A breakthrough that shows the technology underpinning ChatGPT and other similar programs can generate information that goes beyond what is known to humans has been discovered by researchers in the field of artificial intelligence. They claim to have made the world’s first scientific discovery utilizing a big language model.
The discovery was made by Google DeepMind, which is a research facility where researchers are examining whether or not massive language models, which are the foundation of contemporary chatbots like OpenAI’s ChatGPT and Google’s Bard, are capable of doing more than simply repackaging knowledge that was learnt during training and coming up with fresh insights.
According to Pushmeet Kohli, who is the head of artificial intelligence for science at DeepMind, “When we started the project, there was no indication that it would produce something that produces something that is truly new.” To the best of our knowledge, this is the very first time that a huge language model has demonstrated the ability to make a true and novel scientific discovery.
Large language models, also known as LLMs, are very effective neural networks that are able to learn the patterns of language, including computer code, by analyzing enormous amounts of text and other data for example. Since the sudden arrival of ChatGPT a year ago, the technology has been used to debug problematic software and produce a wide variety of content, ranging from college essays and travel itineraries to poetry about climate change written in the style of Shakespeare.
However, despite the fact that chatbots have become incredibly popular, they do not provide any new information and are prone to confabulation. As a result, the responses they provide are fluent and reasonable, but they are severely flawed. This is similar to the replies that are provided by the finest pub bores.
An LLM was utilized by DeepMind in order to develop “FunSearch,” which is an abbreviation for “searching in the function space.” FunSearch is a computer software that provides computational answers to challenges. An “evaluator” is paired with the LLM, and it is responsible for automatically ranking the programs based on how well they perform respectively. After that, the finest programs are integrated and then sent back to the LLM so that it may continue to progress. It is because of this that the system is able to gradually transform weak programs into more strong ones that are capable of discovering new information.
The following is a statement made by Jordan Ellenberg, a professor of mathematics at the University of Wisconsin–Madison and a co-author on the paper: “What I find really exciting, even more so than the specific results we found, is the prospects it suggests for the future of human-machine interaction in mathematics.”
“FunSearch does not generate a solution; rather, it generates a program that actually finds the solution. It is possible that a solution to a particular problem will not provide me with any insight on how to solve other problems that are related. But a computer program that discovers the answer is something that a human being can read and understand, and it is hoped that this will result in the generation of ideas for the next challenge, as well as the next, and the next, and the next.Both of the puzzles were given to FunSearch to solve by the researchers. The first was a task in pure mathematics that has been around for a long time and is considered to be obscure. It is known as the cap set problem. Finding the greatest set of points in space that does not consist of three points that make a straight line is the focus of this problem. FunSearch is responsible for the production of multiple programs that are capable of generating new huge cap sets that surpass the finest that mathematicians have developed.
