WEBVTT 00:00:00.000 --> 00:00:03.060 Welcome to the deep dive. Today, we are looking 00:00:03.060 --> 00:00:06.379 at this stack of research that basically proves 00:00:06.379 --> 00:00:09.480 your computer's most powerful tool isn't its 00:00:09.480 --> 00:00:12.980 flawless logic. It's actually, well, its ability 00:00:12.980 --> 00:00:15.919 to make random terrible mistakes. Right, which 00:00:15.919 --> 00:00:18.079 completely flips the script on how we usually 00:00:18.079 --> 00:00:20.760 think about technology. I mean, we expect computers 00:00:20.760 --> 00:00:23.219 to be these rigid, you know, highly engineered 00:00:23.219 --> 00:00:25.980 machines that just follow explicit instructions 00:00:25.980 --> 00:00:28.339 to spit out one perfect answer. Yeah, exactly. 00:00:29.039 --> 00:00:31.500 But the incredibly detailed Wikipedia deep dive 00:00:31.500 --> 00:00:33.780 we were exploring today shows us a totally different 00:00:33.780 --> 00:00:36.420 paradigm. We are diving into evolutionary computation. 00:00:37.060 --> 00:00:39.539 This is a field where computer scientists literally 00:00:39.539 --> 00:00:42.159 stole Mother Nature's playbook. They really did. 00:00:42.579 --> 00:00:45.789 They're using Darwinian evolution. you know, 00:00:46.049 --> 00:00:47.909 survival of the fittest to solve the world's 00:00:47.909 --> 00:00:50.990 most complex computational problems. So our mission 00:00:50.990 --> 00:00:52.770 today is to figure out how they're actually doing 00:00:52.770 --> 00:00:55.289 this, why making random mistakes is somehow the 00:00:55.289 --> 00:00:57.609 secret to hyperintelligence, and ultimately what 00:00:57.609 --> 00:00:59.969 this means for you and the technology you use 00:00:59.969 --> 00:01:02.479 every single day. Because the underlying philosophy 00:01:02.479 --> 00:01:05.980 here, it turns a rigid calculator into a trial 00:01:05.980 --> 00:01:08.260 and error problem solver. It's literally the 00:01:08.260 --> 00:01:11.159 process of artificial natural selection. The 00:01:11.159 --> 00:01:14.379 machine isn't calculating a single answer through 00:01:14.379 --> 00:01:18.640 a fixed formula. It's growing an answer organically 00:01:18.640 --> 00:01:21.319 over time. OK, let's unpack this. Because the 00:01:21.319 --> 00:01:23.379 mechanics of how this actually works in silicon 00:01:23.379 --> 00:01:25.959 are just wild to me. If you look at how this 00:01:25.959 --> 00:01:29.180 is modeled, there are three foundational forces 00:01:29.180 --> 00:01:31.859 at play. You've got recombination, mutation, 00:01:32.060 --> 00:01:34.340 and selection. Yeah, think of those three forces 00:01:34.340 --> 00:01:37.719 as the engine of evolution. So recombination 00:01:37.719 --> 00:01:39.680 is taking traits from two successful parents 00:01:39.680 --> 00:01:42.640 and just mixing them together. Then mutation 00:01:42.640 --> 00:01:45.480 introduces a totally random, unpredictable change 00:01:45.480 --> 00:01:48.700 to the data. Just a mistake, basically. And selection 00:01:48.700 --> 00:01:50.920 is the environment ruthlessly killing off the 00:01:50.920 --> 00:01:53.260 weak cord, allowing only the highest quality 00:01:53.260 --> 00:01:55.989 solutions to survive to the next round. To visualize 00:01:55.989 --> 00:01:59.129 this, there is this classic animation used to 00:01:59.129 --> 00:02:01.810 demonstrate the concept. The computer is asked 00:02:01.810 --> 00:02:04.549 to evolve an image of Charles Darwin's face. 00:02:05.129 --> 00:02:07.010 And the fascinating part is that it doesn't start 00:02:07.010 --> 00:02:09.449 with a sketch. No, not at all. It starts with 00:02:09.449 --> 00:02:12.530 a massive population of completely random grayscale 00:02:12.530 --> 00:02:15.900 patches. Like