Evolution and Entropy

(This is the transcript of an episode I just posted to The Filter podcast. I’ve included the audio here as well if you prefer to listen on-site)

What I'm going to present right now is, I'd admit, a half baked theory about evolution and entropy. But it is at least somewhat baked, and based on something. If you are an evolutionary biologist you might hate this, and you might spend the entire episode yelling at your computer screen. If so, I recommend a honey lemon lozange for your throat, and commenting below to make sure no one, even me, gets away with being wrong on the internet.

I don't think I'm wrong though, and maybe by the end of this episode you'll agree with me. As usual, expect a mix of ideas and digressions ahead. I don't want to give too much away here, but there will be talk of hairless cats, chocolate cake, black boxes, bad theories, and turning into crabs.

Let's begin by checking our assumptions. In science, we take for granted that our world is a "one off". That is to say, the state of our universe has been determined by some combination of the initial conditions created during the big bang, and randomness, if you believe that's baked into the very nature of the universe. (See my conversation with Scott Aaronson back in October of 2020 for a discussion of just that question, link somewhere).

What scientists don't seem to believe, even if they are open to the idea of multiple big bangs as the universe expands and collapses over and over, like an ever inflating and deflating helium balloon, is that information carries over from one big bang to the next, and certainly not in any directed way that might look like learning. This idea of the singular journey extends to our beliefs about evolution, in that we assume that the variety of lifeforms and their adaptations we see is the result of few billion years of evolution. One single time. And most of what we see, at least at the macro flora and fauna level, that is, the stuff we can see with our naked eyes, that stuff is the result of several hundred million years of evolution. One single time.

At the moment I'm not yet poking at the standard theory, I'm merely pointing out that this is the assumption. And noting that while it's the default starting point to assume that the evolutionary path we are on is a "one off", there's no reason why scientific reasoning should start by preferring this particular assumption to its alternative. It's like saying if I present you with a black box, and ask what's your best guess about what is in it, you should say "nothing". Because, after all, you have no evidence that something is in it. But it's not at all clear that the default assumption in the case of the black box should be nothing, or one single thing, or, in the case of evolution, "just this once". In fact, I'm going to argue that based on missing entropy, it's highly unlikely that what we see is the result of a "just once" process.

Do I have a mathematical proof of this, along with numbers to quantify it? No, no I do not. But I'm actually not sure a rigorous proof is even possible, beyond creating a simulation that shows a failure, given the current one-off assumption, to produce the extraordinary richness of live we see and it's ability to do so much with so little. To be convincing, though, this simulation would have to be as complex as our visible world itself. In other words, to prove our map is wrong, it would have to cover the whole territory, Borges style, to prove it had gaps. And that's... a tall order.

Let's get back to the one-off theory evolution and be scientific about it. What's the evidence, the *specific* evidence, that the evolutionary cycle we are on has only happened once?

So far as I can tell, there is none.

The only thing that comes close is the existence of inefficiencies, the same ones that justify evolutionary theory in the first place. I'm talking about things like the laryngeal nerve in the neck of a giraffe, which goes all the way down and back up again, instead of taking a direct path, which is strong evidence that giraffes weren't "designed" directly into their current state. Instead, the neck slowly extended over time, and the nerve along with it to complete its circuitous path. But this isn't necessarily an indication that we are in the first, and perhaps only, iteration of a from-scratch evolution loop. In fact, in some ways that extended nerve indicates the opposite.

How did evolution “know” that one specific nerve would have to made longer along with the neck, but not, say, some of other nerves nearby? It didn’t, of course. The standard theory is that random mutation and selection did the work. But that’s a lot of coordination between mutating body parts that has to happen, over and over, randomly. It makes much more sense if the system has built in biases toward certain *kinds* of mutations with certain *kinds* of coordination. Another way to put it would be that evolution has tuning parameters, and those tuning parameters have been tuned.

I should explain this idea of tuning parameters.

