A breakthrough discovery has been made by researchers in the field of artificial intelligence, revealing the potential of chatbots to generate information beyond human knowledge. Google DeepMind, a leading research facility, has claimed to achieve the world’s first scientific discovery using a big language model, such as OpenAI’s ChatGPT and Google’s Bard.
The research aimed to determine whether massive language models, which serve as the foundation for contemporary chatbots, can do more than simply repackage existing knowledge learned during training. Pushmeet Kohli, head of artificial intelligence for science at DeepMind, expressed their initial disbelief that the project would yield truly new results. However, they have now demonstrated that a large language model can make a genuine and novel scientific discovery.
Large language models, also known as LLMs, are highly efficient neural networks that analyze vast amounts of text and other data to learn language patterns, including computer code. Over the past year, chatbots like ChatGPT have been widely used for various purposes, including software debugging, generating content like college essays and travel itineraries, and even composing climate change poetry in the style of Shakespeare.
Despite their popularity, chatbots have been criticized for lacking the ability to provide new information and often producing flawed responses. However, DeepMind’s implementation of an LLM called FunSearch has shown promise in addressing this limitation. FunSearch is a computer software that offers computational solutions to challenges by incorporating an evaluator that ranks programs based on their performance. The best-performing programs are then integrated back into the LLM to further enhance its capabilities, gradually transforming weak programs into stronger ones capable of discovering new information.
Jordan Ellenberg, a professor of mathematics at the University of Wisconsin-Madison and a co-author of the research paper, expressed excitement about the future implications for human-machine interaction in mathematics. He highlighted that FunSearch does not merely provide solutions but generates programs that can find solutions. This allows humans to read and comprehend the programs, potentially inspiring new ideas for future challenges.
The researchers tasked FunSearch with two puzzles to demonstrate its capabilities. The first puzzle, the cap set problem, is an obscure and long-standing challenge in pure mathematics. It involves finding the largest set of points in space that does not contain three points on a straight line. FunSearch produced multiple programs that generated significantly larger cap sets than what mathematicians had previously developed.
This groundbreaking discovery by DeepMind opens up exciting possibilities for the role of chatbots and language models in generating new knowledge. While chatbots have faced criticism for their lack of originality, this research demonstrates their potential to contribute to scientific advancements. As the field of artificial intelligence continues to evolve, the collaboration between humans and machines in disciplines like mathematics holds great promise.
In conclusion, DeepMind’s research breakthrough showcases the ability of language models like ChatGPT to generate novel information, going beyond human knowledge. This groundbreaking discovery paves the way for future advancements in human-machine interaction and the potential for chatbots to contribute to scientific progress.
