But looking beyond, I see that historically, games also served as powerful testbeds for technical breakthroughs. Not surprisingly, cognitive architectures built by the founders of my field (e.g., Newell and Simon) and influenced our work, found their home in game NPCs. (6/14)
1
1
8
Replies
How might we craft an artificial society that reflects human behavior? My paper, which introduced “generative agents,” will be presented at
#UIST2023
and now has an open-source repo! w/
@joseph_c_obrien
@carriejcai
@merrierm
@percyliang
@msbernst
🧵
10
114
525
In threads of this genre, I usually summarize the technical contribution of our work, but I've noticed that others have already done that well. So instead, I will offer three updates/reflections on generative agents. (2/14)
1
0
4
1. Please find our open-source repo here:
We hope our community can build on this to improve generative agents. There are also many open-source projects (e.g., from
@LangChainAI
) that implement the core ideas from our work, and… (3/14)
1
2
22
… those projects will continue to be useful. Our repo will add to those efforts by showing how we conceptualized and implemented the original generative agents in a fully developed simulation game setting. (4/14)
1
1
3
2. What are exciting application areas for generative agents? Obviously, I am excited about the future of games -- generative agents may unlock new forms of gameplay that were previously impossible. Here is a thoughtful article by
@Tocelot
: (5/14)
1
1
10
So I am excited to see how generative agents, originally conceived as game NPCs, might contribute to defining a new class of interactive applications (e.g., immersive environments, rehearsal spaces for communications, and simulations for testing social theories). (7/14)
1
0
10
3. What will future systems look like with LLMs? We are at an interesting point in computing history. LLMs and other foundation models have become powerful enough to give us a chance to realize our long-standing goals, such as fully general human-like agents. (8/14)
1
0
5
But to achieve those, I posit that we will need to go beyond first-order templates (e.g., simple few-shot prompts or chain-of-thought prompts) to design more sophisticated design patterns and architectures that leverage the power of LLMs in a larger system context. (9/14)
1
0
3
To me, the challenge is reminiscent of those faced before in computing. Though computers are powerful, it was developments such as object-oriented programming, the MVC design pattern, and operating systems that enabled truly complex and useful interactive applications. (10/14)
1
0
4
I hope the architecture of generative agents presents a version of this vision by demonstrating a working system that leverages the power of LLMs to create a complex set of individual and societal human behaviors. (11/14)
1
0
3
Finally, if you want to learn more about our work beyond what is available here, some of my talks on generative agents are publicly available (e.g., at UC Berkeley in early April: , and most recently at London ML: ). (12/14)
1
1
9
A few acknowledgments:
To me, generative agents was a project that was only possible at the intersection of two fields: human-computer interaction and natural language processing. I am grateful to
@StanfordHCI
and
@StanfordNLP
for nurturing and fostering this idea. (13/14)
1
0
7
Thanks also to
@GoogleAI
,
@MSFTResearch
, HPDTRP,
@StanfordHAI
, and
@OpenAI
for their funding support.
And finally, one more thank you to my amazing advisors,
@msbernst
, and
@percyliang
, mentors,
@merrierm
, and
@carriejcai
, and mentee
@joseph_c_obrien
. (14/14)
4
0
6