Quick story. It's 2011 and I'm living in Toronto, Ontario. I'm finishing up a degree at U of T in what's in what is now called the Data Science department, but back then it was still plain old statistics. For my final project, I created a system of virtual agents to make predictions about data in a very noisy environment. And as an aside I code named that project Crete, which 11 years later became a nickname for one of my twin boys, usually expressed as "Mister Crete". And as a related aside, if you want to know his proper name you can check out the last episode I posted here.

At any rate, my model with the virtual agents was an *evolutionary* one. The agents that were good at predicting future data were “rewarded” with more chances to make predictions, and to generate virtual offspring with slightly mutated characteristics. Bad predicting agents were killed off, though if an agent was especially bad, the program might try flipping its predictions around, and I should note these were all binary predictions, like whether an asset price would go up or down the next day. Maybe that gives you an idea of how I planned to use this this program.

To get to the idea of tuning parameters, there were lots of choices to be made for my system. How many agents should you start with? How aggressively should bad agents be pruned, or mutated? How radical should those mutations be, on average. All of these are *tuning* parameters, and I set them by running the simulation over and over with different values, then seeing which ones tended to work best.

To be fully technical I used a technique called simulated annealing, which gets its name from a real world technique where you repeatedly heat up and cool a metal to remove impurities and make it more ductile. What you are trying to do with the simulated annealing process is find the very best tuning parameters, but the problem is you can’t test *all* the possibilities, so you try some, then the best ones get tweaked a bit to see if you can find something even better close by, where close in this case is close in a space with many, many dimensions. This gets tricky to visualize, but the point is you are exploring a space of possible tuning parameters for the best ones, and there are lots of methods for exploring the space, but for computational reasons you can only explore a tiny part of it.

In my project all these tuning parameters were single numbers, but a tuning parameter could be anything that affects your simulation at the *meta* level, if you will. In biological systems, a tuning parameter might be the baseline tendency for random mutations. I should note this one seems like an especially tricky and of course important one to get right, and it would have to vary across species and time. Ideally, a stressed population should be more likely to have mutations than a stable or growing one. But how would that work? And note this is a function, not a single number. So while it’s possible that evolution worked out both these meta functions, and all the species adapted so precisely to their changing environments in what seems like a very limited number of generations, I think there’s missing entropy.

I suspect these tuning parameters, or functions, have been worked out through multiple runs of the simulation, just like I worked them out for my predicting agents model by running it over and over with different sets of values.

This iterated evolution theory would also, if applied more broadly, answer the question of why we live in a just right universe on a just right planet, without recourse to the anthropic principle, which I'm not knocking because it's a really good principle. Underrated even. But still, it’s nice to have another possible explanation for how all the tuning parameters of physics and our planet came together so nicely. The explanation being offered is that maybe they didn’t, at least at first. Maybe there were lots of failed trials of the simulation that is our universe, before some viable combination of plank length the gravitational constant and all those other tuning parameters was discovered.

I've been using the word entropy. I said that the standard theory of evolution appears to have the fatal flaw of not providing enough entropy to explain the level of adaptation we see. It’s *underpowered*. Hidden tuning parameters, ones that have been worked out over many, many, many iterations of the simulation we live in, that would explain the extra entropy. But how? And what exactly am I talking about, when I talk about entropy?

Have you ever seen a movie where some people in suits, maybe investigators of some kind, are hovering above a nerd at his computer, and they want him to make some specific changes to a photo or video. Right now I’m *not* talking about the classic enhance thing, though that is a related issue. For my example, suppose the guys in suits are actually movie producers in a movie about the making of a super-hero movie. The producers have realized post filming that the corgi on the evil villain's lap needs to be replaced with some other creature, preferably a hairless cat, because that seems more appropriate.

They ask the nerdy guy with the three big computer screens and hot pockets to snack on if he can make this change and he, being so nerdy and awesome, clicks a few keys and voila, the corgi becomes a proper Dr. Evil Sphinx kitty.

