
Francois Chollet, Mike Knoop - LLMs won’t lead to AGI - $1,000,000 Prize to find true solution
Dwarkesh Podcast
Mix of Pattern Matching and Reasoning in Learning
In learning processes, individuals typically engage in a mixed approach of pattern matching and reasoning. For example, even in tasks that may appear to heavily rely on reasoning like proving a mathematical theorem, intuition, pattern matching, and prior knowledge strongly influence the process. This blending of approaches is evident in the utilization of both type one and type two thinking. Larger models tend to be more sample efficient because they possess a wide array of reusable building blocks that assist in recognizing new patterns within the training data. The model's ability to continue recognizing patterns effectively as it grows depends on the alignment of new patterns with its existing knowledge base.