Similarly, these AI models (GPT-4 in particular) are the most complex software that we’ve ever created—and unlike many types of software, they are trained into existence rather than built line-by-line. Like our brains, we still lack the tools for probing and explaining everything that’s going on.

Generative Agents

It turns out that language is also an effective method of interfacing artificial intelligence subsystems with each other. When these systems are imbued with specific goals and objectives, they become “agents.”

A recent paper from Stanford (Park et al) demonstrated the ability to populate a virtual world with characters who interact with each other. As noted in the paper:

A society full of generative agents is marked by emergent social dynamics where new relationships are formed, information diffuses, and coordination arises across agents.

From: Generative Agents: Interactive Simulacra of Human Behavior

Similarly, agents such as AutoGPT and babyagi use language to establish goals, record their memories, conduct google searches, and iteratively form and improve their plans. Using AutoGPT, I created an agent that seeks out ways to help humans flourish:

The Search for Creativity

Human creativity was very limited before humans had good means of communicating with lots of other humans. Now we connect billions of minds—and the acceleration of technology is the result: AI in particular, but it applies to everything—energy, space exploration, metaverse, AR/VR, etc.

It wasn’t that individual minds were limited (we haven’t evolved much since the Paleolithic)—it was just that throughout the aeons of human evolution, few of us could stand on the shoulders of others.

The “bicycle for the mind”—the personal computer—made us far more efficient at creativity and networked communication.

And now we are moving towards a new phase in human civilization: one that involves not only enhancing our own creativity with computers, but working alongside a network of generative models and agents that will help along the path of discovery. These systems will not only be collaborators, they’ll help us filter through the vast ocean of data and information and and applications and practices of all the creativity that happened of the past.

We will experience a huge boost in emergent creativity due to:

• A boost in creative efficiency: user interfaces, language

• Composability: the ease of integrating, linking and combining creative content

• The exponential scale-up in the number of creative actors present in the civilizational noosphere.

What is creativity?

Humans are currently better at creativity than anything else we’ve encountered.

But rather than think of creativity as something unique to our genes or our brains, or divinely inspired, or based on some other vital magic—it may be helpful to think of creativity as a search.

If the universe is a nearly-infinite number of possibilities, parameters and variables—then perhaps creativity is about applying efficient processes towards this search for effective solutions.

AI meets efficiency: from Tenachi paper on how constraining the search space for symbolic regression models leads to exponentially-improved results.

This search is one that results in all manner of discoveries: not only scientific discoveries, but engineering problem-solving and the production of artistic works and cultural products.

We don’t want to explore the entire variable-space, because that would require infinite computation and therefore infinite time and energy. Accessing this knowledge involves making the keys to unlock useful information and new ways of seeing the world. We want results that are lower-entropy than the random noise that must occupy most of infinity—while looking beyond the zero-entropy information that simply rehashes what we already know.

Good means of conducting this search is what we might call intelligence.

As we continue to scale-up the number of minds, network with each other, and create better algorithms for conducting the search, we will produce useful outputs: the kind we call creative.

In one inconceivably complex cosmos, whenever a creature was faced with several possible courses of action, it took them all, thereby creating many distinct temporal dimensions and distinct histories of the cosmos. Since in every evolutionary sequence of the cosmos there were very many creatures, and each was constantly faced with many possible courses, and the combinations of all their courses were innumerable, an infinity of distinct universes exfoliated from every moment of every temporal sequence in this cosmos.

—Star Maker, Olaf Stapledon

Graphic Novel of Star Maker (with AI-generated art

Further Reading

• You might enjoy my article, the Direct from Imagination Era Has Begun.

• You may also enjoy the series of conversations I’m holding with pioneers in generative AI on Metavert.tv.

• I also wrote about composability and “standing on the shoulders of giants” in Composability is the Most Powerful Creative Force in the Universe.

1. I want to credit a few people with this prompting me to this idea. First, Olaf Stapledon in his terrific science-fiction-philosophy novel Star Maker, where he wrote about the infinite possibilities within hypercosmic reality. David Deutsch, in The Beginning of Infinity, wrote about searching through an infinite space of possibilities, retaining solutions to problems while discarding things that don’t. And Erik Bethke, CEO of Million on Mars, who put it succinctly as all discoveries as a “search through the infinite variables in the universe” at the meeting of the Game Industry Cocktail Hour at GDC 2023.