Science

Large language models are capable of answering a wide range of questions – but not always accuratelyJamie Jin/Shutterstock Large language models (LLMs) seem to get less reliable at answering simple questions when they get bigger and learn from human feedback. AI developers try to improve the power of LLMs in two main ways: scaling up – giving them more training data and more computational power – and shaping up, or fine-tuning them in response to human feedback. José Hernández-Orallo at the Polytechnic University of Valencia, Spain, and his colleagues examined the performance of LLMs as they scaled up and shaped up. They looked at OpenAI’s GPT series of chatbots, Meta’s LLaMA AI models, and BLOOM, developed by a group of researchers called BigScience. The research...

Growing up in the US during the oil embargo of the early 1970s, I was bombarded by public service announcements encouraging people to conserve energy. But at a very young age, I also read that “energy is always conserved”, according to physics. This baffled me. If nature automatically conserves energy, why would human efforts to do so be needed? I soon realised that physicists don’t exactly speak English. They employ a dialect full of familiar-sounding terms with unfamiliar meanings (including “conserve” and “energy”). Worse still, many words, including simple ones like “force” and “mass”, don’t even signify what physicists originally intended. Consequently, the language we use to talk about physics obscures some of our most beautiful and fascinating discoveries about...

Last year was the hottest on record. Around the world, we are witnessing more extreme events—from massive forest fires to floods to “hot tub’” ocean temperatures—with devastating consequences for human life and our planet’s biodiversity. The economic costs are already tremendously high. Despite mounting feelings of eco-anxiety, my colleagues and I firmly believe that there is still much that can be done to save our planet.  There is still a lot of uncertainty in climate projections; shortcomings in both climate process understanding and computing capacities limit the accuracy of climate projections needed for optimal adaptation and resilience to climate change. Using AI helps to narrow down the uncertainties in climate modeling that can, in turn, be used to optimize adaptation, and pot...

Long-lasting nuclear batteries could provide power on remote missions like deep-sea explorationEB Adventure Photography/Shutterstock A nuclear battery powered by radioactive decay rather than chemical reactions could last for decades. The most efficient design yet may bring this concept closer to reality. Researchers have wanted to use radioactive atoms to build exceptionally long-lasting and damage-resistant batteries since the 1900s. While some prototypes have been assembled and even used in space missions, they were not very efficient. Now Shuao Wang at Soochow University in China and his colleagues have improved the efficiency of a nuclear battery design by a factor of 8000. They started with a small sample of the element americium, which is usually considered...

The CMS detector at the Large Hadron ColliderSciTech Image/James King-Holmes/Alamy Stock Photo A possible crack in the standard model of particle physics seems to be shrinking, as new data from CERN’s Large Hadron Collider (LHC) contradicts a previous puzzling result that had physicists excited about the possibility of new, exotic physics – but some mysteries remain. “The standard model survives for the moment,” Josh Bendavid at the Massachusetts Institute of Technology told a packed seminar room at CERN, the particle physics laboratory near Geneva, Switzerland, on 17 September. He was presenting new data on the mass of the W boson, a fundamental particle that is crucial for processes like nuclear decay and setting the mass of the Higgs boson. Questions about the ...