Artificial intelligence models are starting to succeed in science. In the past two years they have shown that they can analyze data, design paves and can even come up with new hypotheses. The pace of progress has convinced some researchers that artificial intelligence (AI) could compete in the coming decades with the greatest spirits of science.
Chatgpt broke the Turing test – The race is looking for new ways to judge AI
In 2016, Hiroaki Kitano, a biologist and chief executive at Sony AI, challenged researchers to achieve exactly that: to develop an AI system that is so advanced that it could make a discovery that is worthy of a Nobel Prize. Kitano called it the Nobel Turing -challenging and presented the aim as the big challenge for AI in science1. A machine wins if it can achieve a discovery on an equal foot with human research at top level.
That is not something that the current models can do. But by 2050, the Nobel Turing -challenging an AI system proposes that, without human intervention, combines the skills of generating hypotheses, experimental planning and data analysis to make a breakthrough that is worth a Nobel Prize.
It may not even take until 2050. Ross King, a researcher of chemical engineering at the University of Cambridge, VK, and an organizer of the challenge, thinks that such a ‘AI scientist’ could rise earlier to the laureate status. “I think it is almost certain that AI systems will be good enough to win Nobel Prizes,” he says. “The question is whether it will take 50 years or 10.”
But many researchers do not see how current AI systems, which are trained to generate words and ideas on the basis of the existing knowledge of man, can contribute new insights. Achieving such a performance can require drastic changes in how researchers develop AI and where AI financing goes. “If you saw a government program investing a billion dollars in fundamental research tomorrow, I think it would go much faster,” says Yolanda Gil, an AI researcher at the University of South California in Los Angeles.
Others warn that there are threatening risks to introduce AI to the research pipeline2.
Price -worthy discoveries
The Nobel Prizes are made to honor those who have “granted the biggest advantage” to humanity, as the namesake, Alfred Nobel, wrote in his will. For the scientific prices Bengt Nordén, a chemist and former chairman of the Nobel Commission for Chemistry, regards Chemistry, three criteria: a Nobel Department must be rich, deductible, are useful to be useful, deductible, feature is useful, deductible.
Although only living people, organizations and institutions are currently eligible for the prizes, AI has had earlier meetings with the Nobel Committee. In 2024, the Nobel Prize for Physics went to Pioniers of Machinal Learning that laid the foundation for artificial neural networks. In the same year, half of the chemical price recognized the researchers behind Alpafold, an AI system from Google DeepMind in London that predicts the 3D structures of proteins from their amino acid sequence. But these prices were not for the scientific steps behind AI systems – not for AI.
For an AI scientist to claim his own discovery, according to the Nobel Turing Challenge, the research should be carried out “complete or very autonomously”. The AI scientist should oversee the scientific process from start to finish and, according to Gil, decide on questions to answer, to turn experiments and analyze data.
Demis Hassabis (left) and John Jumper (center) won a Nobel Prize for the AI model Alphafold.Credit: Jonathan Nackstrand/AFP via Getty
Gil says that she has already seen AI tools that scientists help almost every step of the discovery process, which makes the field very exciting “. Researchers have shown that AI can help decode the speech of animals, supposed to be the origin of life in the universe and predicting when the future is able to help. The canal doddy to optimize quantum computers.
AI also starts to conduct experiments in itself. Gabe Gomes, a chemist at Carnegie Mellon University in Pittsburgh, Pennsylvania, and his colleagues designed a system called Coskientist who relies on large language models (LLMS), the species behind chatgpt and similar systems to plan complex chemical reactions3. And a non -depicted version of Cowscientist can do a computational chemistry with remarkable speed, says Gomes.
AI comes up with millions of new materials. Are they good?
One of Gomes’s students once complained that the software took half an hour to work out a transition state for a response. “The problem cost me for more than a year as a graduate student,” he says.
The Tokyo-based company Sakana AI uses LLMS in an attempt to automate machine learning research4. At the same time, researchers from Google and elsewhere investigate how chatbots can work in teams to generate scientific ideas.
Most scientists who use AI turn it as an assistant or type of employee, often appointed to specific tasks. This is the first of three waves of AI in science, says Sam Rodriques, Chief Executive of Futurehouse – a research laboratory in San Francisco, California, which debuted an LLM that was designed to perform chemistry tasks earlier this year. The and other ‘reasoning models’ learn to mimic a step-by-step logical idea, using a test and subscription process that relates to training on correct examples.
The existing models are useful employees who can make predictions based on data and can accelerate precise types of calculations otherwise. But they tend to need a person in the loop for at least one phase.
Secrets or Deepseek Ai Model Unveiled in Landmark Paper
Then, says Rodriques, AI will become better at developing and evaluating its own hypotheses by searching for literature and analyzing data. James, a biomedical data scientist at Stanford University in California, would have started moving to this empire. He and his colleagues recently demonstrated that a system built on LLMS can search biological data to find insights that researchers miss5. For example, when a published article and a data set of RNA sequences are linked with it, for example, the system thought that certain immune cells in people with COVID-19 are more likely to swell while they die, an idea that was not investigated by the authors of the article. It shows “that the AI agent is starting to find new things autonomously,” says would.
He also helps to organize a virtual meeting called Agents4Science later this month, which he describes as the first AI-all scientific conference. All articles are written and assessed by AI agents, in addition to human employees. And the one -day meeting will include invited conversations and panel discussions (of people) about the future of research generated by AI. Would say he hopes that the meeting will help researchers to assess how capable AI is in doing and revising innovative research.
AI helps AI researchers from Lawrence Berkeley National Laboratory to develop materials. Credit: Marilyn Sargent/Berkeley Lab
There are well -known challenges for such efforts, including the hallucinations that often bully LLMS, says. But he says that these problems can usually be remedied with human feedback.
Rodriques says that the final phase of AI in science, and what Futurehouse strives for, are models that can ask their own questions and design their own experiments and perform their own experiments – no one needed. He regards this as inevitably and says that AI could make a discovery that a Nobel could eventually make ‘in 2030’.
The most promising areas for a breakthrough – by an AI scientist or otherwise – are in material science or in the treatment of diseases such as Parkinson’s or Alzheimer’s, he says, because these are areas with great open challenges and an unfulfilled need.
