Foundation modelif it works, do one thing and do it well, do everything well. Foundation modelif it doesn't work, do one thing and fail at it, fail at everything.
Foundation model ($m$) is colloquially known as common sense, or more precisely, it is the union of everyone ($i$)'s common sense ($c_i$).
$$ m\approx \bigcup_i c_i $$
Foundation models exist because humans themselves are proof of existence. In other words, the world that humans commonly understand ($w$) is called common sense, and common sense clearly exists.
$$ m =m(w) $$
Foundation models are unique (up to isomorphism[4]) because the understood world (earth, mathematics, physics, science, economics, basic history, basic human nature, basic morality, basic values...) is unique [1][2][3], in other words—common sense is unique. Therefore, regardless of how AGI is trained, it will ultimately converge to a unique limit $w$.
$$ \forall a,b \\\lim_{t \rightarrow \infty } m_a(w,t) \cong \lim_{t \rightarrow \infty } m_b(w,t) \cong w $$
(assume AGI technology monotonically increases with time $t$)
Therefore, if AGI is better than humans (assuming we don't view AGI as a human tool), it must be because AGI has better sensors than humans (infrared, ultraviolet, ultrasound, infrasound, etc.) and thus can perceive a broader world.
[1]W. Gurnee and M. Tegmark, “Language Models Represent Space and Time.” 2023. [Online]. Available: https://doi.org/10.48550/arXiv.2310.02207
[2]Allegory of the Cave, Plato, c. 375 BC