Shulou(Shulou.com)11/24 Report--

Q version of Westworld! Generative agents attack, we are independent personality NPC!

Have you seen all the NPC in the game?

No matter what NPC does, take on tasks with tasks and chat without tasks, what they have in common is that they just say those words over and over again.

The reason is also simple: these NPC are not smart enough.

In other words, the traditional NPC is to first arrange the script for them, arrange the speech skills, and say the words as soon as they go.

With the advent of ChatGPT, the conversations of these game characters can be self-generated on the premise of entering only key information.

That's what Stanford and Google researchers are doing-generative agents created with artificial intelligence.

How to generate generative agents? The mechanism of this thing is actually very simple, which can be easily summarized with a picture.

The leftmost Perceive is like the initial input, and the agent begins to receive all kinds of information.

The big frame in the middle is the most important part, which reveals how generative agents "generate" content and spontaneously go to the final step "Act".

The most important feature of generative agent is that it has a Memory Stream, that is, memory stream.

Like the traditional npc mentioned above, their actions and words are prescribed, and there is no room for free play.

If you want to play freely, you must have your own "library". This "library" contains what AI has written down and thinks about what you see and hear every day. Only in this way can you further "Act".

AI has to learn.

The link of extracting memory, planning and reflection is the link of AI dealing with memory. AI can not only remember what happens every day, but also reflect on these events and make plans for each day.

At the same time, they will remember the new feedback caused by AI's behavior and use it as new material to continue a new round of processing.

Observation is each character's observation of the surrounding environment, people and things, while Reflection is the reflection of these observed phenomena.

Then the characters generate Plan based on this content.

At first, they worked out a general plan by themselves, and then slowly refined it. First make an hourly plan, then make a short-term plan of 5-15 minutes.

In addition, the characters will modify their plans according to the changes in the surrounding environment. Completely imitating human behavior, who doesn't frantically revise their plans at the end of the day?

In this way, we see the charm of generative agents.

We can see that the performance of the whole generative agent (including observation, thinking, and planning) is much better than that without any one or two kinds of architecture.

Cute little Emoji there are a total of 25 little AI people living on such a map in the scene designed by Google and Stanford researchers.

The researchers set up a total of 25 roles and gave each role basic information such as name and occupation.

For example, John Lin is a drugstore clerk and is very helpful. He is always thinking about how to make it easier for his clients to get medicine.

In terms of life, John lives with his wife Mei Lin, who is a university professor. Their son Eddy Lin is studying music theory.

At the same time, it also establishes the initial network of relationships between the characters, such as John's relationship with his neighbor, his opinion of the neighbor, his relationship with his colleagues in the drugstore, what he likes to talk about, and so on.

Each character has such a preset text, which initially outlines the role framework and leads to subsequent interaction.

The rest is up to AI to play.

The behavior of each role is represented by emoji and is summarized in one sentence in the system. For example, if so-and-so is keeping a diary, he will use

To show.

One day, Isabella and Tom met. They talked about the campaign.

You know, characters can communicate with each other in a complete and clear natural language. And each character can perceive other characters nearby, and can decide whether to pass by or start a conversation according to their own personality and character relationship.

Isabella: "I was wondering who I should choose." But I talked to Sam Moore about the campaign. What do you think of him? "

Tom: "to be honest, I don't like Sam very much." I think he is a little divorced from our community and will not consider to maximize our interests. "

For the purpose of the experiment, the researchers can also intervene. But this kind of intervention is different from the traditional input script. Instead, it exerts a little bit of force on the direction of some branches, and the rest depends on AI himself.

Specific to the daily life of each day, each character will have its own "routine", this "routine" is preset.

Here is the day of John Lin, the little brother of the pharmacy. And as for what happens every day, that's the most interesting part.

Finally, let's take a look at the interesting interactions between AI on Valentine's Day.

On Valentine's Day, the researchers gave two characters, Isabella and Maria, a preset message each, telling Isabella that she was throwing a party and that Maria liked Klaus.

Then the two characters began to play on their own.

Isabella strolls on the map and invites people to his party when he sees them.

This includes Maria. Maria certainly wouldn't miss the opportunity to spend time with Klaus, so she went to Klaus and invited him to Isabella's party.

Isabella invited everyone to go to the coffee shop to decorate the venue and invited Maria to help.

The picture above shows the relationship between these AI and the interaction caused by the party.

Then a group of people AI happily started the party at five o'clock in the afternoon.

The following picture shows in detail how the party was initiated by AI. It involves memory retrieval, reflection from the observation of the environment, and decision-making.

On the left is Isabella's observation of the environment, such as idle refrigerators, beds, shelves, desks, wardrobes, etc., and then makes decisions about decorating, inviting, designing parties, etc., depending on the environment and the intention of hosting the party.

In this way, the whole process appears to be very full.

Science fiction into reality? The above scenes seem simple and fun, but the more they develop, the more they can show different forces.

Pete said on Twitter that this achievement can be applied in many ways in the future.

For example, in the VR world, with this generative agent acting as npc, the scenario envisioned in the Western World will no longer be out of reach.

"in the Western World, every NPC has its own ideas, personalities, plans, and relationships. Even more things are generated on their own. "

And this thing, made by researchers at Google and Stanford, is a mini Q version of Westworld.

Reference:

Https://twitter.com/nonmayorpete/status/1645355225157615616

Https://www.reddit.com/r/MachineLearning/comments/12hluz1/r_generative_agents_interactive_simulacra_of/

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.