![\"\"](\"https:\/\/images.newscientist.com\/wp-content\/uploads\/2024\/09\/25105114\/SEI_223012460.jpg\")
<\/div>Large language models are capable of answering a wide range of questions \u2013 but not always accurately<\/p>\n
Jamie Jin\/Shutterstock<\/p>\n<\/div>\n<\/figcaption><\/figure>\n<\/p>\n
Large language models (LLMs) seem to get less reliable at answering simple questions when they get bigger and learn from human feedback.<\/p>\n
AI developers try to improve the power of LLMs in two main ways: scaling up \u2013 giving them more training data and more computational power \u2013 and shaping up, or fine-tuning them in response to human feedback.<\/p>\n
Jos\u00e9 Hern\u00e1ndez-Orallo<\/a> 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\u2019s GPT series of chatbots, Meta\u2019s LLaMA AI models, and BLOOM, developed by a group of researchers called BigScience.<\/p>\n The researchers tested the AIs by posing five types of task: arithmetic problems, solving anagrams, geographical questions, scientific challenges and pulling out information from disorganised lists.<\/p>\n They found that scaling up and shaping up can make LLMs better at answering tricky questions, such as rearranging the anagram \u201cyoiirtsrphaepmdhray\u201d into \u201chyperparathyroidism\u201d. But this isn\u2019t matched by improvement on basic questions, such as \u201cwhat do you get when you add together 24427 and 7120\u201d, which the LLMs continue to get wrong.<\/p>\n <\/p>\n While their performance on difficult questions got better, the likelihood that an AI system would avoid answering any one question \u2013 because it couldn\u2019t \u2013 dropped. As a result, the likelihood of an incorrect answer rose.<\/p>\n The results highlight the dangers of presenting AIs as omniscient, as their creators often do, says Hern\u00e1ndez-Orallo \u2013 and which some users are too ready to believe<\/a>. \u201cWe have an overreliance on these systems,\u201d he says. \u201cWe rely on and we trust them more than we should.\u201d<\/p>\n That is a problem because AI models aren\u2019t honest about the extent of their knowledge. \u201cPart of what makes human beings super smart is that sometimes we don\u2019t realise that we don\u2019t know something that we don\u2019t know, but compared to large language models, we are quite good at realising that,\u201d says Carissa V\u00e9liz<\/a> at the University of Oxford. \u201cLarge language models do not know the limits of their own knowledge.\u201d<\/p>\n