Is it easier to build a Westworld than to order pizza? At what point is it deemed appropriate to forget what you ate for breakfast? What is a viable ratio of elapsed simulation time to real world time?
The latest session of our New Ideas in AI Series brought together yet another exciting group of leading engineers, founders, and academics from across the industry to discuss these very questions. At the heart of the event was a conversation with the author of the much-discussed Generative Agents paper, Joon Sung Park. Joon’s work pushes the boundaries of thought on what AI can achieve, suggesting a future where AI is more interactive, personalized, and contextually aware. His observations, however, also pushed us to think about the practical implications of agents in the real world and the challenges that remain.
You can find Joon’s full talk above, and some of the topics we found to be the most exciting:
Soft edge vs. hard edge problems
“Building something like a Westworld or a Matrix-like simulation might be easier than ordering pizza”
Public imagination seems to be fixated on the idea of AI “ordering pizza” and tasks alike. Joon, however, prompts us to reflect on whether or not these types of applications are truly best suited for agent-based solutions today. He underscores the spectrum between hard edge problems with definitive success criteria and soft edge problems where the boundaries are much more fluid. While the idea of AI ordering pizza may seem appealing, softer edge problems, like creating NPCs, might actually be the more practical application today.
The challenge of adequate memory streaming
“You likely don’t remember what you ate for breakfast two weeks ago. But let’s say you broke up with somebody – you likely remember for some time, right?”
While seemingly obvious, Joon underscores the fact that determining what information is relevant, important, or recent enough to be retrieved in a given context requires a non-trivial level of sophistication for an agent. The balance between the size of the memory stream, computational efficiency, and cost is critical. The ideal strategy is also highly dependent on the application and task at hand.
“It’s not cheap. But the answer is not to just continue to pour more money in”
The computational resources required for running agent simulations are becoming a rising concern. In the case of Joon’s Smallville, as the number of agents increases, the computational requirements grow exponentially due to the agents’ socializing and communication dynamics. However, Joon emphasizes the fact that simply pouring more money into the simulation time is not a sustainable solution. To make AI agents commercially viable, there is a pressing need to explore ways to make them cheaper and faster. For Joon, he believes running a year of simulation in a month could be the current North Star.