Stephen Wolfram: Can AI Solve Science?

The speaker begins by introducing the topic of AI's potential in solving scientific problems. They discuss the general belief in AI's expanding capabilities and question whether AI can solve all of science's unanswered questions. The speaker clarifies that while AI will not solve everything, it can significantly aid scientific progress, especially with its ability to provide high-level autocomplete for scientific work. The discussion then shifts towards the transformative ideas in science, such as the mathematical representation of the world and the current shift towards computational representation, which is central to AI's function. The speaker aims to explore AI's limitations and potential in science through specific examples and theoretical discussions.

The speaker delves into the concept of computational irreducibility, which suggests that certain systems require step-by-step computation to understand their behavior, and no shortcut can predict their outcomes. This principle, supported by physics, implies that AI, like humans, cannot infinitely predict or solve a system's behavior without performing the necessary computations. Despite this, science is possible because there are 'pockets' of computational reducibility where limited computational effort can reveal insights. The speaker also touches on the historical resonance of computational irreducibility with Gödel's incompleteness theorems, suggesting that science, like mathematics, cannot be fully mechanically solved.

The speaker shares personal experiences using computation for scientific discovery, such as finding unexpected behaviors in simple systems through enumeration. They discuss AI's potential in making these discoveries by systematically exploring possibilities, which can lead to creative and surprising results. The speaker also addresses AI's ability to predict outcomes, which historically has been a significant aspect of successful science. They illustrate this with examples of using AI for inductive inference and the challenges AI faces in making accurate predictions, such as the limitations of neural networks in forecasting complex systems.

The speaker explores the limitations of AI, particularly neural networks, in modeling and predicting data. They use the example of a spring's motion to demonstrate how neural networks can fail to predict future behavior accurately, even with extensive training. The discussion highlights the inherent structural assumptions in AI models and how different activation functions can influence prediction outcomes. The speaker also touches on the idea that AI's predictions are often based on pattern matching rather than an understanding of the underlying science, suggesting that AI's predictive capabilities are limited by the data and structure it is trained on.

The speaker discusses AI's role in predicting protein folding, a complex scientific problem. They suggest that AI's success in this area might be due to its ability to identify significant aspects of protein structure that align with human-defined criteria for success. The speaker also considers the possibility that AI might be tapping into undiscovered regularities in protein folding, representing new scientific discoveries. However, they note that AI's predictions are often only roughly correct and may not capture the full complexity of the folding process, especially for proteins with more complicated structures.

The speaker considers whether AI can streamline the process of solving equations and scientific problems. They discuss traditional mathematical approaches and how AI might offer an alternative by learning patterns from data. The speaker uses the example of the three-body problem in physics to illustrate the challenges AI faces in providing accurate solutions. They suggest that while AI can handle large datasets and complex equations more efficiently than traditional methods, it still struggles with the intrinsic complexity and computational irreducibility of certain problems.

The speaker explores AI's potential in handling multi-computational processes, such as finding paths in game graphs or proving theorems. They discuss how AI might be used to estimate distances or scores in these processes, guiding the search for solutions. The speaker also highlights the limitations due to multicomputational irreducibility, where AI may not always find optimal solutions. They provide examples of how AI can assist in these processes, such as using neural networks to predict distances or to guide the search for paths, but also note that AI's effectiveness is constrained by the inherent complexity of the problems.

The speaker discusses AI's role in creating scientific narratives, which are essential for human understanding of scientific discoveries. They emphasize the need for AI to identify 'waypoints' or familiar concepts that can be used to construct a narrative. The speaker suggests that AI can assist in this process by reducing complex data to interpretable parameters and by using language models to generate explanations. However, they also note the challenges in aligning AI's internal thought processes with human scientific concepts and the limitations of AI in exploring uncharted conceptual spaces.

The speaker addresses the challenge of defining what is 'interesting' in science, particularly in the context of AI's ability to identify anomalies or surprising patterns. They discuss how AI can be trained to recognize typical distributions of data and highlight outliers, which may be considered interesting. The speaker also considers the limitations of AI in determining interestingness, as this is often a matter of human judgment and context. They suggest that AI can assist in identifying potential areas of interest but that the ultimate decision on what is scientifically interesting remains a human responsibility.

The speaker discusses the concept of 'inter-concept space,' which refers to the vast space of possible computations and ideas that lie between familiar concepts. They explore the idea that AI, particularly when trained on large datasets, can operate in this inter-concept space, generating outputs that may not have a clear human narrative or interpretation. The speaker suggests that while AI can produce novel outputs, these may not necessarily be considered interesting or relevant from a human perspective, highlighting the challenge of aligning AI's operations with human understanding and interests.

The speaker considers AI's potential in identifying patterns and making connections across diverse scientific disciplines. They suggest that AI's ability to process large volumes of data and recognize patterns could facilitate the discovery of analogies and connections that might be difficult for humans due to the high level of specialization in modern science. The speaker also discusses the potential for AI to assist in formalizing these interdisciplinary connections, contributing to a more integrated approach to scientific discovery.

The speaker reflects on AI's role in the future of scientific discovery, emphasizing the potential for AI to assist in various scientific workflows, particularly in identifying patterns, suggesting questions, and providing approximate solutions. They acknowledge the limitations of AI, especially in the face of computational irreducibility, but also highlight the opportunities for AI to contribute to science by leveraging computational reducibility. The speaker concludes by suggesting that the combination of AI and the computational paradigm, including formal computational languages, may offer the greatest potential for advancing scientific discovery.