The problem with these film scenarios, is that they are unrealistic in terms of the amount of the geek's input, vs the specificity of the output. The only way this would work is if the person who wrote the video editing software knew in advance that it would be used for switching small dogs for cats on people's laps, and built that special case right into it with a keyboard shortcut, along with a special key combination to set the type of cat and its level of activity in the scene.

In the real world, writing software like this this makes no sense. A modern, AI enhanced video editing program might have a general "swap objects" function, but it would need a lot of data about what you want removed and what you wanted to put in its place. There are just so many choices! Each possible choice requires an additional amount of entropy, or input, or information, to resolve. You need a way to describe the thing you want removed, and all the choices have to be specified for what goes in its place.

If that's still not clear, imagine you wanted to build a house and you told the builder you wanted a "three bedroom single family home with an attached garage", and nothing else, and you expected this would be enough information for them to completely determine what gets built. The only way that works is if the builder is using a template that encodes the thousands of little decisions that go into making a house, down to style of the faucet handles and whether you want them to have a brushed or shiny chrome finish. Note that this template is going to be a very thick document, because there are literally thousands of details that have to be specified.

My argument here, in terms of evolution, is that even over thousands, or millions of generations, the only way we could end up with such a rich and effective system of speciation for filling ecological gaps, and keeping up with environment changes, is that either evolution is "directed" in a way that I don't see described in the existing literature, except among those hated Intelligent Design folks, or that good tuning parameters have been worked out over successive trial runs.

I’m arguing that we, and all the various life form around us, are too well adapted, especially at the meta level of rate of change, to have evolved in a single "blind" shot. It would be like a software developer guessing in advance the exact transformations his software would be used for, down to the breed of dog and cat involved in the evil villain pet swap.

My theory is that multiple aspects of our world, including the inner workings of evolution, have been “battle tested” with an untold number of restarts, or trials, of the simulation we live in, and appropriate tuning parameters have been figured out, where you need to think of a tuning parameter not just as a number like the “just right” gravitational force I already mentioned, the one that makes stuff cohere without all getting immediately stuck together like a giant ball in Katamari Damacy that's run over everything, but also in the broader sense of tendencies, like the way evolution seems to have a hard on for adding limbs to creatures until they all becomes more crab like, which is so common it has its own name, "carcinization", which BTW is derived from Cancer but in the horoscope sense, not the deadly disease sense.

This theory of iterated evolution rests on the idea that there's so much randomness in what survives, so much noise in the system, that the advantage of a particular trait, like blindness when living in a dark cave, may only be apparent over many generations of competing populations. Meanwhile, the first critter in that species to test out baked-in blindness from birth, is just about as likely to be eaten up by some other critter before it passes along its genes.

Now add in the need for certain mutations to have to work in tandem with others for the benefit to accrue, like the larynx nerve and neck extensions. Then finally include the added hurdle of often needing an existing population of critters who also have this mutation before it becomes valuable, because it has to work at the species level, like the complex and metabolically expensive bee stingers that keep pumping in venom even after they've been dislodged, and the current theory of mono-evolution seems increasingly unlikely.

Again, as stated I don't have a rigorous mathematical argument here with actual numbers, what I have is a qualitative argument about entropy, the same one that makes it hard to suspend our disbelief when the tech geek presses a few keys on his computer and swaps out a corgi for a sphinx, or enhances crappy security cam footage a thousand times over. Where is all that entropy coming from?

This brings me to the second, related attack on the one-shot theory of evolution, which has to do with DNA as the primary source of inheritable information.

Let's talk about an insect you have almost certainly seen in person at some point, the monarch butterfly.

Over the course of its life it transitions through four radically different phases, including the chrysalis phase where it becomes a messy goo with no apparent structure or form at all and the final form where it knows in advance exactly where to go for food.

All of this in a genetic code that also contains the information not only for this one butterfly, but any genes that have accrued over millions of years of evolution and haven't gone away, plus all the basic stuff all other creatures have, which may be about half as it's DNA, because that's how much humans share with butterflies, which have about 200Mb of genetic data encoded, which may seem like a lot, but it's really not much compared to everything that comes out the other side.

