uplifter

There is a distribution of optimism, some people in 2023 were predicting AGI by 2025.

No such thing as trajectory when it comes to mass behavior because it can turn on a dime if people find reason to. Thats what makes civilization so fun.

Has someone turned this into an agent benchmark? Most tokens emitted until the system rm -rf /s

This may be more a problem with how "fit" is defined and used than with natural selection theory itself. Fitness can be hard to define beyond the trivial "these organisms which survived the selection event must be the fit ones," and natural systems are usually so noisy with inputs that its hard to figure out what was actually important in retrospect, or likely to evolve in the future.

Only in situations with a powerful selection pressure (like an asteroid strike causing a nuclear winter, or antibiotic applied to a petri dish) can one have a hope of reliably predicting which variants will be selected for or against.

However, these situations are not irrelevant, especially if we can predict the likelihood of those situations developing. Real predictions of the theory of natural selection can be applied to managing antibiotic resistance in populations of bacteria. For example, we know that antibiotic resistance mechanisms that bacteria evolve will often have an energy metabolism cost to their maintenance. This means that, absent pressure to be resistant to antibiotics, we'd expect a population to gradually lose individuals with the genes for the resistance mechanism, because they would be incurring a metabolic penalty for possessing those genes. So natural selection theory accurately predicts that if you remove the selection pressure of the antibiotic, the bacteria will evolve to lose the resistance mechanism, and become susceptible to the antibiotic again over several generations of natural selection. Using this knowledge, some rural regions will discontinue use of a given class of antibiotics in agriculture to allow for resistant strains to decline, and then resume their use when they are again effective. By intelligently rotating use of antibiotics in this way, we can enjoy their benefits without incurring too much inefficiencies and worse tragedies from antibiotic resistance.

That is real & useful predictive power.

> What is the algorithm implementation when it comes to the physical world?

It is the physical world, nature is the implementation of the natural selection algorithm. Yes, the strong force is part of the implementation, because the strong force can play a role in selection events, cf nuclear bombs and radiation. The gravitational pull of remote galaxies can also influence selection events by changing planetary orbits minutely.

I don’t see these as problems for my argument because I am not the one claiming they encode an objective, I just see them as natural forces which can influence selection without any overarching purpose or goal. It is those claiming natural selection is an optimization process who must show how it could work. The onus is on them to show where their supposed objective of natural selection is encoded in its implementation.

> If you create selection pressure, and have a reproductive system that allows for mutations, then you may end up an "implementation" that encodes the goal implicit in the selection pressure.

What goal are you referring to?

> But anyone who messed around with genetic algorithms or artificial life in the 90s knows that you can trivially start out with no resemblance to "the goal" at all. Where life on earth in aggregate or any specific example of it in particular might be along that pathway is similarly impossible to say.

I am one of these people, but I don’t know what goal you are saying these systems came to demonstrate. Are you saying these artificial evolution systems had objectives they pursued? What caused them to follow these objectives? What is this “pathway”?

>Consider the well-documented case of moth evolution in industrial (and later, post-industrial) northern England. Their camouflaging wing tones changed to respond to the typical color on vertical surfaces, twice within a human generation or three. Was "the goal" flexible coloration across generations, or was it "light, then "dark" and then "light" again?

There was no goal at any point in the process. Moths with colors that matched their contemporaneous environment were less likely to be eaten by predators than those which stood out. Calling it a goal is a confusion, its trying to add a conceptual framing that isn’t necessary and adds nothing to the understanding of the system. Neither the soot levels in the air nor the birds hunting for moths have a goal of adjusting the balance of moth coloration phenotypes. They are just the context, along with everything else in their environment, in which evolution of moth coloration may occur.

In what sense is there any goal in the example? And if it is a goal, why is it not optimization? I claim there is no goal, no optimization objective to natural selection. Its not just a philosophical side question, it is the question.

edit: typo

What part of it don’t you agree with? That an algorithm implementation must encode the goal that it pursues? How can something pursue a goal it has no access to a definition of? If you have an alternative way it could work, please propose it.

I’m not asking rhetorically, I’m truly interested in learning the flaws in my argument for why natural selection cannot be modelled as an optimization process. So if you have the time to reply with a more detailed rebuttal, I’d much appreciate it.

edit: Addendum: I recognize my claim that computer scientists might have an advantage in understanding this is contentious, and I was not implying that they (we) as a group have a better record of understanding evolution’s subtlety than biologists (which I studied in uni) or the average lay person. I just think they could have an advantage in understanding the version of the argument that I gave above, and I am interested in improving it for that purpose.

That's fine, all I'm saying is that if genes don't actually have intention, then the utility of modeling them as though they do must be strictly limited, if not an outright liability in some contexts. Use the heuristic at your own risk, but don't sell it as gospel truth.

This relates to why you will often see multiple mushrooms of the same type blooming at the same time in a ring pattern: the edge of the ring is the periphery of the linearly, radially expanding mat of subterranean fungal fiber weave, which produces fruiting bodies at its edges.

>Thus, one may guess, that first distinctive poisonous mushrooms like the fly agaric developed, then most animals large enough to eat them developed an instinct to avoid all mushrooms, and then the non-poisonous freeloading mushrooms developed.

Just wanted to note that these phenomena are important enough in the study of mimicry in biology to have earned their own names:

Müllerian mimicry is when two species who are similarly well defended (foul tasting, toxic or otherwise noxious to eat) converge in appearance to mimic each other's honest warning signals.

Batesian mimicry is when a harmless or palatable species evolves to mimic a harmful, toxic, or otherwise defended species.

In The Selfish Gene, Dawkins emphasized that the primary unit of evolution was the individual gene, not whole genomes. The genes were replicators and the genomes were just collections of replicators, and the way the selection pressure math worked out, there was too much diffusion of responsibility for whole genomes that typically evolution could not work coherently at that scale, or at least that's my best recollection of the book's main theory.

Regarding intentionality being a good practical assumption, I actually don't recall Dawkins recommending that, and it seems doubtful because that can lead to all kinds of fallacious reasoning. I mostly considered Dawkins a data-based neo-darwininian, so it would surprise me that he would recommend that.

Could you recall a quote or chapter from the book that bolsters your point?

edit: typo