Large language models’ abilities are doubling every seven months, new research suggested.
It’s hard to gauge the relative performance of LLMs, IEEE Spectrum said, because their main output is written text, making objective assessment difficult. Instead researchers gave models tasks of differing complexity, as measured by how long humans take to complete them, and seeing how reliably the models completed them.
The researchers found LLMs’ capabilities were improving exponentially, far outpacing Moore’s law, which says chips double in power roughly every two years.
One of the study’s authors warned against naive extrapolation, but that if the trend continues, artificial intelligence could be taking on months-long projects by the early 2030s, and eventually “you might not need human workers.”