just pure visual static. Great, 00:02:15.900 --> 00:02:18.120 because the programmer never tells the machine 00:02:18.120 --> 00:02:20.740 how to draw a face. They only define a goal, 00:02:20.900 --> 00:02:24.479 which is resembling Darwin, and a mechanism for 00:02:24.479 --> 00:02:27.620 survival. So the computer ranks that static based 00:02:27.620 --> 00:02:30.039 on how closely it accidentally resembles the 00:02:30.039 --> 00:02:33.180 target. It trashes the worst static, takes the 00:02:33.180 --> 00:02:35.819 best performing patches, and breeds them together. 00:02:35.879 --> 00:02:38.479 It mixes their visual data, so that's the recombination. 00:02:38.719 --> 00:02:40.780 Exactly. And then it throws in a few random pixel 00:02:40.780 --> 00:02:43.620 changes, the mutation. Generation by generation, 00:02:43.939 --> 00:02:47.360 the noise evolves. Eyes emerge. A beard forms. 00:02:47.939 --> 00:02:49.860 And eventually, you are staring at a perfect 00:02:49.860 --> 00:02:52.180 portrait of Charles Darwin. Honestly, it reminds 00:02:52.180 --> 00:02:54.159 me of breeding racehorses. That is actually a 00:02:54.159 --> 00:02:55.919 very accurate way to frame it. Right, because 00:02:55.919 --> 00:02:57.879 if you want a fast horse with incredible stamina, 00:02:58.439 --> 00:03:00.800 you don't build one from scratch in a lab. You 00:03:00.800 --> 00:03:03.139 take a ridiculously fast horse and breed it with 00:03:03.139 --> 00:03:04.979 a horse that can run all day. And you just hope 00:03:04.979 --> 00:03:07.620 the offspring inherits both traits through that 00:03:07.620 --> 00:03:10.300 recombination. Yeah, and maybe a random genetic 00:03:10.300 --> 00:03:13.439 mutation gives the new full slightly larger lungs 00:03:13.439 --> 00:03:15.979 or something. You waste them, you keep the winners, 00:03:16.379 --> 00:03:18.580 and you put the losers out to pasture. You are 00:03:18.580 --> 00:03:21.000 mathematically speaking isolating the winning 00:03:21.000 --> 00:03:23.460 traits and combining them. Wait, I have to ask 00:03:23.460 --> 00:03:26.000 about the efficiency here. Yeah. I mean, calculating 00:03:26.000 --> 00:03:28.360 millions of artificial horse races in a computer 00:03:28.360 --> 00:03:32.300 seems incredibly slow and honestly chaotic. It 00:03:32.300 --> 00:03:34.360 can seem that way, yeah. Why go through all this 00:03:34.360 --> 00:03:37.319 trial and error? Why not just write an algebraic 00:03:37.319 --> 00:03:39.960 formula to find the mathematically perfect horse 00:03:39.960 --> 00:03:42.879 or that perfect visual pixel arrangement directly? 00:03:43.039 --> 00:03:45.199 Well, what's fascinating here is that traditional 00:03:45.199 --> 00:03:48.360 math completely breaks down when problems become 00:03:48.360 --> 00:03:51.439 multi -dimensional. Think about standard optimization 00:03:51.439 --> 00:03:54.080 algorithms like gradient descent, for example. 00:03:54.219 --> 00:03:57.360 OK. They are fantastic at finding an answer, 00:03:57.659 --> 00:03:59.840 but they frequently get trapped in what computer 00:03:59.840 --> 00:04:02.560 scientists call local minima. Local minima. Give 00:04:02.560 --> 00:04:04.759 me a visual for that. How does a computer get 00:04:04.759 --> 00:04:07.280 trapped? Okay, imagine you are blindfolded, and 00:04:07.280 --> 00:04:09.800 your only goal is to find the absolute lowest 00:04:09.800 --> 00:04:12.319 point in this massive, rugged mountain range. 