Each time a brand new AI mannequin is launched, it’s usually touted as acing its efficiency towards a sequence of benchmarks. OpenAI’s GPT-4o, for instance, was launched in Could with a compilation of outcomes that confirmed its efficiency topping each different AI firm’s newest mannequin in a number of exams.
The issue is that these benchmarks are poorly designed, the outcomes onerous to duplicate, and the metrics they use are incessantly arbitrary, in accordance with new analysis. That issues as a result of AI fashions’ scores towards these benchmarks decide the extent of scrutiny they obtain.
AI corporations incessantly cite benchmarks as testomony to a brand new mannequin’s success, and people benchmarks already kind a part of some governments’ plans for regulating AI. However proper now, they won’t be ok to make use of that manner—and researchers have some concepts for a way they need to be improved.
—Scott J Mulligan
We have to begin wrestling with the ethics of AI brokers
Generative AI fashions have turn out to be remarkably good at conversing with us, and creating pictures, movies, and music for us, however they’re not all that good at doing issues for us.
AI brokers promise to vary that. Final week researchers printed a brand new paper explaining how they educated simulation brokers to duplicate 1,000 individuals’s personalities with gorgeous accuracy.
AI fashions that mimic you might exit and act in your behalf within the close to future. If such instruments turn out to be low-cost and simple to construct, it is going to increase a lot of new moral issues, however two particularly stand out. Learn the total story.
—James O’Donnell
