Next part (2)

---

A week ago I came across this damned paper (Generative Agents: Interactive Simulacra of Human Behavior). Also check out this explanation of the paper and its implementation. I called the paper damned because the moment I understood its possibilities (creating intelligent virtual agents has always been a dream of mine), the spark of obsession flared up in me, and for someone this obsessive, that could only mean one thing:

I figured that I was capable enough of implementing that paper using the Python programming language. In my first try, I thrust ahead, with very few unit tests, to implement the memories database, the planning and reflective functionalities, as well as the navigation system. It was when I hit the paper’s promise that GPT-3.5-turbo was going to be able to output the name of the chosen destination when I realized that the writers of the paper must have simplified the process of navigation (if not finagled it quite a bit), because large language models don’t work like that. And without the agents being able to move around in the environment tree, this system wouldn’t work remotely as intended.

I gave it some thought and I ended up opting for utility functions. Used in basic AI for decades, utility functions output a rating value between a bunch of options, from which the code ends up choosing the best rated option (with some room for randomness). Thankfully, if you ask GPT-3.5-turbo to give a rating for how well a “farmhouse” would fit the purpose of “eating lunch”, then that location may be chosen, and from there you could ask the AI to rate the individual locations of the farmhouse to continue the path.

Implementing the navigation of the agents in such a way gave me the opportunity to start the code from scratch. This time I relied mostly on test-driven development (meaning that one or more unit tests support the part of the code you are working on), very important for a system that can suffer from plenty of regressions. I have 51 unit tests supporting the code so far.

Before I delve into further details, here’s the link to my GitHub repository with all the code I have written.

Here’s as far as I’ve gotten:

• The user writes the environment of the simulation in a json file. I’m using the Python library ‘anytree’ to represent the graph, which also comes with this fancy command-line method of displaying any given graph:

Location: plot of land (plot_of_land)
├── Location: barn (barn)
│ └── Sandbox object: tools (tools) | description: lots of tools for farming
├── Location: field (field)
│ ├── Sandbox object: crop of corn (corn_crop) | description: a crop of growing corn
│ └── Sandbox object: crop of potatoes (potatoes_crop) | description: a crop of growing potatoes
└── Location: farmhouse (farmhouse)
├── Location: bedroom (bedroom)
│ ├── Sandbox object: bed (bed) | description: a piece of furniture where people sleep
│ └── Sandbox object: desk (desk) | description: a piece of furniture where a person can read or write comfortably
├── Location: kitchen (kitchen)
│ └── Sandbox object: fridge (fridge) | description: an appliance that keeps food cold
└── Location: bathroom (bathroom)
├── Sandbox object: toilet (toilet) | description: a simple toilet
└── Sandbox object: bathtub (bathtub) | description: a tub in which to take baths

• For this test, I pictured a farmhouse, a nearby barn, and a couple of nearby crops.
• I added in the corresponding json file the needed information for the agents involved:

{
“Joel”: {
“age”: 38,
“planned_action”: null,
“action_status”: null,
“current_location_node”: “farmhouse”,
“destination_node”: null,
“using_object”: null
},
“Betty”: {
“age”: 24,
“planned_action”: null,
“action_status”: null,
“current_location_node”: “farmhouse”,
“destination_node”: null,
“using_object”: null
}
}

• As depicted, agents can have a planned action, an action status, their current location, their destination, and what object they are using, apart from their age and name.
• As the paper says, each agent’s memories are seeded from a few memories that are written in a simple text file.

Betty started living with Joel six years ago;Betty enjoys planting crops and tending to them;Betty wishes that she and her boyfriend Joel could afford to buy a few horses;Betty wishes she didn’t have to travel to the nearby town often;Betty loves peace and quiet;When Betty feels overwhelmed, she lies down on her bed and listens to ASMR through her noise-cancelling headphones;Betty wants to have a child, hopefully a daughter, but Joel told her years ago that he didn’t want children;Joel and Betty saw strange lights in the skies in the 11th of May of 2023;On the 11th of May of 2023, the local news broadcast was cut off in the middle of a transmission as they were reporting on the strange lights in the sky;Joel and Betty have heard ominous explosions in the distance throughout the morning of the 12th of May of 2023

Joel is a farmer who has lived in his plot of land for ten years;Joel is living with his girlfriend, Betty, whom he loves;Joel enjoys listening to music from the sixties;Joel wishes he and Betty could buy a few horses, but he doesn’t think they can afford them;Joel and Betty saw strange lights in the skies in the 11th of May of 2023;On the 11th of May of 2023, the local news broadcast was cut off in the middle of a transmission as they were reporting on the strange lights in the sky;Joel and Betty have heard ominous explosions in the distance throughout the morning of the 12th of May of 2023

• These seed memories are inserted into a memories database. Its implementation is one of the most important parts of this system. The database is composed of a vector database (which stores values according to some eldritch, multidimensional distance between them), and a json file that records readable values of those memories. You can query the vector database with simple text, and it will return the most relevant text it contains regarding the query you sent it. Sort of like a recommendation system.
• In addition, when a query is made to the agents’ memories, they are rated according to the memory’s importance and recency. The recency is a normalized value in the range of [0.0, 1.0] depending on when that memory was accessed last, and importance is a rating that the AI itself gave to how poignant the memory was.
• Once the agents are loaded in the simulation, a character summary is created for each agent. The AI is prompted to write all the segments of these summaries only from the memories present in their vector databases.