00:04:12.960 --> 00:04:15.599 The traditional computer logic is simple. You 00:04:15.599 --> 00:04:17.439 feel the ground with your foot, and you always 00:04:17.439 --> 00:04:20.019 take a step downhill. I mean, that makes intuitive 00:04:20.019 --> 00:04:22.540 sense. If you always go down, you— eventually 00:04:22.540 --> 00:04:25.360 hit the bottom of the valley. Usually, yes. But 00:04:25.360 --> 00:04:27.220 what if you walk downhill and hit the bottom 00:04:27.220 --> 00:04:29.939 of a small crater halfway up the mountain? Ooh! 00:04:30.079 --> 00:04:32.120 Right. Every direction you step from the bottom 00:04:32.120 --> 00:04:34.339 of that crater is uphill. So the traditional 00:04:34.339 --> 00:04:36.980 algorithm feels the ground, realizes it can't 00:04:36.980 --> 00:04:39.220 go downhill anymore, and declares, I must be 00:04:39.220 --> 00:04:41.399 at the absolute lowest point on earth, problem 00:04:41.399 --> 00:04:44.550 solved. So it settles for a good but not best 00:04:44.550 --> 00:04:47.149 solution. It's physically trapped by its own 00:04:47.149 --> 00:04:49.829 rigid logic. Precisely. It stops searching entirely 00:04:49.829 --> 00:04:51.949 because the formula tells it the job is done. 00:04:52.069 --> 00:04:54.250 Wow. And this is where evolutionary computation 00:04:54.250 --> 00:04:57.310 comes in. It avoids this trap by using mutation. 00:04:57.859 --> 00:05:00.980 Mutation is a random, unpredictable jolt. Like 00:05:00.980 --> 00:05:03.360 an earthquake. Exactly. It's a sudden earthquake 00:05:03.360 --> 00:05:05.439 that bounces you entirely out of that crater, 00:05:05.879 --> 00:05:07.759 throwing you to a completely different part of 00:05:07.759 --> 00:05:09.480 the mountain range so you can keep exploring 00:05:09.480 --> 00:05:13.079 and find the true global valley floor. Without 00:05:13.079 --> 00:05:16.519 that random mistake, the system literally cannot 00:05:16.519 --> 00:05:19.459 find the best answer. To realize that a computer 00:05:19.459 --> 00:05:21.759 needed to make random mistakes to avoid those 00:05:21.759 --> 00:05:24.839 craters, that is a massive leap in logic. Especially 00:05:24.839 --> 00:05:26.980 when you consider the history. I mean, the computers 00:05:26.980 --> 00:05:30.399 of the 1950s and 60s were these massive, room 00:05:30.399 --> 00:05:33.060 -sized, incredibly expensive calculators. Yeah, 00:05:33.199 --> 00:05:35.160 they were precious resources. You did not just 00:05:35.160 --> 00:05:37.800 tell a mainframe to act randomly. So who had 00:05:37.800 --> 00:05:40.180 the guts to actually try this? Well, the concept 00:05:40.180 --> 00:05:42.579 of mimicking evolution in a machine goes all 00:05:42.579 --> 00:05:45.459 the way back to Alan Turing in 1948. Turing. 00:05:45.709 --> 00:05:48.509 Like the code breaker, the father of modern computing. 00:05:48.610 --> 00:05:51.050 The very same. He was decades ahead of his time. 00:05:51.470 --> 00:05:53.949 Turing proposed what he called B -type and P 00:05:53.949 --> 00:05:56.829 -type unorganized machines. Unorganized machines? 00:05:56.910 --> 00:05:58.850 Yeah, he was essentially mapping out primitive 00:05:58.850 --> 00:06:01.170 neural networks where the connections between 00:06:01.170 --> 00:06:04.110 nodes were learned via a genetic search. He designed 00:06:04.110 --> 00:06:06.329 them to use pleasure and pain signals to learn 00:06:06.329 --> 00:06:09.089 behaviors. Wait, pleasure and pain signals in 00:06:09.089 --> 00:06:12.899 1948? Why isn't Turing universally recognized 00:06:12.899 --> 00:06:15.319 as the founder of evolutionary computation? Oh, 00:06:15.459 --> 00:06:17.639 it is a matter of tragic timing. Turing passed 00:06:17.639 --> 00:06:21.620 away in 1954, and the brilliant 1948 paper of 00:06:21.620 --> 00:06:24.879 his went entirely unpublished until 1968. Oh, 00:06:24.980 --> 00:06:27.199 wow. It just sat in a drawer while the rest of 00:06:27.199 --> 00:06:29.079 the world caught it. Exactly. It had essentially 00:06:29.079 --> 00:06:31.420 zero impact on the early development of the field. 