Theorize how In-Context Learning (ICL) functions in Large Language Models. Discuss the hypothesis that ICL is actually meta-learning or Bayesian inference performed at inference time. Compare and contrast this with traditional gradient-based learning theories.