Joel’s AI-written character summary

Joel (age: 38)

Innate traits: Hardworking, resilient, curious, observant, caring, nostalgic, pragmatic.

Joel is a farmer who has lived in his plot of land for ten years and wishes he and Betty could buy a few horses, but he doesn’t think they can afford them. He enjoys listening to music from the sixties and is living with his girlfriend, Betty, whom he loves. They have experienced ominous explosions in the distance and saw strange lights in the skies on the 11th of May of 2023, and the local news broadcast was cut off in the middle of a transmission as they were reporting on the strange lights.

Joel is a farmer who has lived in his plot of land for ten years.

It is difficult to accurately describe Joel’s overall feeling about his recent progress in life based on these statements alone. However, we can infer that he enjoys listening to old music, has heard ominous explosions and seen strange lights with Betty, is a farmer, loves his girlfriend, and wishes they could afford to buy horses. These statements do not provide enough information to determine his overall feelings about his recent progress in life.”

Betty’s AI-written character summary

Betty (age: 24)

Innate traits: resilient, domestic, nurturing, longing, introverted, anxious, curious, intuitive, calm

Betty is a woman who enjoys living a peaceful and simple life with her partner Joel. She is interested in farming and wishes to own horses. Betty also has a desire to have a child, but this conflicts with Joel’s wishes. She dislikes having to travel to nearby towns and finds solace in listening to ASMR when she feels overwhelmed. Betty is aware of strange occurrences happening around her, such as explosions and lights in the sky, but she seems to maintain a calm and composed demeanor.

Betty is a homesteader or a farmer.

Betty is feeling somewhat unsettled and uncertain about her progress in life, as indicated by her mixed feelings about her living situation with Joel, her desire for a more stable and fulfilling life with horses and a child, and her occasional need for escapist relaxation techniques like ASMR. She enjoys the sense of purpose that comes from tending to her crops, but is also wary of the potential dangers and disruptions of the outside world, as shown by the ominous sounds and sightings of explosions and strange lights in the sky. Overall, she seems to value the simple pleasures of a quiet country life, but is also aware of the limitations and difficulties of this lifestyle.

• If the simulation detects that the agents haven’t decided upon an action, it generates one through a complicated process.

2023-05-13T09:30:00 Joel produced action: Saturday May 13 of 2023, 9 AM. Joel is going to check with his neighbors to see if they have any information about the strange lights and ominous explosions they have been experiencing. The length of time it takes for Joel to check with his neighbors is not specified, so it is impossible to provide an answer in minutes.

2023-05-13T09:30:00 Betty produced action: Saturday May 13 of 2023, 9 AM. Betty is going to talk to Joel about their conflicting desires regarding having a child and try to come to a mutual understanding or compromise. The duration of this action is not specified in the given information. It can vary based on the discussion and resolution reached by Betty and Joel.

• The AI model (GPT-3.5-turbo in this case) went over each agent’s memories and decided what action the agents would take, according to their circumstances. A problem with the paper’s original implementation already comes up: the agents decide a plan, then they search for the most fitting location, and then object, to fulfill that plan. But in this case, the AI decided that Joel should check with his neighbors to see if they have any information about weird happenings. But there are no neighbors in this environment. In addition, Betty decided that she wants to talk to Joel, but the paper doesn’t offer a system of locating other agents that may not be in the immediate surroundings. There’s a system for dialogue, but I haven’t gotten that far.
• Through the rating system for locations and objects (going from the root of the environment tree deeper and deeper), Joel came to this conclusion about what should be his destination given the action he came up with:

2023-05-13T09:30:00 Joel changed destination node to: Node(‘/Location: plot of land (plot_of_land)/Location: farmhouse (farmhouse)/Location: kitchen (kitchen)/Sandbox object: fridge (fridge) | description: an appliance that keeps food cold’)

• Apparently Joel (therefore, the AI model) thought that the neighbors could be found in the fridge. From the root of the environment tree, between the choices of “farmhouse”, “barn” and “crops”, obviously farmhouse was the most adequate option to find people. Inside that house, it opted for the kitchen between all the possible rooms. Inside there was only a fridge, so the navigation system had no choice but to choose that. If I had added a telephone in the kitchen, the AI model likely would have chosen that instead to communicate with the neighbors.
• Regarding Betty, to fulfill her action of talking to Joel, she chose the following destination:

2023-05-13T09:30:00 Betty changed destination node to: Node(‘/Location: plot of land (plot_of_land)/Location: barn (barn)/Sandbox object: tools (tools) | description: lots of tools for farming’)

• Interestingly enough, the AI model believed that her farmer husband was more likely to be in the barn than in the farmhouse. Once the barn was chosen, the tools were the only option as to which object would help the agent fulfill her action.

