Gary Marcus

GPT-4 successor models will be shown to have fundamental reliability limitations preventing deployment in high-stakes medical settings without heavy human oversight.

Current large language models will fail to achieve robust reasoning comparable to a 10-year-old child by 2027.