Nicholas Carlini on the future of large language models: the error bars should be wide

Nicholas Carlini writes arguing that expectations for the future of large language models should be held with broad uncertainty. He believes that within three to five years, language models might be capable of most economically useful cognitive tasks beyond human expert level, or they might only develop along a normal path of declining costs and gradual capability improvement without a fundamental paradigm shift. He uses his own underestimation of language model potential from 2018 to 2021 as an example of how fallible predictors are, and notes that current models can already solve early-PhD-level math problems, reach the coding level of top competitive programmers, and benchmarks are constantly being refreshed, so both extreme possibilities cannot be ignored.