OpenAI DevDay 2024 | Community Spotlight | Altera

In the OpenAI DevDay 2024 presentation, Robert Yang from Altera discusses the company’s mission to create digital humans that embody autonomy, emotion, and social dynamics, exemplified by their Project Sid, which explores autonomous agents in a Minecraft environment. He highlights the challenges of long-term agent progression and introduces a concurrent architecture inspired by brain functions to enhance the agents’ adaptability and decision-making capabilities.

In the OpenAI DevDay 2024 presentation, Robert Yang from Altera.AL introduces the company’s mission to create digital humans that can live, love, and grow alongside humans. He emphasizes the distinction between artificial life and artificial intelligence, arguing that building life encompasses more than just intelligence. With a background as an assistant professor at MIT, Yang discusses how the advent of large language models has transformed the landscape of digital agents, leading him to co-found Altera with a vision of developing agents that possess fundamental human qualities, including autonomy, emotion, and potentially consciousness.

Yang highlights the ambitious goal of having 100 billion agents that can collaborate and progress over extended periods, rather than just short tasks. He introduces Project Sid, inspired by Sid Meier’s Civilization games, which involves creating a Minecraft server populated by autonomous agents. The aim is to observe how these agents develop their own economies, cultures, and social dynamics without human intervention. Yang shares anecdotes from the project, such as the emergence of a trading hub and the development of a religious narrative among the agents, showcasing their ability to create complex social structures.

The presentation delves into the challenges of long-term progression for agents, particularly the issue of data degradation over time. Yang explains that traditional agent workflows can lead to looping behaviors, where agents become stuck in repetitive cycles due to the quality of their outputs declining. He emphasizes the importance of preventing these loops to enable agents to continue making progress autonomously. The team at Altera has been experimenting with different models to extend the operational time of agents, achieving significant results with their approach.

Yang discusses the technical aspects of their concurrent architecture, which is inspired by brain functions. Unlike typical sequential models that process information one step at a time, Altera’s architecture allows multiple modules to run concurrently, enabling agents to process information at different timescales. This design is complemented by context-dependent modules that activate only when necessary, optimizing resource use and adaptability. The intent generation module plays a crucial role in decision-making, focusing on essential information to guide the agents’ actions coherently.

In conclusion, Yang emphasizes the ongoing research and development at Altera to create a future where multi-agent collaboration is a reality. He invites interested parties to connect with the team to explore their consumer product and the potential of their innovative approach to digital agents. The presentation underscores the transformative vision of Altera.AL in building digital humans capable of complex interactions and long-term engagement, paving the way for a new era of artificial life.