4 ms·

I think people in HN (and in general) have to remember the central axes of evolution before digesting this kind of article. I've already seen comments like it'

I think people in HN (and in general) have to remember the central axes of evolution before digesting this kind of article. I've already seen comments like it's "odd that no species has ended up evolving a continuous maintenance system" and "sleep has a higher opportunity cost in humans" in this comments section. This is a common, but poor framing of evolution. Remember: the only real rule with evolution is "did this evolutionary change mean they survived / reproduced more"? Evolution is not a local or global maxima solving function. I find this kind of attempt at interpreting selection common with engineers, programmers, etc.

yuretz 2mo ago
I can see why it isn't a local or global maxima solving function, but it surely can be modelled by one, right?
- teekert 2mo ago
  I'm a biologist so forgive me the poor wording, but there are "infinitely high" peaks/walls in the evolutionary landscape, as in: "the human urinary system has some evolutionary constraints and anatomical features that are not optimally designed from an engineering perspective, but they are generally functional and reflect our evolutionary history rather than deliberate design." <- this is from AI, I didn't remember the details, but the ureters or urethra take a long way around the inner part of the body because at some point it was a good route, but as we evolved it became suboptimal, but we can't just evolve to untangle all the pipes in our body and do some cable management... We're stuck with what we have (no tunneling electrons ;)). So in that sense, a global maxima (minima?) solving function will probably not arrive at what we have, it has to take into account our evolutionary path, in its entirety (as in the whole tree.), because it is all encoded in 1 (or several) long molecules, and a radical changes would take a rewrite of many parts at the same time, which just does not happen. Maybe at some point we can do it ourselves... though I imagine there being a lot of unintended consequences in other places in the body :). Another example is that sometimes complexity in a cell signalling cascade is simply there to get some timing right (ie delay some response), and the complexity is just there, and won't simplify by itself -> But then later the complexity may start to serve some purpose. Just like proteins that are "copypasta-ed" in the genome, leaving one copy free to mutate and attain a function, or not... etc etc.
  - dtech 2mo ago
    That is the difference between global and local optima. We're not at a local optimum though, our bodies can probably evolve to produce half the urine or something without changing the route, but there is no evolutionary pressure to do so.
    - teekert 2mo ago
      Yes, but we cannot evolve to have our urinary tracts take a more efficient route (or at least the probability is exceedingly low), even though there would be pressure to do so. (Maybe I'm misunderstanding your point)
    - mindslight 2mo ago
      You're implying that "half the urine" describes something that could be a local optimum, but a local optimum would be something more in terms like urine output versus energy expenditure required to achieve that. But that doesn't even capture it, because you can directly control how much urine you produce - water in, water out. So what's really being optimized for is the pressure/stress/disutility function of the kidneys in the context of how much water you are intaking. Which is presumably moderated by your access to water, and higher level conscious optimization like how much water you're wanting to intake and how much urine you're wanting to produce. And that's not even capturing all the constraints in play! But it hopefully illustrates how the scope of what is being optimized for is quite large - it's not each metric in isolation, but rather some (probabilistic) utility function over the whole.
  - albert_e 2mo ago
    Interesting I did not know human urinary tract was suboptimal. Are there examples of "discontinuous" evolution where a chance mutation stumbled upon a much better way to do something that allowed a species to jump the gap? I imagine it would be very rare because chance mutations in a single individual that confer extraordinary survival / mating advantage to offset entire population would be rare like being a superman among men.
