Large Language Models (LLMs) are entirely controllable through human prompts and lack ’emergent abilities’; that is, the means to form their own insights or conclusions. Increasing model size does not lead LLMs to gain emergent reasoning abilities, meaning they will not develop hazardous abilities and therefore do not pose an existential threat. A new study sheds light on the (until now unexplained) capabilities and shortcomings of LLMs, including the need for carefully engineered prompts to exhibit good performance.
