This article was provided to us with kind permission from Stephen Wolfram: Stephen Wolfram (2024), “Can AI Solve Science?”, Stephen Wolfram Writings.
Advertisement
Especially in light of recent surprise successes, there is widespread belief that artificial intelligence will one day be able to do “everything” – or at least everything we currently do. And what about science? Over the centuries, we humans have made incremental progress, gradually building what is now essentially the greatest intellectual edifice of our civilization. But despite our best efforts, there are still all sorts of scientific questions that remain unanswered. Can artificial intelligence now come along and simply solve them all?
When it comes to the question of whether AI can completely solve science, the answer is inevitably and resoundingly no. However, this does not mean that AI cannot make a significant contribution to scientific development. Large Language Models (LLMs), for example, offer an innovative linguistic interface to computer technologies, such as those implemented in the Wolfram Language. Because of their understanding of established scientific knowledge, LLMs have the potential to work at the highest level like an intelligent autocompleter, suggesting conventional answers or next steps in scientific processes.
However, I would like to address deeper questions about AI in science here. Three centuries ago, the idea of using mathematics to represent the world transformed science. And today we are in the midst of a major shift towards a fundamentally computer-based representation of the world (and yes, that is exactly what our computer language Wolfram Language is all about). So what about AI? Should we view it essentially as a practical tool that provides access to existing methods, or does it offer something fundamentally new to science?
Stephen Wolfram is a British physicist, mathematician and entrepreneur, known for developing the Wolfram Language and Mathematica, as well as his book “A New Kind of Science”.
What can you expect from AI in science?
I want to explore what AI can and cannot do in science. To do this, I look at specific, simplified examples to make the key points clear. I will share my thoughts and expectations based on our observations so far. I also discuss some theoretical and philosophical aspects about what is possible and what is not.
What do we actually mean by AI? In the past, anything seriously related to computers was often considered artificial intelligence. In this case, what we have long been doing with our Wolfram Language computer language would in particular be considered AI – as would all my studies of simple programs in the computer universe. But here I will mostly use a narrower definition and say that AI is something based on machine learning (and usually implemented with neural networks) that is gradually trained based on examples given to it. Additionally, I often take into account, that the training examples contain either a large amount of human scientific texts or real world observations. This means that AI not only learns raw, but also from a wide range of knowledge based on human insights.
Now we have a common definition of AI. But what does science mean and what does it mean to do science? At its core, it is about capturing or translating phenomena that exist “out there in the world” (usually in the natural world) into concepts that can be thought about or thought through. However, the actual scientific work follows several, quite different and widespread “work processes”. Some are concerned with prediction: predicting future events based on observed behavior; a model is sought that clearly describes how a system will behave; Concrete implications should be derived from an existing theory. Other work processes focus on explanations: a behavior is translated into an explanation that people can understand; Analogies between different systems or models are sought. Still other workflows focus on the creation of something new: the discovery of something that has certain properties; the discovery of something “interesting.”
Below, I'll explore these workflows in more detail and show how they can (or can't) be transformed by AI. But before I begin, I want to discuss something that overshadows any attempt to “solve the science”: the phenomenon of computational irreducibility.