Grant Sanderson suggests a key advantage of AI in research is the ability to systematically reset an agent's context or spawn multiple agents with different contexts, allowing them to escape cognitive ruts that humans struggle with.

Grant Sanderson suggests that the autoregressive, token-by-token generation process of current LLMs is a significant bottleneck that inhibits their ability to make unlikely, creative connections between different fields of knowledge.

Grant Sanderson predicts that the most useful progress from AI models in mathematics in the next 5 years will be their ability to act as "supercharged connectors," finding and mapping connections between different fields.