|||
Over the last year Generative AI has astounded people with the ability to create human-like text, images, and even videos. But that initial wonder has faded into hope for future use cases, at the same time that conversations about LLM hallucinations and egregious energy consumption have grown in the public's awareness.
The primary problem, beyond energy use which we’ll come to later, is that Generative AI lies. All the time. Everyone who’s played with ChatGPT or Claude or Gemini has been presented with “factual information” that is not true. In fact, it is *expected* that responses will be wrong. It's right there on the label!
The Problem of Generative AI's "Hallucinations"
Bullshit is baked into Generative AI. It has no respect for the truth. It can’t explain which answers are wrong or why they are wrong because these systems are inscrutable black boxes, even to themselves. We literally don’t know why an AI system decides on a specific output for a specific query at a specific time, or what the probability is that each answer is true or not.
There are lots of factors that create these erroneous outputs, including architecture, objective functions, optimization techniques, among others. But I want to focus on one in particular.
Human knowledge is grounded in our ability to engage with our environment, learn from experience in real-time and apply that learning to achieve our goals. Without this fundamental capability, Generative AI systems will consistently produce inaccurate answers without the ability to recognize or explain its mistakes.
It’s obvious at this point that the problems with Generative AI should be a bigger deal. AI companies and their customers are betting the farm on Generative AI before we know how to assess and control the validity of its outputs. And we’re already seeing it show up everywhere in our digital lives, spreading misinformation and dangerous solutions to mundane problems.
Touching Grass: Grounding AI in Reality
To mitigate the limitations of Generative AI, we need new approaches to keep systems connected and responsive to the real world instead of their training data. We need to give AI the ability to "touch grass." Touching grass means stepping outside your digital bubble and engaging with the real world. It's a reminder to reconnect with the tangible, living context of the world around us. This same kind of “reality grounding” is essential for developing reliable and trustworthy AI systems.
Living systems have an innate ability to assess and respond to uncertainty in their environment. When faced with a decision, they consider the potential outcomes and weigh the certainty or uncertainty of each option. This constant evaluation of uncertainty is critical for perception, interaction, and learning.
But uncertainty is just one piece of the puzzle. Living systems also rely on internal models of the world to navigate their surroundings. These models are built on an understanding of cause and effect - how different factors influence each other and lead to specific outcomes. By constantly updating these causal models based on feedback from the environment, living systems align their understanding and behavior with the ever-changing realities they face.
It's an ongoing cycle of perception, interaction, learning, and adaptation. Living systems perceive their environment, interact with it based on their current models, learn from the outcomes of those interactions, and then adapt their models accordingly. This continuous loop is what allows living systems to thrive in diverse and dynamic contexts. It keeps them grounded in reality, even as the world around them shifts and changes.
This kind of grounding is missing in Generative AI.
Active Inference: A Framework for Grounded AI
In recent years, physicist and neuroscientist Karl Friston and his colleagues have formalized these underlying principles into a mathematical framework known as Active Inference. Active inference is a framework that describes how living organisms perceive, learn, and interact with their environment. It supposes that organisms constantly update their internal causal models of the world based on sensory inputs and use these models to generate actions that minimize surprise and maximize their chances of survival.
This framework offers a new lens through which we can understand how dynamic emergent systems can perceive, learn, and adapt to their surroundings. By grounding these “natural intelligence” systems in the real world, we can begin to understand and predict how complex systems behave within their contexts. Through this lens, we can start to explore how the principles of Active Inference offer a potential solution to mediating and mitigating the faults of Generative AI.
Integrating Active Inference with Generative AI
To create more transparent and trustworthy AI systems, we can leverage Active Inference to help Generative AI develop a sense of causal understanding and uncertainty about its own outputs in local contexts. One approach involves creating a mediating Active Inference layer that grounds an AI's interactions in reality, acting as a filter to constrain the model's outputs and align them with the specific knowledge and context relevant to the user or task at hand.
Active Inference allows AI systems to reason about the underlying relationships between variables and make informed decisions based on the causal structure within a complex system. By explicitly representing causal dependencies, Active Inference enables efficient belief updating, action selection, and model refinement. Moreover, by incorporating uncertainty estimation, AI systems can quantify the confidence in their predictions and actively seek out additional information when necessary.
Real-World Applications of Active Inference in AI
Let's explore a few examples where Active Inference has the potential to transform various domains where AI is applied.
1. Energy Grid Orchestration: An AI agent responsible for managing energy distribution across a smart grid can utilize a causal graph that models the relationships between energy supply, demand, weather conditions, and grid performance. By reasoning about the causal impact of different energy allocation strategies on grid stability and considering the uncertainty associated with each action, the agent can make informed decisions and seek human input when necessary.
2. Cybersecurity Threat Detection: An AI system employing Active Inference for detecting security breaches can use a causal graph that captures the relationships between network events, user behaviors, and potential threat indicators through time. By reasoning over causal dependencies and quantifying the uncertainty of its predictions, the system can identify anomalous patterns, adapt to evolving threat landscapes, and actively seek feedback from security experts to refine its detection capabilities.
3. AI Datacenter Workload Optimization: An AI agent optimizing workload distribution in a datacenter can construct a causal graph representing the dependencies between workload characteristics, resource utilization, and performance metrics. By leveraging the causal structure and incorporating uncertainty estimation, the agent can reason about the impact of different optimization strategies, adapt to variability in workload patterns, and ensure optimal resource allocation in real-time.
The Path Forward: Embracing Active Inference in AI Workloads
Since Generative AI is already here and will inevitably take a chunk of our attention and energy, we need tools to overcome its in inherent shortcomings. Active Inference offers a promising approach. By integrating principles that mimic the adaptive and self-updating nature of living organisms, AI systems can become more reliable, efficient, and grounded in reality.