• The navigation system detected that the agents weren’t at their destination, so it started moving them:

2023-05-13T09:30:00 Joel needs to move to: Node(‘/Location: plot of land (plot_of_land)/Location: farmhouse (farmhouse)/Location: kitchen (kitchen)/Sandbox object: fridge (fridge) | description: an appliance that keeps food cold’)
2023-05-13T09:30:00 Joel changed the action status to: Joel is heading to use fridge (located in Location: kitchen (kitchen)), due to the following action: Saturday May 13 of 2023, 9 AM. Joel is going to check with his neighbors to see if they have any information about the strange lights and ominous explosions they have been experiencing. The length of time it takes for Joel to check with his neighbors is not specified, so it is impossible to provide an answer in minutes.

2023-05-13T09:30:00 Betty needs to move to: Node(‘/Location: plot of land (plot_of_land)/Location: barn (barn)/Sandbox object: tools (tools) | description: lots of tools for farming’)
2023-05-13T09:30:00 Betty changed the action status to: Betty is heading to use tools (located in Location: barn (barn)), due to the following action: Saturday May 13 of 2023, 9 AM. Betty is going to talk to Joel about their conflicting desires regarding having a child and try to come to a mutual understanding or compromise. The duration of this action is not specified in the given information. It can vary based on the discussion and resolution reached by Betty and Joel.

• The simulation advances in steps of a specified amount of time, in this case thirty minutes, and the navigation system will move the agents a single location each step. This means that an agent would take thirty minutes to move from a bedroom to the entrance of the same house, but whatever.

2023-05-13T10:00:00 Joel changed the current location node to: Node(‘/Location: plot of land (plot_of_land)/Location: farmhouse (farmhouse)/Location: kitchen (kitchen)’)
2023-05-13T10:00:00 Joel changed the current location node to: Node(‘/Location: plot of land (plot_of_land)/Location: farmhouse (farmhouse)/Location: kitchen (kitchen)/Sandbox object: fridge (fridge) | description: a piece of furniture that keeps food cold’)

• Joel moved from the entrance of the farmhouse to the kitchen, and once the agent is in the same location as his or her destination sandbox object, he or she starts using it. In this case, the fridge.

2023-05-13T10:00:00 Betty changed the current location node to: Node(‘/Location: plot of land (plot_of_land)’)

• Betty has moved from the farmhouse to the root of the environment tree.

• In the next step of the simulation, Joel continues using the fridge, while Betty moves further to the barn, where she starts using the tools:

2023-05-13T10:30:00 Joel continues using object: Node(‘/Sandbox object: fridge (fridge) | description: a piece of furniture that keeps food cold’)
2023-05-13T10:30:00 Betty changed the current location node to: Node(‘/Location: plot of land (plot_of_land)/Location: barn (barn)’)
2023-05-13T10:30:00 Betty changed the current location node to: Node(‘/Location: plot of land (plot_of_land)/Location: barn (barn)/Sandbox object: tools (tools) | description: lots of tools for farming’)

• From then on, the agents will continue using the tools they have chosen to perform their actions:

2023-05-13T11:00:00 Joel continues using object: Node(‘/Sandbox object: fridge (fridge) | description: a piece of furniture that keeps food cold’)
2023-05-13T11:00:00 Betty continues using object: Node(‘/Sandbox object: tools (tools) | description: lots of tools for farming’)
2023-05-13T11:30:00 Joel continues using object: Node(‘/Sandbox object: fridge (fridge) | description: a piece of furniture that keeps food cold’)
2023-05-13T11:30:00 Betty continues using object: Node(‘/Sandbox object: tools (tools) | description: lots of tools for farming’)
2023-05-13T12:00:00 Joel continues using object: Node(‘/Sandbox object: fridge (fridge) | description: a piece of furniture that keeps food cold’)
2023-05-13T12:00:00 Betty continues using object: Node(‘/Sandbox object: tools (tools) | description: lots of tools for farming’)

That’s unfortunately as far as I’ve gotten with my implementation, even though it took me a week. The paper explains more systems:

• Observation: as the agents use the objects, they are prompted with observations about what other objects in the same location are doing, and about the other agents that may have entered the location. That gives the busy agent an opportunity to stop what they’re doing and engage with other objects or agents.
• Dialogue: the paper came up with an intriguing dialogue system that became one of my main reasons for programming this, but it requires the Observation system to be implemented fully.
• Planning: once a day, the agents should come up with day-long plans that will be stored in memory, and that will influence the actions taken.
• Reflection: when some condition is triggered, the agents retrieve the hundred most recent memories from their databases, and prompt the AI model to create high-level reflections of those memories. The reflections will be stored in the memory stream as well, so that they influence their planning, their decision-taking, as well as the creation of further reflections.

That’s as far as I remember of the paper at this moment. I wanted to get the navigation system done as early as possible, because I considered it the hardest part, and it annoyed the hell out of me partly due to how Python handles references (at one point I was losing the list of ancestors and descendants of the agents’ destination nodes between simple function calls to the same instance of Simulation, for no fucking reason that I could determine). I may move directly onto the Observation system, or take a rest. I already feel my sanity slipping quite a bit, likely because I took a sabbatical from my ongoing novel to do this.