In a new survey, 76% of scientists said that scaling large language models was "unlikely" or "very unlikely" to achieve AGI.