00:06:32.040 --> 00:06:34.000 Evolutionary computation had to be completely 00:06:34.000 --> 00:06:36.120 reinvented by independent researchers who had 00:06:36.120 --> 00:06:38.220 absolutely no idea Turing had already mapped 00:06:38.220 --> 00:06:41.240 the Turin. Which brings us to 1964, Germany. 00:06:41.470 --> 00:06:44.550 Two researchers, named Ingo Reckenberg and Hans 00:06:44.550 --> 00:06:47.209 -Paul Schweiffel, are trying to solve incredibly 00:06:47.209 --> 00:06:50.209 complex optimization problems in fluid dynamics. 00:06:50.670 --> 00:06:52.209 And they are the ones who really kick off a branch 00:06:52.209 --> 00:06:54.750 called evolution strategies. Right, because fluid 00:06:54.750 --> 00:06:57.129 dynamics is notoriously difficult to calculate. 00:06:57.449 --> 00:07:00.410 The math is highly non -linear. If you are trying 00:07:00.410 --> 00:07:02.850 to design the perfect shape for a pipe or, say, 00:07:02.910 --> 00:07:05.290 an airplane wing, a tiny change in the curve 00:07:05.290 --> 00:07:07.810 creates massive unpredictable changes in how 00:07:07.810 --> 00:07:10.709 the air or water flows. You cannot easily use 00:07:10.709 --> 00:07:13.319 a rigid formula to find the perfect shape. So 00:07:13.319 --> 00:07:15.560 they decided to use evolution. But the craziest 00:07:15.560 --> 00:07:17.660 part is, they didn't even have a computer to 00:07:17.660 --> 00:07:19.779 run the simulations at first. Oh, they didn't. 00:07:19.879 --> 00:07:22.939 They were sitting in a lab, literally rolling 00:07:22.939 --> 00:07:25.800 physical dice to determine the random mutations 00:07:25.800 --> 00:07:28.740 for their pipe shapes. Rolling dice to introduce 00:07:28.740 --> 00:07:31.000 that earthquake, that random jolt out of the 00:07:31.000 --> 00:07:33.660 crater, they would change a parameter based on 00:07:33.660 --> 00:07:36.420 the dice roll, manually test the physical result, 00:07:36.439 --> 00:07:38.379 and then just keep the shape that performed best. 00:07:38.480 --> 00:07:41.740 It sounds incredibly tedious. It does. But it 00:07:41.740 --> 00:07:43.980 proved that the fundamental math of evolution 00:07:43.980 --> 00:07:46.959 works regardless of the medium. You do not need 00:07:46.959 --> 00:07:49.800 silicon to run an evolutionary algorithm. Wow. 00:07:50.819 --> 00:07:52.939 And as computers did become more powerful through 00:07:52.939 --> 00:07:56.560 the 1960s and 70s, it seems like different scientists 00:07:56.560 --> 00:07:59.360 across the globe started building their own distinct 00:07:59.360 --> 00:08:02.220 evolutionary religions, so to speak. They all 00:08:02.220 --> 00:08:04.519 realized Darwinian math worked, but they had 00:08:04.519 --> 00:08:06.920 wildly different grades of coding it. Yeah, the 00:08:06.920 --> 00:08:09.480 global isolation during that era led to three 00:08:09.480 --> 00:08:12.279 distinct historic branches developing simultaneously. 00:08:13.040 --> 00:08:15.199 We have Reckenberg and Schreiffel in Germany, 00:08:15.680 --> 00:08:18.160 rolling dice for engineering parameters. But 00:08:18.160 --> 00:08:21.480 over in the United States around 1962, you have 00:08:21.480 --> 00:08:24.259 Lawrence J. Fogel. Developing evolutionary programming. 00:08:24.899 --> 00:08:27.259 Yes. Fogel was focused on prediction problems. 00:08:27.319 --> 00:08:30.420 Yeah. And the documentation notes he wasn't mutating 00:08:30.420 --> 00:08:33.529 physical parameters. He was mutating finite state 00:08:33.529 --> 00:08:35.429 machines. Can you break that down for us? Sure. 00:08:35.750 --> 00:08:38.029 A finite state machine is a mathematical model 00:08:38.029 --> 00:08:41.129 of computation. Think of it like a subway turnstile. 