    - 2mo ago
- dtech 2mo ago
  Evolution is the ultimate "good enough". Once the required threshold is passed, pressure stops. The threshold might constantly be moving due to competition, environment etc.
- i_love_limes 2mo ago
  Great question! It maybe can be, but what is the maxima that is being solved for exactly? There are usually tradeoffs, the most common and prevalent is our body's immune response. If it's under-active, we might die from an infection. If it's over-active, we get diabetes, lupus, chron's, and (probably) lots of mental health disorders. So, it's a bit context dependent, and definitely dependent on other mutations that have occurred in other parts of your DNA. Most complex diseases are "polygenic", meaning it's a culmination of quite a few factors that would contribute to a specific good or bad outcome. So, yes, it could be modeled as a sort very context dependent with a lot of highly correlated non-independent covariates. We do use quite a lot of statistical and ML methods to understand the genome (I work in statistical genetics), but the complexity of biology has so far proved a tough nut to crack.
  - yuretz 2mo ago
    Great answer! Thank you!
  - codethief 2mo ago
    > It maybe can be, but what is the maxima that is being solved for exactly? Survival rate/reproduction rate/rate of genes being spread?
    - PetitPrince 2mo ago
      I think you're correct, but only if you add "for a given environment". It's good to develop the ability to store super efficiently fat when you're a Pacific Islander and food is scarce, but this advantageous trait becomes a liability with modern and plentiful (junk) food. see thrify gene hypothesis; I know it's not a good explanatory theory but it's a good illustration of my point
      - codethief 2mo ago
        Sure, the form of the function being maximized will usually depend on the environment.
- tempfile 2mo ago
  It can't be a global one because it is inherently local. Evolution is a process which determines the next time step (generation) from only the previous generation. It has no memory at all nor "knowledge" of other reproducing pairs.
hackernud3s 2mo ago
Did their genes survive more, you mean. Hyper-vigilance could be maladaptive for the individual and still survive if it benefits others that carry the gene for it.
- vintermann 2mo ago
  A bit of a digression, but I recently learned about the concept of pangenomes, of trying to map not just the genes in one individual, but the entire set of genes from a clade. It seems it matters a lot in practice for e.g. the immune system that we don't all have the same surface proteins on our cells, even if done some surface proteins are clearly better than others.
FallCheeta7373 2mo ago
To be precise inclusive genetic fitness not idealized bayesian optimality.
Kim_Bruning 2mo ago
> "did this evolutionary change mean they survived / reproduced more" You're just renaming the terms. More/less is a fitness gradient, and the thing that lives there is called an optimizer. Evolution is a tunable-scale optimizer. It's not perfectly local, because it shotguns to avoid local optima on a (bumpy/noisy) fitness landscape. If you tune the 'shot pattern' really wide (infinite copies with infinite variation), you'd genuinely get a one-shot global optimizer at the limit. But that's not efficient nor realistic, so typically it's somewhere in between. Uh, compare maze solvers: you've got your greedy direct route (gets stuck in the first corner) , Dijkstra (finds all the answers but takes ages), and then stuff like A*, which is a happy medium. Evolutionary algos can be used to solve mazes just the same. I think it's a bit heavier on resources than A*, but well suited to embarrassingly parallel optimization. edit: Heh, I actually had claude build a comparative simulation. Turns out evolution is actually very slow on a single optimization axis, and gets stumped in mazes where the detour length is greater than the "temperature"/"spread"/"noise". The actual performance metrics are fascinating. But once tuned, it is able to solve mazes. Thus an optimizer. QED :-P

