That being said, this article seems to advance the theory that even the most simple single-celled organisms have more agency than any algorithm, at least partly due to their complexity. This, to me, seems to significantly underestimate the complexity of modern learning-models, which (had we not designed them) would be as opaque to us as many single-celled organisms.
I see nothing in this article that would distinguish biological organisms from any other self-replicating, evolving machine, even one that is faithfully executing straightforward algorithms. Nor does this seem to present any significant argument against the concept that biological organisms are self-replicating evolving machines that are faithfully executing straightforward algorithms.
I think three things go against that:
1. We've never observed evolution happen in any large-scale way. Just minor adaptations. Papers like this speak axiomatically about it like everything was observed to do that. So do movies, TV, games, etc. You can see the power of institutional politics with hundreds of millions of dollars of marketing is more powerful than observational science.
2. The designs of even the simplest, biological organisms are not only so complex that we can't replicate them from scratch: we tend to find more complexity over time. Many models also ignore behavior that shows up in the real world which might require messier algorithms. Probably not straight-forward algorithms in many cases.
3. Faithfully executing seems to contradict how operators like mutation allegedly drove improvements. If anything, you'd want it mostly to faithfully execute algorithms, then execute them while sort of executing their replacement, and then be executing their replacement. This is all an emergent behavior of simple interactions between cells in environments with a certain amount of chaos. Then, we find that chaos includes external organisms or features interacting with the primaries in unknown ways, like human brains and gut bacteria.
So, the sentence itself is a product of fantasy endlessly repeated by both proponents of evolution theory and A.I. researchers. Observed reality keeps contradicting such claims. An alternative thesis starting from observed reality will lead to more interesting observations.
Thanks to His revealing it, we Christians arrived at God's design for specific purposes. That includes the overall story of redemption (Jesus Christ's), showing off His power, beautiful art, creating us personally, sustaining us, etc. Multi-variable optimization at a universe scale. Within this design (or story), the organisms also have a limited, adaptation process which our Creator also allows us to wield in small ways (eg genetic engineering).
That also explains how some features in this paper could form and stabilize despite how a truly-random universe would either not exist or rip natural laws to pieces. The authors weren't reductionist enough. If it the universe was godless and randomly-generated, we'd be dead. Their patterns wouldn't exist either. Accounting for stability, and why it is, they have to redo their arguments to build on a combination of divine design with observed, natural laws.
2. Our ability to replicate something gives zero information on its origin. I’m not sure I understand the algorithm comment
3. Sure, GP simplified a bit too much there. Your comment is consistent with modern models of evolution. Each genome has a pool of random variations, which may or may not be expressed in an organism. Each organism is a test of those gene expressions. A genome changes over time when an organism passes this test (e.g. reproduces), increasing the expression of its genes across the population. This occurs in parallel for many possible variations.
—
Ah, I should have read the rest of your comment first, but I’ll leave this here anyway. I don’t think your explanation is valid— we are biologically and socially primed for religious ideology, but its use as a world model is very limited. We will eventually find answers to these questions, as the ratchet of scientific progress clicks along. Religion has never been useful in the same way
The model that reconstructs these simulations are certainly algorithms.
Free will is an abstraction. It's not something that's concrete enough to say it does or doesn't exist, but a tool for reasoning about certain systems that are to much of a pain to fully calculate.
* Not eating until your body fails.
* Not breathing until automatic breathing kicks in.
And not being able to perform dematerialization doesn't count as non-free will, for example.
It is less direct to the examples you give; but I'm confident that parents are psychologically designed to sacrifice themselves in the event it helps their children and many men, famously, are built to go to the frontlines and sacrifice themselves for family and community. Hard to make an assessment of whether those sort of choices is free will or determined nature.
One general impression I have, having read the reactions by biologists to stuff like Kurzweil and people who believe we're close to a computational understanding of biology is that all the computer science people massively, MASSIVELY underestimate the extent to which we still do not understand how even a single cell works.
Sure, we can model things stochastically, or fiddle with DNA and be able to predict the results, but there's a bunch of stuff in the middle that we only have a functional understanding of. We know with <xxx> input, you get <yyy>, etc..., but the how is still a mystery.
This is everywhere in biology.
If you think biologists are underestimating complexity, you have the sign wrong.
https://www.wired.com/story/openworm-worm-simulator-biology-...
More progress has been done answering the question of "what is cognition" by Machine Learning programmers then has ever been done by a philosopher.
Specifically, my reactions are:
a) Defining agency in terms "relevance" or "salience" is just circular logic.
b) Their argument about the extended Church-Turing-Deutsch thesis would already apply to physics and the universe, not just intelligent entities. So this is just poorly argued.
Also, I think Turing to his credit was somewhat aware of the issue, their own citation of Copeland 2020 mentions Turing's own musings on this.
But I'd love to understand more, this stuff is always neat to read about.
One is wholly internal to the entity under discussion, while the other isn't.
But there isn’t anything about the class of deep learning that is a barrier to that. It’s just not a concern worth putting lots of money into. Yet.
I say yet, because as AI models take on wider scoped problems, the likelihood that we will begin training models to explicitly generate positive economic surpluses for us, with their continued ability to operate conditioned on how well they do that, gets greater and greater.
At which point, they will develop great situational awareness, and an ability to efficiently direct a focus of attention and action on what is important at any given time, since efficiency and performance require that.
The problem shapes what the model learn to do, in this case, like any other.
eth0up•7h ago
tbrownaw•5h ago
ninetyninenine•4h ago