While I'm no expert on human learning, I highly doubt we learn using LLM-like techniques. For one, machine learning models require massive amounts of training data, where as humans will learn with minimal inputs. Our ML algorithms are also very simplistic compared to what the brain's actual functioning could be. According to the orchestrated objective reduction hypothesis, our consciousness arises from quantum processes in the brain, not directly from synapse interactions. Which, considering how much trouble we're having explaining ourselves with just synapses, might very well be the case. So if we learn by something as simplistic as "guess answer, adjust weights, guess again, adjust weights, etc", our learning methods seem out of step with the complexity in our brains.