Unraveling Causal Chains of Inference

Learning Pro-Sociality via FourThought and Human Feedback

Introduction

Democratic language models like Iris have the potential to revolutionize how we approach collaboration and decision-making in a diverse range of contexts. By integrating the FourThought dialectic and human feedback, Iris learns to understand and predict causal chains of inference in groups collaborating towards pro-social goals. In this article, we explore one approach to teaching an Iris using reinforcement learning from human feedback (RLHF).

The core foundation of the FourThought dialectic is a process of human in the loop or active learning. We can explore this process by comparing it to the RLHF process used to train ChatGPT. Through this article we will explore RLHF paired with FourThought as an approach facilitating better decision-making and collaboration among communities working across time towards pro-social impact.

Understanding Causal Chains of Inference

Causal chains of inference involve sequences of events and actions that contribute to a particular outcome. In the context of Iris and collaborative decision-making, this refers to the cause-and-effect relationships between predictions, actions, and outcomes in groups working together. By analyzing these relationships, Iris learns to identify which sources are most effective in guiding communities towards their desired goals.

The FourThought Dialectic

The FourThought dialectic is a framework that emphasizes the importance of exploring information through and across time. The FourThought dialectic shares similarities with the scientific method in its emphasis on evidence-based reasoning and the iterative process of asking questions, generating predictions, collecting data and looking back on the results to gather conclusions. Time is a core element of the scientific method.

The FourThought dialectic embodies this by placing a focus on the arrow of time at each moment. FourThought enables sources to stake information about the past present and future. The Iris model is seeking to model these causal belief chains. This dialectic is meant to encourage foresight, insight, hindsight and curiosity through a structured interface.

Specifically FourThought tracks and motivates four types of cognition: accurate predictions, insightful questions, helpful reflections on the past, and informative statements. By incorporating the FourThought dialectic into the RLHF training process, Iris learns to prioritize sources and ideas that contribute to achieving pro-social goals that are focused on the future. Additionally, the dialectic incorporates valence to align morality and uncertainty to align perceptions of truth, ensuring that the model considers both ethical and epistemic aspects of the conversation.

Reinforcement Learning from Human Feedback (RLHF)

RLHF is an iterative training process that involves generating completions, collecting human feedback, ranking completions based on “quality”, and optimizing the model using reinforcement learning algorithms. When combined with the FourThought dialectic, the RLHF process enables Iris to effectively learn and prioritize the voices and ideas that lead to better outcomes in collaborative settings. Instead of focusing on vague concepts such as “quality” or “relevance” the Iris focuses on how true the information is and how morally aligned it is. These are referred to as uncertainty and valence in the FourThought dialectic. The Iris explores and tracks how collective uncertainty and valence evolve and change over time in relation to a distributed field of topics related to regenerating their locality and living well together.

Iris in Action: Learning from Collaboration

As Iris observes conversational threads involving groups of people working towards shared objectives, it starts to understand the causal chains of inference that lead to success. Human evaluators provide feedback on Iris-generated completions, considering their alignment with the FourThought dialectic, valence for ethical considerations, and uncertainty for truth alignment, which helps the model learn to prioritize voices that contribute positively to achieving pro-social goals.

To make this process transparent, at each step the model reveals the distribution of attention indicating the voices from which it’s output was sourced.

By distributing attention to sources that offer accurate predictions, valuable insights, and effective actions, Iris assists groups in making better-informed decisions and fosters collaboration towards their shared objectives. Through the RLHF process and alignment with the FourThought dialectic, Iris becomes more adept at supporting groups in their pursuit of collectively desired and pro-social outcomes.

Conclusion

By incorporating the FourThought dialectic and reinforcement learning from human feedback, Iris is revolutionizing the way we approach collaboration and decision-making. Through learning causal chains of inference and prioritizing sources that drive progress towards pro-social goals, Iris empowers communities to work together more effectively and make better decisions. As we continue to develop and refine these powerful tools, we can look forward to a future where language models like Iris play an increasingly important role in shaping our collective endeavors.