Why AI Needs a Mental Model of the Physical World
View original- current
Summary (TL;DR)
AI language models like ChatGPT lack a mental model of the physical world, which limits their ability to verify accuracy and reason about cause and effect. They rely on patterns in human text, not simulations. Adding an internal physics simulation could allow AI to test assumptions, improve accuracy, and even surpass human reasoning. Learning physics from textbooks is insufficient because abstractions miss real-world complexity. A mental model of the physical world could be the next big leap in AI, enabling novel insights rather than just pattern matching.
Artificial intelligence has come a long way in recent years, with many breakthroughs in natural language processing, image recognition, and decision-making. However, practical differences between AI and humans still affect their ability to reason and make accurate conclusions. In this post, we will explore the importance of a mental model of the physical world for AI and how it can impact the accuracy of their answers.
The Limitations of AI
Unlike humans, AI language models do not have a physical simulation of the world to compare against; they find patterns in words and create connections from inputs. This lack of a mental model of the physical world can limit the ability of AI to verify the accuracy of their responses. Even advanced AI models like ChatGPT can provide erroneous answers or assumptions, and nobody knows how the AI arrives at these conclusions. This is because the reasoning behind an answer cannot be debugged like regular software.
The Importance of a Mental Model of Physics
If the AI had the ability to test things in an internal simulation of the physical world, it could provide more accurate answers. With a mental model of the physical world, AI could verify assumptions like humans do when we verify an incorrect assumption made by AI to improve its accuracy in generating responses. This is because it would be able to simulate physical scenarios closer to how the human world works and test its conclusions against the results of those simulations.
The Knowledge Gap: Can AI Learn Physics Instead?
I was asked whether AI could gain a mental model of the physical world by studying physics. That approach is limited by the training data available to AI, which consists of human explanations and formulas. These abstractions represent a subset of the complexity of reality and cannot fully capture the intricacies of the physical world (The Map is Not the Territory). To truly understand the cause and effect of the physical world and how it relates to its patterns, AI needs to learn from actual simulations rather than relying solely on human interpretations of reality.
The idea is to find patterns humans haven't, not simply repeat what already exists, which is what AI language models do currently. Also, although you could input with a simulation of physics, you could also input with a simulation that includes other aspects of the real human world environment, such as a social system, an evolutionary system with natural selection, etc.
The Next Big Jump
Inducing a mental model of the physical world into an AI’s artificial brain could be the next big jump in artificial intelligence. This would allow AI to simulate physical scenarios and test their assumptions, leading to more accurate conclusions. AI could even exceed human reasoning capabilities in some areas if they test against a mental model of the physical world.
Does it make sense?
The lack of a mental model of the physical world is a limitation of AI that affects its ability to verify assumptions and make accurate conclusions, which we humans do ourselves after interacting with AI. If we can input a mental model of the physical world into its artificial brain, it could significantly improve its accuracy and capability to reason and make predictions. That would have the potential to create new ideas, not merely create connections and find patterns of information already there.
