You don't think it does because what you described is only the optimistic take on how much farther LLMs will be able to advance :). The pessimists would look at previous "AI" and point out each new approach quickly rises to prominence and then drastically tapers off in how much more can be squeezed out of it.
I'm somewhere in the middle. I think there is more to squeeze out of LLMs still, but not nearly the kind of growth we had from GPT-2 to multimodal reasoning models. Part of the equation is, as you say, a willingness to experiment on radical ideas. The other part is a willingness to find when the growth curve is slowing rather than bet it will always grow enough for a novel architecture lead to be meaningful.
I'm not sure I think the progress is about multimodality. After all, Mistral's approach hasn't involved multimodality, and they've kept up.
An efficient model, then data curation, then post-training. Where things are slowing down is of course necessary to know to be efficient, at least in the short term competition.
I'm somewhere in the middle. I think there is more to squeeze out of LLMs still, but not nearly the kind of growth we had from GPT-2 to multimodal reasoning models. Part of the equation is, as you say, a willingness to experiment on radical ideas. The other part is a willingness to find when the growth curve is slowing rather than bet it will always grow enough for a novel architecture lead to be meaningful.