even before this, i've often thought about machine learning methods as lossy compression schemes at their core. llms to me bear a striking resemblance to how my

even before this, i've often thought about machine learning methods as lossy compression schemes at their core. llms to me bear a striking resemblance to how my own human memory works - you have to "jog" it with a prompt, the recall is imperfect and it often takes the shortest apparent path to an answer unless you really try to take it step-by-step. and before you dismiss this theory you should consider that 1) during the course of its training the model gets exposed to a lot more information than any human over their entire lifetime; 2) the substrate on which the model runs, is more robust that squishy human neurons.