I do not know at all how AI works.
After checking out the first open AI system available to the public, ChatGPT, I am curious whether systems like this could contribute to scientific theory in the future.
Artificial Intelligence Stack Exchange is a question and answer site for people interested in conceptual questions about life and challenges in a world where "cognitive" functions can be mimicked in purely digital environment. It only takes a minute to sign up.Sign up to join this community
This already happened:
What directly comes to my mind is AlphaTensor that developed a novel method for matrix multiplication. Matrix multiplication is a computational expensive and often used operation. So it is safe to assume that over the past decades many humans have put effort into finding faster algorithms for this.
There are also scientific papers written by an AI. But without a close look at these papers, I would hesitate to claim that they contain novel scientific principles or theories.
Yes, from first principles.
Discovering new scientific theories is a mixture of following a set of mathematical logic principles and using creativity to come up with something new and useful.
Regarding the first criteria, logic and math, it is clear that computers are superior to humans in carrying out long chains of logic inferences, thanks to the precision of the hardware on which they are built.
Creativity is often mentioned as something differentiating humans from machines, but actually machines have access to the highest level of creativity: random number generators. Nothing is more creative than just randomly picking an option!
While current systems are still primitive in manipulating results and having the big picture necessary to put them to use, the state of the art is rapidly advancing. For instance, given the recent advances in NLP AIs will soon be able to chain together different concepts according to common patterns, such as the one used by scientists. Chain-of-Thought techniques are a good example. Now imagine gluing this NLP system to a source of truth, such as a tool (say, an equation solver or a microscope) able to test the proposed theory with an experiment. By brute forcing over many hypotheses you can imagine how these systems may eventually come up with relevant theories.
Assuming the above, that is LLMs or chatGPT-like systems are in principle able to formulate valid theories, are these systems in practice able to tame the combinatorial explosion in logic formulas arising when probing non trivial theories? I would lean on saying no, models which are limited to predicting the next word would be inefficient in performing the smart reasonings required to avoid combinatorial explosions, even though you may imagine that clever tricks could be devised to remind the language models the laws and mental models needed to trim the tree of possibilities at each step of the reasoning process.
What could do it then? I believe a substantial leap in capabilities would come from a fully multimodal model, able to also Chain-of-Thought reason with images, 3d representations, embodiments and physics simulations. In a sense these modalities are summarising many words at once, cutting down a continuum of possible inference chains. This is often called Visual Thinking.