I don't see gradient descent or back propagation in biology.
As for back propagation, I'm sure there are aspects in nature in which translate - maybe the nerves system of some species.
Way to see it is that evolutionary change of any form could in theory be modeled in machine learning. There may be aspects of evolution and how it operates that help refine how we do machine learning.
But regarding evolution implementing back-propagation, I agree that is just bonkers, as it would require cells to back-propagate in time information about death or procreation! If our biology implements information time machines, biology would be a lot weirder than it is (and it would imply that we could use that by sequencing an organisms genome (or ensemble of them) to read from the future, and keeping it fed without allowing reproduction, and later when we want to write information back either kill or let it reproduce).
That's just plain nonsense. There can't be back-propagation in evolution without biological time-travel of information.
And machine learning without back-propagation is basically a genetic algorithm, so the claim that "evolution has done a lot of work that we equate as machine learning" would then reduce to a widely accepted fact that nobody disputes and is thus uninformative: "evolution has done a lot of work that we equate as genetic algorithms"
To equate this into evolution, if you viewed a species as the model, each generation adapts to the previous generations experience, hence it feeds back into the next generation. That evolution may prevail or that whole lineage dies off. It gets down to the perspective of how you measure that time aspect and define the start point of your time perspective. In a simulation you are arbitrary resetting that time baseline on each model, so the equivalent would be to view the human as a baseline of each birth. yes it gets confusing, but to dismiss it just because you have no actual time-travel from a cosmological perspective and compare to an arbitrary time perspective that runs in a linear temporal realm of the universe is unfair.
But many things are hard to equate, you could say it is like comparing apples to pears, but then there are many perspectives in which you can compare them and find them equal. But you can't take one single perspective and transpose that upon another perspective and presume a conclusion.
if natural selection were to use back-propagation, then killing or allowing procreation would be measurable not only in the genome right after birth of an IVF organism, but even be measurable in the half genomes of the sperm and egg cells before they met (assuming we reach some non-destructe genome sequencing technology). That would simply make actual natural selection useless. Moreover, if all cells had this capacity, and our neurons are cells too, why don't I have this skill inately?
I stand by my opinion that natural selection does not employ timetravel, and hence also does not employ backpropagation
EDIT: adding, one does not need timetravel for backpropagation in the brain or in a computer, just enough memory or delay bandwidth to store the intermediate values for backpropagation. i.e. the feedback is nearly instantaneous, or low bandwidth from example to scoring, but scoring in natural selection depends on how much offspring one makes before one dies, how can this inform the choice mutations??
Also interesting that 11 is also a prime number. But equally, may just be one of life's coincidences and the human drive to see patterns and understand exactly what is going on - a trait that still advances humanity today.