Computer science and technology

Using a new technique developed to examine the risks of multimodal large language models used in robots, MIT researchers were able to have a “simulated robot arm do unsafe things like knocking items off a table or throwing them by describing actions in ways that the LLM did not recognize as harmful and reject,” writes Will Knight for Wired. “With LLMs a few wrong words don’t matter as much,” explains Prof. Pulkit Agrawal. “In robotics a few wrong actions can compound and result in task failure more easily.”

Researchers from MIT and elsewhere have compared 12 large language models (LLMs) against 925 human forecasters for a three-month forecasting tournament to help predict real-world events, including geopolitical events, reports Tomas Gorny for Forbes. "Our results suggest that LLMs can achieve forecasting accuracy rivaling that of human crowd forecasting tournaments,” the researchers explain.

Writing for Forbes, Senior Lecturer Guadalupe Hayes-Mota SB '08, MS '16, MBA '16 shares insight into how entrepreneurs can use AI to build successful startups. AI “can be a strategic advantage when implemented wisely and used as a tool to support, rather than replace, the human touch,” writes Hayes-Mota. 

Prof. Armando Solar-Lezama speaks with New York Times reporter Sarah Kessler about the future of coding jobs, noting that AI systems still lack many essential skills. “When you’re talking about more foundational skills, knowing how to reason about a piece of code, knowing how to track down a bug across a large system, those are things that the current models really don’t know how to do,” says Solar-Lezama.

Prof. Daniela Rus, director of CSAIL, and Prof. Russ Tedrake speak with the Financial Times about how advances in AI have made it possible for robots to learn new skills and perform complex tasks. “All these cool things that we only dreamed of, we can now begin to realize,” says Rus. “Now we have to make sure that what we do with all these superpowers is good.”

Forbes contributor Michael T. Nietzel spotlights the newest cohort of Rhodes Scholars, which includes Yiming Chen '24, Wilhem Hector, Anushka Nair, and David Oluigbo from MIT. Nietzel notes that Oluigbo has “published numerous peer-reviewed articles and conducts research on applying artificial intelligence to complex medical problems and systemic healthcare challenges.” 

Yiming Chen '24, Wilhem Hector, Anushka Nair, and David Oluigbo have been named 2025 Rhodes Scholars, report Brian P. D. Hannon and John Hanna for the Associated Press. Undergraduate student David Oluigbo, one of the four honorees, has “volunteered at a brain research institute and the National Institutes of Health, researching artificial intelligence in health care while also serving as an emergency medical technician,” write Hannon and Hanna.

Research from the Data Provenance Initiative, led by MIT researchers, has “found that many web sources used for training AI models have restricted their data, leading to a rapid decline in accessible information,” reports Gary Drenik for Forbes. 

Researchers at MIT have developed a new AI model capable of assessing a patient’s risk of pancreatic cancer, reports Erez Meltzer for Forbes. “The model could potentially expand the group of patients who can benefit from early pancreatic cancer screening from 10% to 35%,” explains Meltzer. “These kinds of predictive capabilities open new avenues for preventive care.” 

Arago, an AI startup co-founded by alumnus Nicolas Muller, has been named to the Future 40 list by Station F, which selects “the 40 most promising startups,” reports Romain Dillet for TechCrunch. Arago is “working on new AI-focused chips that use optical technology at the chipset level to speed up operations,” explains Dillet.

Craft in America visits Prof. Erik Demaine and Martin Demaine of CSAIL to learn more about their work with computational origami. “Computational origami is quite useful for the mathematical problems we are trying to solve,” Prof. Erik Demaine explains. “We try to integrate the math and the art together.”

Neural Magic, an AI optimization startup co-founded by Prof. Nir Shavit and former Research Scientist Alex Matveev, aims to “process AI workloads on processors and GPUs at speeds equivalent to specialized AI chips,” reports Kyle Wiggers for TechCrunch. “By running models on off-the-shelf processors, which usually have more available memory, the company’s software can realize these performance gains,” explains Wiggers. 

Researchers at MIT have developed a “new type of transistor using semiconductor nanowires made up of gallium antimonide and iridium arsenide,” reports Alex Knapp for Forbes. “The transistors were designed to take advantage of a property called quantum tunneling to move electricity through transistors,” explains Knapp. 

Researchers at MIT have developed a new virtual training program for four-legged robots by taking “popular computer simulation software that follows the principles of real-world physics and inserting a generative AI model to produce artificial environments,” reports Jeremy Hsu for New Scientist. “Despite never being able to ‘see’ the real world during training, the robot successfully chased real-world balls and climbed over objects 88 per cent of the time after the AI-enhanced training,” writes Hsu. "When the robot relied solely on training by a human teacher, it only succeeded 15 per cent of the time.”

Research Scientist Nick van der Meulen speaks with Financial Times reporter Bethan Staton about how automation could be used to help employers plug the skills gap. “You can give people insight into how their skills stack up . . . you can say this is the level you need to be for a specific role, and this is how you can get there,” says van der Meulen. “You cannot do that over 80 skills through active testing, it would be too costly.”