> We can observe that cells/DNA are programmed to adapt to local changes; we have no idea how they make sudden jumps for a more efficient design;

This totally misrepresents evolution. DNA isn’t programmed to adapt to local changes, but rather it’s not programmed to perfectly reproduce without errors. Some errors are useful and get passed on, most errors are either useless or harmful and don’t. And lots of small changes equals a large change — we don’t have to be able to leap a mountain in a single bound to reach the top, either literally or metaphorically, just make a lot of small steps.

The problem with this thinking is that it completely disregards the time complexity/probability of that happening. So having certain "bits flipped" in DNA in a certain sequence in multiple representatives of the same species in the same location for the change to propagate has too low probability, especially when the number of instances of biological species was usually quite low due to constant struggle for survival; the complexity of growing e.g. a new leg because you somehow need it is again unthinkable with our current computing power (we can't even model that properly with reinforcement learning on fastest supercomputers). So as a conclusion, theory of evolution is fitting for small adaptive changes but is either too optimistic or outright wrong for large jumps inside species; we are obviously missing something very important there. Listening to Dawkins I can agree with his local adaptations arguments but then he suddenly makes a jump and says "and this happens all over millions of years" and that somehow makes that happen, which is a position of faith and not science. Better would be if biologists finally admitted we have no clue about mechanism that allows these jumps; references to random sampling are laughable and fitting freshmen at a university and I am saddened this is accepted - we are likely going to miss some great discoveries because somebody will be insisting on this ridiculous hypothesis instead of figuring out the proper algorithm behind it.

> the complexity of growing e.g. a new leg because you somehow need it is again unthinkable with our current computing power (we can't even model that properly with reinforcement learning on fastest supercomputers).

On the contrary, the complexity is so trivial it’s something I did in 30 minutes, on a laptop, in 2008, as a “prove it to myself” response to a young-Earth Baptist Creationist.

What we can’t do is simulate full-lifecycle organ-scale multi-cell protein folding in real time, but that’s a very different thing.

What do you mean? You trivially simulated the process of figuring out that an animal needs to grow a new leg/organ in response to some environmental situation or to gain "an upper leg" against its adversaries? Are you serious?

Genetic algorithms are a very easy to implement form of AI. The hard part is to create a fitness function which does what you actually want — In the wild, reality and competitors are your fitness function, and none needs to be created; conversely, in a simulation the fitness function is necessarily artificial, and can be made to reach any goal you want, from “compose a song” to “write the rules for an AI for this video game so I don’t have to” to “discover a new superconductor” to “design a clock” to “learn to walk”… etc., and all of those are real examples.

Separately, as an algorithm, evolution also has easily characterisable failure modes (local minima), and we do observe those failure modes in organic examples.

But, to emphasise, the fact artificial evolution requires us to write our own fitness functions is merely an implementation detail; these algorithms work because evolution works. They produce novel solutions without the authors of the programs creating the solutions.

Genetic/memetic algorithms are just a rough approximation of a certain selection mechanism we see in the nature, like what Deep Learning is to brain. Making a conclusion that some trivial algorithm with even complicated custom cross-over function can be translated to real-world is as much fantasy as thinking one could upload consciousness to TensorFlow. I never understood how anyone calling themselves a scientist could make such a leap, but obviously it's popular and people made a "scientific" career doing just that...

> Genetic/memetic algorithms are just a rough approximation of a certain selection mechanism we see in the nature

This is irrelevant to my argument. If the simplified version is capable of doing what you have said is unexplained, it is unreasonable to assume that the more complicated version will fail to be capable of the same.

> I never understood how anyone calling themselves a scientist could make such a leap, but obviously it's popular and people made a "scientific" career doing just that...

Occam’s Razor. Start with the simplest possible model, make a prediction, look carefully at reality to see if you were wrong (an act which is easier the simpler the model), and only update the model when reality disagrees with it.

Me coding a genetic algorithm can easily reproduce the entirety of any specific human’s genome if I pick the correct fitness function. Pretty pointless to do so beyond proof of concept, but proof of concept is enough to make the point in this case.

Humans have a history of occasionally choosing bad fitness functions that predates computers, for example look at the awful things we've done to persian cats or chihuahua dogs.

Well yes. Also, the fact that brains (not even just human ones) influence the real world’s organic evolutionary fitness functions does make it somewhat impossible to perfectly replicate in a computer.

Fortunately my point has no dependency on any specific fitness function, natural or programmed — my argument has been that the original claim “We can observe that cells/DNA are programmed to adapt to local changes; we have no idea how they make sudden jumps for a more efficient design;” totally misrepresented evolution, that in fact evolution totally can make sudden-looking jumps (a point which in retrospect I don’t think I properly justified, even if my list of examples was intended that way), and that it’s trivial to demonstrate this by writing a genetic algorithm.