00:08:41.169 --> 00:08:44.009 Okay. It has a finite number of states. It is 00:08:44.009 --> 00:08:46.440 either locked or unlocked. If you insert a coin, 00:08:46.720 --> 00:08:48.720 the state transitions from locked to unlocked. 00:08:49.120 --> 00:08:51.259 If you push the arm, it transitions back to locked. 00:08:51.480 --> 00:08:54.399 It is a system of rigid rules and triggers. So 00:08:54.399 --> 00:08:56.100 Fogle would take an artificial intelligence running 00:08:56.100 --> 00:08:58.500 on a complex version of those rules. Instead 00:08:58.500 --> 00:09:00.799 of changing the data, he would randomly mutate 00:09:00.799 --> 00:09:03.399 the rules themselves. Exactly. Like spontaneously 00:09:03.399 --> 00:09:05.960 adding a new state. or deleting the rule that 00:09:05.960 --> 00:09:08.179 says inserting a coin unlocks the gate. That 00:09:08.179 --> 00:09:11.379 is the exact mechanism. He was mutating the logical 00:09:11.379 --> 00:09:13.879 flowchart of the machine itself to see if a slightly 00:09:13.879 --> 00:09:16.620 corrupted set of rules could predict a sequence 00:09:16.620 --> 00:09:19.139 of numbers better than the original rules. That 00:09:19.139 --> 00:09:21.679 is wild. But then you have John Henry Holland 00:09:21.679 --> 00:09:23.440 at the University of Michigan around the same 00:09:23.440 --> 00:09:26.879 time and his approach. genetic algorithms that 00:09:26.879 --> 00:09:30.120 feels much closer to literal biology. Yeah, Holland 00:09:30.120 --> 00:09:32.179 wanted to understand the actual mechanics of 00:09:32.179 --> 00:09:35.139 adaptation. He was asking, how does a biological 00:09:35.139 --> 00:09:37.700 organism physically pass down winning traits? 00:09:38.100 --> 00:09:41.080 To simulate this, he represented data as chromosomes 00:09:41.080 --> 00:09:44.539 using strings of binary bits, zeros and ones. 00:09:44.860 --> 00:09:47.720 So if we visualize that, a single solution in 00:09:47.720 --> 00:09:49.879 the computer's memory might look like the sequence 00:09:49.879 --> 00:09:54.169 0, 1, 1, 0. And another solution might be 1001. 00:09:54.409 --> 00:09:56.909 Exactly. And each position in that string represents 00:09:56.909 --> 00:10:00.210 a specific trait. Holland's true genius was simulating 00:10:00.210 --> 00:10:02.490 crossover. He wouldn't just mutate a single bit. 00:10:02.549 --> 00:10:04.669 He would take those two parent strings, cut them 00:10:04.669 --> 00:10:06.850 both in half, and swap their tails. Oh, wow. 00:10:07.029 --> 00:10:08.950 So the first half of parent A gets glued to the 00:10:08.950 --> 00:10:11.190 second half of parent B. creating a completely 00:10:11.190 --> 00:10:14.350 novel offspring that inherits complex combinations 00:10:14.350 --> 00:10:17.309 of traits from both parents. Furthermore, Holland 00:10:17.309 --> 00:10:19.289 wasn't just working with one parent and one child 00:10:19.289 --> 00:10:22.450 like Fogel. Holland created massive populations 00:10:22.450 --> 00:10:24.950 of these bit strings and had them compete in 00:10:24.950 --> 00:10:28.129 a giant ongoing tournament every single generation. 00:10:28.450 --> 00:10:31.450 An absolute battle royale for binary code. It 00:10:31.450 --> 00:10:33.710 is so fascinating that these brilliant minds 00:10:33.710 --> 00:10:36.690 were all stumbling toward the exact same biological 00:10:36.690 --> 00:10:39.950 truth from completely different angles. Fogel 00:10:39.950 --> 00:10:42.690 mutating logic gates, Holland cross -breeding 00:10:42.690 --> 00:10:45.549 data strings, Rechenberg mutating engineering 00:10:45.549 --> 00:10:47.649 shapes. And they were totally separate until 00:10:47.649 --> 00:10:50.110 the term evolutionary computing was finally coined 00:10:50.110 --> 00:10:53.730 in 1991 to unite the field. Unification is huge. 00:10:53.830 --> 00:10:56.509 It was. It allowed researchers to share concepts 00:10:56.509 --> 00:10:58.870 across disciplines, and it paved the way for 00:10:58.870 --> 00:11:01.289 the fourth major branch to emerge, which is perhaps 00:11:01.289 --> 00:11:03.090 the most radical of them all. Here's where it 00:11:03.090 --> 00:11:06.070 gets really interesting. In the 1990s, a researcher 00:11:06.070 --> 00:11:09.230 named John Koza pushes the boundary with genetic 00:11:09.230 --> 00:11:11.769 programming. Up until this point, researchers 00:11:11.769 --> 00:11:14.110 were evolving data or evolving the parameters 00:11:14.110 --> 00:11:17.269 of a solution. COSA decided to evolve the software 00:11:17.269 --> 00:11:20.590 itself. Yes. The subject of the evolution was 00:11:20.590 --> 00:11:23.750 a functional computer program written in a high 00:11:23.750 --> 00:11:26.409 -level language called Lisp. Evolving actual 00:11:26.409 --> 00:11:28.730 source code. Yeah. But wait, if you randomly 00:11:28.730 --> 00:11:31.610 slice two pieces of computer code together, wouldn't 00:11:31.610 --> 00:11:34.809 you just get a syntax error like... 99 % of the 00:11:34.809 --> 00:11:36.929 time? You would think so, yes. Because if you 00:11:36.929 --> 00:11:39.129 randomly delete a bracket or swap a variable 00:11:39.129 --> 00:11:41.830 in traditional code, the entire program just 00:11:41.830 --> 00:11:44.590 crashes and refuses to compile. Yeah. How does 00:11:44.590 --> 00:11:47.009 the software survive long enough to evolve? Well, 00:11:47.009 --> 00:11:49.929 that is the brilliant reason COSA used Lisp. 00:11:50.110 --> 00:11:52.529 Lisp utilizes something called s -expressions, 00:11:52.870 --> 00:11:55.490 which are structured as nested lists. You can 00:11:55.490 --> 00:11:57.690 visualize them as tree branches. OK, tree branches. 00:11:57.929 --> 00:12:00.190 The trunk of the tree is a mathematical operation, 00:12:00.509 --> 00:12:02.899 like addition. The branches coming off that trunk 00:12:02.899 --> 00:12:05.039 are the numbers being added. So instead of cutting 00:12:05.039 --> 00:12:08.419 text blindly, like a word processor, the algorithm 00:12:08.419 --> 00:12:11.460 is snapping off an entire functionally complete 00:12:11.460 --> 00:12:14.320 branch of logic, say, a branch that calculates 00:12:14.320 --> 00:12:17.279 a specific trajectory and plugging it into the 00:12:17.279 --> 00:12:20.549 socket of another program's tree. Exactly. Because 00:12:20.549 --> 00:12:23.370 it swaps whole functional branches, the grammar 00:12:23.370 --> 00:12:26.490 of the code remains completely intact. The new 00:12:26.490 --> 00:12:29.110 Frankenstein program might produce a terrible, 00:12:29.330 --> 00:12:32.090 completely wrong answer, but it will successfully 00:12:32.090 --> 00:12:34.649 run without a syntax error. That's incredible. 00:12:34.750 --> 00:12:36.970 It allows the computer to cross -breed its own 00:12:36.970 --> 00:12:39.429 instructions for sequence induction, pattern 00:12:39.429 --> 00:12:43.570 recognition, and complex planning. Software that 00:12:43.570 --> 00:12:45.970 writes and breeds better versions of itself. 00:12:46.240 --> 00:12:48.379 I mean, it sounds like a science fiction utopia. 00:12:48.539 --> 00:12:50.039 You just tell the computer what you want and 00:12:50.039 --> 00:12:53.299 it breeds the perfect app. The literature actually 00:12:53.299 --> 00:12:56.820 calls out a major modern problem in this field, 00:12:57.360 --> 00:12:59.799 a dark side to all this biological inspiration. 