With the invention of the internet, one thing became common, the interaction between computers and thereby people increased. These days there is a growing trend of using APIs to interact with computers running different kinds of software to seamlessly deliver a service. It is becoming more of an orchestration between computers rather than a single point service delivery.
Let's take it a step further, what if we codify API response codes into hex values? And these values are specific to every computer? What I mean is this: while the major API methods remain the same, every computer or rather the IP has its little hex digit for itself. This can be random or generated, but remains unchanged and non-transferable. Now imagine again that a group of computers talk to each other with this quirk. We have now successfully invented what in the land of linguistics is called a Pidgin language where all the computers in the group know exactly what this special character means and tries to perform / retrieve the certain action within the group.
Taking it a step further, let us now also have RNNs that are pre-programmed into the group of computers that are interacting with each other. The more a computer uses its specific hex code and the more "attention" it has to using this code to perform a certain operation, the more likely it is to be picked up with the other members of the group. In this way digital language takes shape. One major disadvantage of this way of networking / communication is the loss of control to us, programmers. These networks will essentially become black boxes, albeit faster and perhaps better.
- the RNNs can be the same number of parameters / variables but with a random distribution of attention for the RNNs in the initial phase.
attention could also be used to reinforce memory much like the human experience, the more successful the codes are, the more likely it is used, and has more attention to it.
the GPT-3 computerphile video: https://www.youtube.com/watch?v=_8yVOC4ciXc
as with human beings, there is communication and then action based on the communication. i.e information and action on that information. These are two distinct things. While a computer can memorise (GPT-3 esp.) addition in the memorisation sense, it does not link that with the binary operation, this therefore calls for another approach on top of the GPT to translate from NLP type models or "action" based models. These actions are inherently internal communications and therefore are not needed to be communicated with the outside world, while the results can be.