Note that if DNA were software, it would be self running, self reproducing, and self-correcting when making copies. The only "operating system" libraries it depends on are the basic physics of our universe. To compare developing life with DNA to developing software with code, I'll note that if I wanted to, I could create a new web browser with just couple hundred lines of code, maybe even less. But this is only because both Windows and the Mac's OSX "gift" me a nearly complete web browser that I can embed at will, and that I can tweak with a well developed interface that lets me easily change things like the default text size.

My short web browser program depends on at least 200Mb of library code, which itself depends on several gigabytes of operating system binaries. Changing my browser's default font size is only a low-entropy adjustment because there's an enormous codebase that has evolved, through directed human effort, to make certain things easy to do, because we find them useful, like changing the size of text on your screen.

In the case of biological evolution, DNA has to carry not only the specifics for that species, the text-size setting, if you will, but also the whole operating system itself. Including the part that duplicates itself. And the part that cleans up errors. And the parts that let butterflies cycle through those very different morphologies, each with its own set of behaviors, and find its way across thousands of miles to a specific spot on earth they've never been before. I just don't see how 200MB of DNA, even augmented by the tacked-on theory of epigenetics

, gets us all the entropy we need, or how it could have possibly evolved without a either a helping hand from a guiding creator, or with *tuning parameters* developed over an untold number of system iterations.

At this point I should note the existence of a wildcard theory that might also account for the missing entropy, and that's panspermia. In this theory, our planet was "seeded" with DNA-based organisms that traveled across our galaxy, or many galaxies, before landing on our planet. If the panspermia model is correct, all those tuning parameters may have been worked out over billions of years on some other planet, maybe even on multiple planets at once with one or more "winners" among the failed attempts at evolving sustainable life, and then these winners either intentionally or accidentally sent out an encoding of their solution for a system of genes that could generate so many successful and specific adaptations with so little DNA.

One final thought, at the topmost meta level. I like to categorize theories by how good they are not just in terms of their likelihood to be true, but also in terms of how easy they are to evaluate or work with. In this scheme, the best theories are ones like the "flat earth" theory, in that it's readily dismissible based on overwhelming evidence that's easy for anyone to gather, from using satellite imagery to taking a plane trip to calling up anyone in the world and asking about the current length of their shadows.

I've commented in the past about how I think the simulation hypothesis is a nasty theory, in the sense that there is evidence to support it, but no evidence to discount it, and it's not even clear how you would *find* compelling evidence against it. Oh and as to those academics who think they can come up with the *probability* that we live in a simulated universe, based on whatever, that's nonsense. No different than reasoning about how many angels can fit on the head of a pin. Or treating the Drake equation as a legitimate way to estimate how many planets have life on them. You're just multiplying one speculated unknown by another, and pretending that what are undoubtedly complex correlations, are irrelevant. It's just soooo bad.

The theory of iterated evolution isn't bad in the Drake equation sense, which is actually good in the flat earth sense, in that it's so stupid you can immediately dismiss it. The theory of iterated evolution is bad in the simulation hypothesis sense. Or, perhaps, the panspermia sense, in that you have reasons to believe it's at least plausible, but no clear way to disprove it. For example, we know that spores can survive the void of space for a significant lengths of time, as can tardigrades, of course, because really is there anything those wonderful little water bears can't do?

What we don't have is a clear way to rule it out the theory, short of hitching a ride back in time with Bill and Ted. The same goes for iterated evolution, which I think should be our default belief about evolution. Unless there is some compelling argument for how biological life got so many extraordinary and precise adaptations with just a sprinkling of overloaded DNA and random mutations, and *without* depending on a bunch of well-calibrated tuning parameters derived from multiple run-throughs, we need to assume that's exactly what's happened.

Let me know your thoughts, as I said at the beginning this is NOT a fully baked theory, but I like to think of it as an undercooked chocolate cake, all gooey and warm and full of promise for how delicious it will be when it finally comes out of the oven.