Google has presented FunSearch as a new AI for the field of scientific research, achieving its first milestones within mathematics.
2023 can be considered the year of AI, this can be seen in what has been the most searched term on Google or the most visited Wikipedia entry. The advances made have allowed the launch of Copilot or Gemini, but AIs at the service of ordinary users are not the only ones that exist and the new FunSearch AI for the scientific field is proof of this.
FunSearch comes from Google DeepMind. This new research method has already borne fruit, managing this week to do discoveries in open mathematical problems through the use of LLMs or large language models, being able to generate, evaluate and verify that the created code works.
FunSearch: a new AI tool to make discoveries in science
The way FunSearch operates uses a specific LLM, although since it is from Google it is PaLM 2, this artificial intelligence language model. is trained within the area to which he will be exposed and is integrated into the operating scheme developed for problem resolution.
This scheme serves as a researcher, developer and evaluator. FunSearch aims to generate solutions, test that they work and evaluate their operation. You could say that it is a research machine in a loop fed back by the information with which it has been trained and by the erroneous solutions found.
The solutions, as we have indicated, may or may not work. But this is part of the process and in this case it is the key to how FunSearch works, When working on programming code what is done is to test the generated solutions and check their viability. What is discarded serves to indicate the paths that should not be followed.
This learning or he feedback obtained from FunSearch is capable of generating a constant improvement process in which the solutions that have worked can be implemented in the large language model (LLM) in order to continue their evolution and obtain better solutions or in a shorter time.
The FunSearch application has already shown that it is capable of obtaining new discoveries in scientific matters. Despite all the capabilities of FunSearch, it also has limitations such as All the information that must be fed to this AI must be generated by humans.
FunSearch cannot work on pure concepts, it is necessary to establish your own knowledge and define what you want to achieve. AI will not be able to dedicate itself to researching a completely new field. FunSearch is a tool with which facilitate research through an evolutionary process of finding solutionsopening up the use of AI generally in research.
To always be up to date with the latest in technology, subscribe to our official and verified Andro4all channel on WhatsApp.