I often wonder if the future of OpenAI is using GPT as a meta interface for the rest of their models. Interpret and extrapolate human intent and then select a more appropriate model to execute requests and return the asks in a human understandable format.
It seems that's indeed the direction. And it makes sense. When saying that their overarching goal is building _an AGI_ there's an implication that there is really only one model / interface / entry point, and that how it does its magic (one neural network, several similar neural networks, an ensemble of various neural networks of different architecture and training + some glue code to tie them together) is an implementation detail and not something we should concern ourselves with as consumers.
Isn't that sort of how GPT-4 already works? It's a Mixture of Experts architecture that bundles together a bunch of models and selects one for each query.
do you have a source for that info? (not accusing, i've heard this before, i just want to see it from openai) i'm curious because i believed that myself but i'm not sure how some of the developments from this release would work in that architecture