00:12:59.919 --> 00:13:01.820 Yeah, unfortunately. It's essentially academic 00:13:01.820 --> 00:13:04.419 spam. It is a significant hurdle right now. Because 00:13:04.419 --> 00:13:07.100 evolutionary computation is so deeply tied to 00:13:07.100 --> 00:13:10.120 biological metaphors, researchers in recent years 00:13:10.120 --> 00:13:12.379 have started proposing dubious novel algorithms 00:13:12.379 --> 00:13:15.159 with incredibly poor experimental validation. 00:13:15.419 --> 00:13:17.179 Yeah, there is an actual catalog keeping track 00:13:17.179 --> 00:13:19.779 of this called the Evolutionary Computation Bestiary. 00:13:20.100 --> 00:13:23.159 It sounds like a medieval book of monsters, but 00:13:23.159 --> 00:13:25.740 it's really a list of bad code. It highlights 00:13:25.740 --> 00:13:28.600 a really frustrating trend. Researchers will 00:13:28.600 --> 00:13:31.659 take a highly successful mathematically proven 00:13:31.659 --> 00:13:34.740 algorithm, frequently something called particle 00:13:34.740 --> 00:13:37.159 swarm optimization. Oh, break down how the swarm 00:13:37.159 --> 00:13:39.799 works. Particle swarm optimization mimics a flock 00:13:39.799 --> 00:13:42.539 of birds searching for food. You have a massive 00:13:42.539 --> 00:13:44.679 number of data points moving through a mathematical 00:13:44.679 --> 00:13:48.850 space. Every single point remembers its own personal 00:13:48.850 --> 00:13:52.250 best location. But it also constantly communicates 00:13:52.250 --> 00:13:54.610 with the rest of the flock to know the absolute 00:13:54.610 --> 00:13:57.049 best location found by the group so far. Right, 00:13:57.269 --> 00:14:00.250 sharing intel. Exactly. Every point adjusts its 00:14:00.250 --> 00:14:02.870 velocity and direction based on both its personal 00:14:02.870 --> 00:14:05.269 memory and the flock's collective knowledge. 00:14:05.710 --> 00:14:08.110 It's a gorgeous piece of math. The flock naturally 00:14:08.110 --> 00:14:10.590 swoops in and converges on the perfect solution. 00:14:10.809 --> 00:14:13.110 So what are these dubious researchers doing to 00:14:13.110 --> 00:14:15.610 it? They are stealing that exact mathematical 00:14:15.610 --> 00:14:18.350 engine, copying it line for line, and simply 00:14:18.350 --> 00:14:20.970 slapping a brand new biological metaphor on top 00:14:20.970 --> 00:14:23.169 of it. You're kidding. No, they swap the word 00:14:23.169 --> 00:14:26.830 bird for bat or cuckoo bird or, you know, gray 00:14:26.830 --> 00:14:30.409 wolf. They claim they have invented a revolutionary 00:14:30.409 --> 00:14:33.490 new gray wolf optimizer, published a paper, and 00:14:33.490 --> 00:14:36.070 just had their academic resumes. So they're looking 00:14:36.070 --> 00:14:39.330 at a solved problem and saying, behold. I've 00:14:39.330 --> 00:14:43.090 invented a new animal -themed algorithm. When 00:14:43.090 --> 00:14:45.090 under the hood, the engine hasn't changed at 00:14:45.090 --> 00:14:47.429 all. Does this mean the field has stalled out? 00:14:47.470 --> 00:14:50.629 Like are we just out of actual foundational ideas? 00:14:51.000 --> 00:14:53.820 Not at all. The copycats generate a lot of noise, 00:14:53.940 --> 00:14:56.659 but the true science is accelerating. Foundational 00:14:56.659 --> 00:14:58.820 evolutionary algorithms are legitimately being 00:14:58.820 --> 00:15:01.519 used right now to solve multi -dimensional logistics. 00:15:01.600 --> 00:15:04.440 They optimize the aerodynamics of jet engines 00:15:04.440 --> 00:15:07.220 and route global telecommunications networks. 00:15:07.639 --> 00:15:09.860 They are solving problems that are far too complex 00:15:09.860 --> 00:15:12.240 for human software designers to map out manually. 