Why why?

From navigating us through busy streets to managing the intricacies of our smart homes, our new digital companions promise to make our lives easier, safer, and more efficient. However, for AI to truly excel in these roles, a deep understanding of cause-and-effect is fundamental. Yet current AI technologies, like Large Language Models (LLMs) and Reinforcement Learning (RL), fall short in this aspect.

The Essence of Cause-and-Effect. At the heart of effective decision-making lies the ability to understand the causal relationships between events. One everyday example that underscores the importance of looking beyond statistical correlations is something as simple as observing someone walking past your window holding an umbrella. At first glance, you might predict it's raining outside—a reasonable assumption based on a common correlation. However, this assumption doesn't account for alternative explanations, such as the individual using the umbrella as a sunshade on a bright day. Without understanding the 'why' behind what we observe, we risk making poor predictions and decisions. Causal reasoning pushes us to consider not just the correlation but also the context and possible alternative explanations for the observed phenomena, leading to more accurate interpretations and better decision-making. The question ‘What If?’ can help us to improve our predictions and judgments. ‘What if it’s sunny outside, would someone still use an umbrella?’.

This principle is not just applicable in everyday life but is critically important in the development of AI co-pilots that aim to understand and navigate the complexities of the real world alongside humans. This goes beyond recognising patterns or correlations; it's about grasping why things happen. For AI agents tasked with assisting humans, this understanding is crucial. It enables them to predict outcomes, make informed decisions, and take actions that align with our best interests. But how well do our current AI technologies fare in this domain?

‍

Limitations of Current AI Technologies

Large Language Models (LLMs) such as OpenAI’s GPT-4, Meta’s Llama 2, Google’s Gemini, have astounded the world with their ability to generate human-like text. They can write poems, draft emails, and even create code. However, their operation is grounded in identifying statistical correlations within vast datasets. They lack an intrinsic understanding of the causal relationships between the events they describe. While LLMs can mimic human language and generate plausible responses, they do not encode the ‘why’ behind the information, making their assistance superficial in scenarios requiring deep causal understanding.

Reinforcement Learning (RL) has shown promise in environments where the goal is clear and actions lead to direct outcomes, such as in games or controlled tasks. By optimising actions to maximise rewards, RL agents learn through trial and error. However, this learning process is fundamentally different from understanding causality. RL agents are adept at navigating environments with well-defined rules but struggle in the real world, where causality is nuanced, and the relationship between actions and outcomes is not always direct.

‍

The Challenge of 'Why'

For AI agents to transition from mere assistants to true co-pilots, they must evolve to comprehend the complex web of cause-and-effect that characterises the real world and the work that we do. This means moving beyond pattern recognition and reward maximisation to predict the long-term consequences of actions and understand the causal mechanisms at play. The question then becomes, how can we equip AI with this level of understanding?

Encoding Expert Wisdom into Causal Models. Our approach to this challenge involves encoding the real-world wisdom of experts into a causal model of the task domain and the expert's decision-making process within that domain. By doing so, we can leverage the nuanced understanding that comes from years of experience and domain-specific knowledge. Experts can help identify which causal variables are most relevant, which intervention targets are desirable, and how to simplify complex causal structures without losing essential information for decision-making.

Let’s unpack this. In the context of making changes or improvements, a variable is anything that can change or be changed. It's a factor that can influence outcomes in different situations. For example, in a school setting, the way teachers teach reading could be a variable because it can change from one method to another. An intervention goal is the specific outcome we're aiming for when we make a change. It's what we hope to achieve by adjusting variables. For instance, if we want students to read better, improving their reading levels is our intervention goal.

Let's illustrate these concepts with a simple example. Imagine a school wants to help its students read better (this is the intervention goal). To achieve this, the school decides to try a new teaching method for reading. The change in teaching method is the variable we're adjusting to reach our goal. If, after trying this new method, students' reading improves, we can see that changing our teaching method (the variable) helped us achieve our goal of improving reading levels.

This approach of identifying what we can change (variables) and what we hope to achieve (intervention goals) helps us make informed decisions and track whether our efforts are leading to the desired outcomes. This is how we train our AI co-pilots to behave.

This integration of expert insight is crucial for developing AI systems that can make sense of the vast amount of data they encounter and apply it in a way that aligns with human reasoning and values. It keeps humans in the loop, ensuring that the AI co-pilot’s causal reasoning is grounded in practical reality and actionable insights.

‍

Towards a Solution

Addressing the 'Why why?' challenge requires a paradigm shift in AI research. We need to explore new models that prioritise causal inference and understanding. This might involve integrating the strengths of LLMs and RL with advanced causal models and leveraging insights from psychology, cognitive science, and other fields. Ongoing research in causal AI and machine learning is promising, suggesting a future where AI agents can navigate the complexities of cause-and-effect in the real world.

The quest to equip AI agents with a deep understanding of causality is more than an academic pursuit—it's a necessity. Only by solving the 'Why why?' and incorporating the wisdom of seasoned experts can we unlock the full potential of AI as our co-pilots, capable of navigating the complexities of the human world with the same nuance and understanding that we do. The journey is undoubtedly challenging, but the rewards - a world where AI truly understands and anticipates our needs - are immeasurably valuable.