That's an assumption. Most/all neural network based tech faces a similar problem of exponentially diminishing returns. You get from 0 to 80 in no time. A bit of effort and you eventually ramp it up to 85, and it really seems the goal is imminent. Yet suddenly each percent, and then each fraction of a percent starts requiring exponentially more work. And then you can even get really fun things like you double your training time and suddenly the resultant software starts scoring worse on your metrics, usually due to overfitting.

And it seems, more or less, clear that the rate of change in the state of the art has already sharply decreased. So it's likely LLMs have already entered into this window.