That was my experience playing with the older neural net models back in the day. Usually the initial models and datasets are nearly the best it will get. Trying to feed it more data or trying to tweak the model only gains marginal improvements. Unless you've made a critical error in the initial work or miraculously happen upon a much better dataset then not much will change. I mean the whole premise that the machine finds the optimal solution within the limits of its capabilities. Beyond that you're, rolling the RNG until the output lands on a result you like better, but the capabilities remain the same.

That's why I suspect LLMs have peaked and these companies must be applying smoke and mirrors to keep the this iteration of AI going. There's not going to be another leap forward until scientists put out another revolutionary paper for everyone to copy. That seems to be the cadence of AI.