Generative AI Isn’t So Smart After All: Study
The threat that large language models will reach human levels of performance on common tasks is overblown, according to research co-authored by Meta’s top AI scientist Yann LeCunn. A new benchmark of real-world problems found that humans succeeded at 92% while ChatGPT-4 with plug-ins solved only 15%. A cynic would note that Meta has a strong interest in reducing pressure for AI regulation, and that other experts have argued the debate over human-like performance is a distraction from concrete problems that AI poses today.