Brain Explored

Why its hard to make LLMs work for real-life usecases (not toy-benchmarks / demos): if you're gonna push a piece of "machinery" to the limit, and expect it to hold together – you have to have some sense of where that limit is. (that limit can't be read on twitter, cant be logically deduced – it has to be "felt" by actually pushing the machinery to the limit)

In the brain, some neurons adapt easily, while others remain resistant to change. This can be likened to our beliefs, where certain deeply ingrained convictions, like the historical belief that the sun revolves around the earth, required substantial evidence and effort to alter. Similarly, in machine learning, we could enhance efficiency by incorporating a "resistance to change" factor for each neuron during training. This factor would determine how readily a neuron or set of neurons can adapt or how firmly they maintain their learned patterns (or weights). By drawing on the concept that some brain regions or neuron patterns encode hard-set beliefs through repeated reinforcement, we can apply this principle to improve the adaptability and stability of artificial neural networks. This could be a pathway for the Alignment Problem. PS: Dropout is a somewhat overlapping, but a completely different concept, introduced for a different need.