Maria del Rio-Chanona

Maria del Rio-Chanona, Nadzeya Laurentsyeva, Johannes Wachs

Large language models like ChatGPT efficiently provide users with information about various topics, presenting a potential substitute for searching the web and asking people for help online. But since users interact privately with the model, these models may drastically reduce the amount of publicly available human-generated data and knowledge resources. This substitution can present a significant problem in securing training data for future models. In this work, we investigate how the release of ChatGPT changed human-generated open data on the web by analyzing the activity on Stack Overflow, the leading online Q\&A platform for computer programming. We find that relative to its Russian and Chinese counterparts, where access to ChatGPT is limited, and to similar forums for mathematics, where ChatGPT is less capable, activity on Stack Overflow significantly decreased. A difference-in-differences model estimates a 16\% decrease in weekly posts on Stack Overflow. This effect increases in magnitude over time, and is larger for posts related to the most widely used programming languages. Posts made after ChatGPT get similar voting scores than before, suggesting that ChatGPT is not merely displacing duplicate or low-quality content. These results suggest that more users are adopting large language models to answer questions and they are better substitutes for Stack Overflow for languages for which they have more training data. Using models like ChatGPT may be more efficient for solving certain programming problems, but its widespread adoption and the resulting shift away from public exchange on the web will limit the open data people and models can learn from in the future.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.