Posted by Xiao Xiao from Aufeisi Qubit
| Public Account QbitAI
Letting 25 AIs live together, they simulated human actions!
Not only do they get up, brush their teeth, eat, and sleep like people, but they are also forced to "go out to work" and even celebrate festivals together in their spare time.
All this, from a new study from Stanford and Google -
They did a very new experiment, created 25 AI NPCs, each with different identities and action decisions, and let them live together in a sandbox environment, relying on the ChatGPT large model to complete action decisions .
It turns out that these AI NPCs will not only live like people, such as writers to write, shopkeepers to run stores, but they will also interact with each other and even have dialogues:
As soon as the research was released, it exploded. Ethan Mollick, a professor at the Wharton School of Business at Penn University, said:
These actions are more like actions that normal humans would make than "role-playing".
OpenAI scientist Andrej Karpathy even said that "this dwarfs open world games":
So, what kind of "AI society" does this research construct?
When AI Plays The Sims
Judging from the demo display, these 25 AIs have different identities, personalities and ages:
They will live in a communal community, which is very detailed, including residential areas, universities, parks, cafes, etc.:
It can be seen that each AI has its own basic itinerary, including waking up, washing, eating, sleeping, etc. According to different identity settings, it will take different actions to jointly maintain the operation of the community.
Among them, the emoji in the demo simply represent the ongoing operations of AI, such as sending and receiving letters and eating:
But among them, there will also be interactive behaviors:
It is not the language of AI, but the way humans speak:
Not only that, these AIs that can interact with each other can even listen to the requirements of human players and take corresponding actions.
For example, the authors of the paper mentioned that they told one of the AI NPCs as a human being that they hoped to hold a "Valentine's Day Party" event in the community.
Unexpectedly, this AI NPC not only enthusiastically agreed to the requests of the authors, but spontaneously began to organize this event, including the time, place and people present:
Seeing this, some netizens have already expressed "too crazy"——
For players, rather than follow the steps, they may really need to persuade NPCs or think and act in order to complete the task in the game.
There are also netizens who have opened their minds and thought of a more "evil" gameplay:
Imagine you killed an NPC, then drove back to the city and watched them hold his funeral...
So, how do these AIs act spontaneously and build a self-running AI interactive society?
Control Action Decisions with Large Models
Specifically, Stanford and Google have AIs take action and interact with other AIs in two steps .
The first step is to construct a variety of AI "personal settings" and let them take daily actions that conform to the settings.
Here we take John Lin, one of the AI characters, as an example. You can find his setting list in the Demo.
It can be seen that his settings are very detailed, including basic information (name, age, location, time), attention settings (field of vision, etc.), personal circumstances (personality, family background, work, etc.).
Using this as a reference, the AI named John Lin can form a fixed daily living habit, including getting up early to wash up routinely, then going to work, and finally going home to sleep.
But in the process of his life, he will definitely communicate with other AIs or meet new people, so——
The second step is to let the AI character respond to the external environment and the actions of other people to generate interactions .
Thus, this research builds an architecture that empowers AI with memory, retrieval memory, perception, feedback, and interaction.
Judging from the architecture diagram, AI will first perceive the surrounding environment, record events that occur within a certain field of vision (attention), and add them to its own memory stream (memory stream);
Subsequently, whether it is what the AI plans to do (plan) or give feedback (reflect) on what happened outside, it will be affected by this memory flow. The specific way is to retrieve the corresponding events from the memory and rely on large Model ChatGPT to make decisions ;
(Of course, whether it is planning or feedback, the corresponding event will also be added to the memory flow and become a new memory)
Finally, the results of these decisions will be translated into concrete solutions, prompting AI to think and take concrete actions.
The study mentioned that the reason why the large model + memory architecture is adopted is that although the large model already has very good action logic, whether it is GPT-4 or other large models, long-term planning and maintaining a clear memory are still a problem. kind of challenge.
Based on this, a memory system is needed as a "reference" for the large model, so that it has a clear list of events that can be traced back when making decisions.
Still, the study raises some questions about AI safety .
Including how to prevent its impact on society, and at the same time being used by people with intentions to generate Deepfake and "customized version" AI fraud, etc.——
For example, the AI tea seller who chats with you in the future may not only remember everything you said, every birthday, but can even use this as a reference to generate more "intimate" answers, so that "her" looks more like a person...
One More Thing
It is worth mentioning that a domestic company called Hyperparameter Technology has already made a similar AI NPC system GAEA .
Like the Stanford Google study, GAEA also has a memory system and relies on a large language model to help NPCs make action decisions.
However, there are some differences in the incentive mechanisms that drive AI to act. Compared to this study relying on large models to complete action decisions:
The AI in GAEA will also rely on more specific values to further affect the NPC's action decisions (hunger value, mood value, etc.):
For this kind of AI NPC technology, what brain-opening gameplay do you think of?
Demo address:
https://reverie.herokuapp.com/arXiv_Demo/#
Paper address:
https://arxiv.org/abs/2304.03442
Reference link:
[1]https://twitter.com/emollick/status/1645499660402925576
[2]https://twitter.com/karpathy/status/1645485475996790784