yuretz 2mo ago

I can see why it isn't a local or global maxima solving function, but it surely can be modelled by one, right?

teekert 2mo ago
I'm a biologist so forgive me the poor wording, but there are "infinitely high" peaks/walls in the evolutionary landscape, as in: "the human urinary system has some evolutionary constraints and anatomical features that are not optimally designed from an engineering perspective, but they are generally functional and reflect our evolutionary history rather than deliberate design." <- this is from AI, I didn't remember the details, but the ureters or urethra take a long way around the inner part of the body because at some point it was a good route, but as we evolved it became suboptimal, but we can't just evolve to untangle all the pipes in our body and do some cable management... We're stuck with what we have (no tunneling electrons ;)). So in that sense, a global maxima (minima?) solving function will probably not arrive at what we have, it has to take into account our evolutionary path, in its entirety (as in the whole tree.), because it is all encoded in 1 (or several) long molecules, and a radical changes would take a rewrite of many parts at the same time, which just does not happen. Maybe at some point we can do it ourselves... though I imagine there being a lot of unintended consequences in other places in the body :). Another example is that sometimes complexity in a cell signalling cascade is simply there to get some timing right (ie delay some response), and the complexity is just there, and won't simplify by itself -> But then later the complexity may start to serve some purpose. Just like proteins that are "copypasta-ed" in the genome, leaving one copy free to mutate and attain a function, or not... etc etc.
- dtech 2mo ago
  That is the difference between global and local optima. We're not at a local optimum though, our bodies can probably evolve to produce half the urine or something without changing the route, but there is no evolutionary pressure to do so.
  - teekert 2mo ago
    Yes, but we cannot evolve to have our urinary tracts take a more efficient route (or at least the probability is exceedingly low), even though there would be pressure to do so. (Maybe I'm misunderstanding your point)
  - mindslight 2mo ago
    You're implying that "half the urine" describes something that could be a local optimum, but a local optimum would be something more in terms like urine output versus energy expenditure required to achieve that. But that doesn't even capture it, because you can directly control how much urine you produce - water in, water out. So what's really being optimized for is the pressure/stress/disutility function of the kidneys in the context of how much water you are intaking. Which is presumably moderated by your access to water, and higher level conscious optimization like how much water you're wanting to intake and how much urine you're wanting to produce. And that's not even capturing all the constraints in play! But it hopefully illustrates how the scope of what is being optimized for is quite large - it's not each metric in isolation, but rather some (probabilistic) utility function over the whole.
- albert_e 2mo ago
  Interesting I did not know human urinary tract was suboptimal. Are there examples of "discontinuous" evolution where a chance mutation stumbled upon a much better way to do something that allowed a species to jump the gap? I imagine it would be very rare because chance mutations in a single individual that confer extraordinary survival / mating advantage to offset entire population would be rare like being a superman among men.
  - 2mo ago
dtech 2mo ago
Evolution is the ultimate "good enough". Once the required threshold is passed, pressure stops. The threshold might constantly be moving due to competition, environment etc.
i_love_limes 2mo ago
Great question! It maybe can be, but what is the maxima that is being solved for exactly? There are usually tradeoffs, the most common and prevalent is our body's immune response. If it's under-active, we might die from an infection. If it's over-active, we get diabetes, lupus, chron's, and (probably) lots of mental health disorders. So, it's a bit context dependent, and definitely dependent on other mutations that have occurred in other parts of your DNA. Most complex diseases are "polygenic", meaning it's a culmination of quite a few factors that would contribute to a specific good or bad outcome. So, yes, it could be modeled as a sort very context dependent with a lot of highly correlated non-independent covariates. We do use quite a lot of statistical and ML methods to understand the genome (I work in statistical genetics), but the complexity of biology has so far proved a tough nut to crack.
- yuretz 2mo ago
  Great answer! Thank you!
- codethief 2mo ago
  > It maybe can be, but what is the maxima that is being solved for exactly? Survival rate/reproduction rate/rate of genes being spread?
  - PetitPrince 2mo ago
    I think you're correct, but only if you add "for a given environment". It's good to develop the ability to store super efficiently fat when you're a Pacific Islander and food is scarce, but this advantageous trait becomes a liability with modern and plentiful (junk) food. see thrify gene hypothesis; I know it's not a good explanatory theory but it's a good illustration of my point
    - codethief 2mo ago
      Sure, the form of the function being maximized will usually depend on the environment.
tempfile 2mo ago
It can't be a global one because it is inherently local. Evolution is a process which determines the next time step (generation) from only the previous generation. It has no memory at all nor "knowledge" of other reproducing pairs.

hackernud3s 2mo ago

Did their genes survive more, you mean. Hyper-vigilance could be maladaptive for the individual and still survive if it benefits others that carry the gene for it.

vintermann 2mo ago
A bit of a digression, but I recently learned about the concept of pangenomes, of trying to map not just the genes in one individual, but the entire set of genes from a clade. It seems it matters a lot in practice for e.g. the immune system that we don't all have the same surface proteins on our cells, even if done some surface proteins are clearly better than others.

FallCheeta7373 2mo ago

To be precise inclusive genetic fitness not idealized bayesian optimality.

Kim_Bruning 2mo ago

> "did this evolutionary change mean they survived / reproduced more" You're just renaming the terms. More/less is a fitness gradient, and the thing that lives there is called an optimizer. Evolution is a tunable-scale optimizer. It's not perfectly local, because it shotguns to avoid local optima on a (bumpy/noisy) fitness landscape. If you tune the 'shot pattern' really wide (infinite copies with infinite variation), you'd genuinely get a one-shot global optimizer at the limit. But that's not efficient nor realistic, so typically it's somewhere in between. Uh, compare maze solvers: you've got your greedy direct route (gets stuck in the first corner) , Dijkstra (finds all the answers but takes ages), and then stuff like A*, which is a happy medium. Evolutionary algos can be used to solve mazes just the same. I think it's a bit heavier on resources than A*, but well suited to embarrassingly parallel optimization. edit: Heh, I actually had claude build a comparative simulation. Turns out evolution is actually very slow on a single optimization axis, and gets stumped in mazes where the detour length is greater than the "temperature"/"spread"/"noise". The actual performance metrics are fascinating. But once tuned, it is able to solve mazes. Thus an optimizer. QED :-P