They aren’t held up by any ancillary aids - it’s simply a very popular flavour of ANN which dominates the forefront of DL-based AGI research. As for “Do not learn”, I would go on to argue that they are the only systems to be able to do so: Pathways Language Model (PaLM): Scaling to 540 Billion Parameters for Breakthrough Performance – Google Research Blog The ability to logically infer things, understand complex jokes, translate between languages, manipulate concepts, write stories, or do a wide variety of tasks alone with a single model is evidence enough that if they aren’t learning, they are at least at this moment the only workable prototypes which appear to do so.
Very true, but in another aspect, DL models pay with high power but trade-off with superhuman memory which can potentially be very helpful in achieving the very long term goal of ASI; I feel its this symbiosis between memory (able to manipulate thousands of concepts, perform cross-correlation) and strong reasoning skills that would actually get AGI.