These agents have the ability to draw conclusions about themselves, other agents, and their city by storing new information in memory. At the same time, the bots demonstrated “plausible” human behavior, for example, they coordinated plans, had meaningful conversations, and even “unwound” in the evenings at a bar.
The secret sauce behind these digital beings lies in the fusion of large language models with computational, interactive agents that possess the memory, reflection, and planning prowess of a chess grandmaster, simulating human behavior that seems real.
– Memory and retrieval allows agents to store a complete record of their experiences using natural language in a memory stream containing observations of their actions and others’ actions that they’re able to retrieve for current situations based on recency, importance, and relevance, selecting the top-k memories to pass to the language model.
– Reflection and summarization allows them to synthesize their memories over time into higher-level reflections that capture their opinions, feelings, and insights and use reflection triggers to decide when to generate a reflection, such as at the end of the day or after an important event. The agent then uses a summarization technique to compress the selected memories into a concise reflection that is added to the memory stream.
– Planning and execution allows them to use their memories and reflections to plan their future actions and achieve their goals using a planning trigger to decide when to generate a plan, such as at the start of the day or when facing a problem. They use a generation technique to produce a sequence of actions that form a plan that is added to the memory stream and executes that plan by selecting and performing one action at a time, while updating the memory stream with new observations.
The most important aspect here is the memory stream, because without it, nothing would be feasible and it would just be The Sims, with mindless bots roaming around with no intention or thought. That’s what allows these agents to not only engage in complex behavior, but having their actions rated more human than humans roleplaying.
Paper: https://arxiv.org/pdf/2304.03442.pdf
Demo: https://reverie.herokuapp.com/arXiv_Demo/#
#chatgpt #ai #technology
————————
🐱🚀 GitHub: https://github.com/forrestknight
🐦 Twitter: https://www.twitter.com/forrestpknight
💼 LinkedIn: https://www.linkedin.com/in/forrestpknight
📸 Instagram: https://www.instagram.com/forrestpknight
📓 Learning Resources:
My Favorite Machine Learning Course: https://imp.i384100.net/YgYEBJ
Open Source Computer Science Degree: https://bit.ly/open-source-forrest
Python Open Source Computer Science Degree: https://bit.ly/python-open-source
Udacity to Learn Any Coding Skill: http://bit.ly/udacity-forrest
👨💻 My Coding Gear:
My NAS Server: https://amzn.to/3brqO7b
My Hard Drives: https://amzn.to/3aKetMi
My Main Monitor: https://amzn.to/3siQfPa
My Second Monitor: https://amzn.to/3keHT84
My Standing Desk: https://amzn.to/3boAcbC
My PC Build: https://bit.ly/my-coding-gear
My AI GPU: https://amzn.to/3uvmUmz
