AI doesn’t know how to forgive and cannot forget

Machine Memory vs. Human Memory: The Forgotten Cost

The notion that “AI never forgets” is a common misconception. In reality, machine memory is a complex system that holds information in four different places, each with its own way of “forgetting.” The weights, context window, retrieval store, and logs all contribute to a machine’s ability to retain and recall information. However, unlike human memory, machine memory lacks the ability to forgive and forget.

This is not just a philosophical difference, but a technical one. Machine unlearning, or the removal of specific training data without retraining from scratch, is an unsolved problem at any real scale. This is because the weights of a model are smeared across billions of parameters, making it impossible to delete a specific piece of information. This is in stark contrast to human memory, which is capable of forgetting and forgiving through a process of active compression and reconsolidation.

The implications of this difference are significant. As we deploy perfect-recall systems into human relationships, we are introducing a new dynamic that is fundamentally at odds with the way humans interact. Second chances, which are a natural result of human forgetfulness, are not available to machines. This raises important questions about the consequences of deploying machines that remember everything at full fidelity forever.

The Technical Limitations of Machine Forgetting

Machine forgetting is a complex and multifaceted issue. Neural networks forget constantly, but they do so in a way that is different from human forgetting. Catastrophic forgetting, a phenomenon discovered by McCloskey and Cohen in 1989, occurs when a neural network is trained on a new task and overwrites the old one. This is a result of the way neural networks are designed, and it highlights the challenges of creating machines that can forget and forgive.

Furthermore, transformers, a type of neural network, forget by accident. The “Lost in the Middle” work by Liu et al. (2023) showed that models recall information placed at the beginning and end of a long context far better than the same information buried in the middle. This emergent dead zone is an unprincipled forgetting that lands wherever the architecture happens to be weak.

The technical limitations of machine forgetting are a result of the way machines are designed. Unlike humans, machines do not have a budget for memory, and storage is effectively free. This means that there is no pressure to prune, compress, and let go of information, which is a fundamental aspect of human memory.

The Winners and Losers of Perfect-Recall Systems

The deployment of perfect-recall systems has significant implications for human relationships. Those who benefit from these systems are likely to be individuals and organizations that value efficiency and accuracy above all else. However, those who are impacted by these systems, such as individuals who rely on human relationships and empathy, may find themselves at a disadvantage.

Furthermore, the use of perfect-recall systems raises important questions about accountability and responsibility. If machines are capable of remembering everything at full fidelity forever, who is responsible for the consequences of their actions? This is a complex issue that requires careful consideration and nuance.

The impact of perfect-recall systems on adjacent markets and job categories is also significant. As machines become increasingly capable of performing tasks that were previously the domain of humans, there is a risk of displacement and disruption. This highlights the need for careful consideration and planning as we deploy these systems.

The Skeptical Case: The Limits of Machine Forgetting

While the idea of perfect-recall systems may seem appealing, it is essential to consider the limitations of machine forgetting. Machines are not capable of forgiving and forgetting in the same way that humans are. This is a fundamental difference that has significant implications for the way we deploy these systems.

Furthermore, the assumption that machines can be designed to forget and forgive is based on a flawed understanding of human memory. Human memory is a complex and multifaceted system that is capable of forgiving and forgetting through a process of active compression and reconsolidation. Machines do not have this capability, and it is unclear whether they ever will.

The Signal to Watch Next: The Development of Machine Forgetting

As we move forward with the development of perfect-recall systems, it is essential to watch the development of machine forgetting. This is a critical area of research that has significant implications for the way we deploy these systems.

One key area to watch is the development of machine unlearning. This is an unsolved problem at any real scale, and it is essential to consider the implications of machines that can forget and forgive. Furthermore, the development of transformers and other neural networks that are capable of forgetting by accident is also an area to watch.

What’s your take on this? Drop your perspective in the comments below.

By Alex Mercer, Senior Tech Analyst at TrendFlashy