00:15:12.700 --> 00:15:14.960 The field hasn't stalled. It is simply dealing 00:15:14.960 --> 00:15:17.549 with the growing pains of its own success. That's 00:15:17.549 --> 00:15:19.450 reassuring. And while some researchers might 00:15:19.450 --> 00:15:21.830 be faking new algorithms, the original algorithm 00:15:21.830 --> 00:15:24.250 biological life itself is actually revealing 00:15:24.250 --> 00:15:26.690 secrets we never expected, all thanks to this 00:15:26.690 --> 00:15:28.909 computer science research. If we connect this 00:15:28.909 --> 00:15:31.710 to the bigger picture, genetic algorithms have 00:15:31.710 --> 00:15:35.230 completely revolutionized how we understand systems 00:15:35.230 --> 00:15:38.710 biology. This was the absolute biggest mind bender 00:15:38.710 --> 00:15:41.980 in the entire deep dive. The research suggests 00:15:41.980 --> 00:15:44.559 that biological systems aren't just like computers. 00:15:45.139 --> 00:15:47.700 It argues that biological systems are literal 00:15:47.700 --> 00:15:50.200 computational machines. Yeah, for a very long 00:15:50.200 --> 00:15:52.580 time we viewed biology through the lens of classical 00:15:52.580 --> 00:15:55.460 dynamical systems. We assumed development was 00:15:55.460 --> 00:15:58.200 orderly, well controlled, and governed by a top 00:15:58.200 --> 00:16:00.860 -down structure. Like a central brain or a master 00:16:00.860 --> 00:16:03.659 blueprint aggressively telling every single cell 00:16:03.659 --> 00:16:06.610 exactly what to do and where to go. Right. But 00:16:06.610 --> 00:16:08.970 studying evolutionary computation has proven 00:16:08.970 --> 00:16:11.330 that there is no central control in biological 00:16:11.330 --> 00:16:14.289 development. An embryo develops as a result of 00:16:14.289 --> 00:16:16.370 purely local interactions. Just cell -to -cell? 00:16:16.549 --> 00:16:18.669 Yes. A single cell doesn't know what the whole 00:16:18.669 --> 00:16:21.269 body looks like. It is essentially taking chemical 00:16:21.269 --> 00:16:23.330 inputs from its immediate neighbors, running 00:16:23.330 --> 00:16:26.409 an algorithm, and computing its next state. The 00:16:26.409 --> 00:16:29.289 cell is acting exactly like a logic gate in a 00:16:29.289 --> 00:16:32.509 silicon chip. It is a literal low -level computational 00:16:32.509 --> 00:16:35.549 operation. So what does this all mean? The documentation 00:16:35.549 --> 00:16:38.809 states that if you apply concepts from computational 00:16:38.809 --> 00:16:43.009 theory to biology, microprocesses in biological 00:16:43.009 --> 00:16:46.649 organisms are fundamentally incomplete and undecidable. 00:16:46.970 --> 00:16:49.509 Yes. Does that mean natural evolution operates 00:16:49.509 --> 00:16:52.129 mathematically like an infinite, unpredictable 00:16:52.129 --> 00:16:54.850 computer program? It does. This is proven through 00:16:54.850 --> 00:16:57.710 a theoretical framework called evolutionary automata. 00:16:57.830 --> 00:17:00.350 An automaton is essentially an abstract machine 00:17:00.350 --> 00:17:02.990 that follows a set of rules to process input, 00:17:02.990 --> 00:17:05.049 right? Like a Turing machine. A Turing machine 00:17:05.049 --> 00:17:07.930 is a highly advanced automaton that can simulate 00:17:07.930 --> 00:17:11.430 any step -by -step computer algorithm, but computer 00:17:11.430 --> 00:17:13.589 scientists have generalized this into evolutionary 00:17:13.589 --> 00:17:16.430 Turing machines. Okay. And the math proves that 00:17:16.430 --> 00:17:18.829 natural evolution and the algorithms we run on 00:17:18.829 --> 00:17:21.349 computers are mathematically undecidable. Meaning 00:17:21.349 --> 00:17:23.630 you literally cannot write a formula to predict 00:17:23.630 --> 00:17:25.990 the outcome of evolution. The only way to know 00:17:25.990 --> 00:17:28.170 what an evolutionary system will do is to press 00:17:28.170 --> 00:17:30.809 play, let it run, and watch what happens. It 00:17:30.809 --> 00:17:34.529 is deeply unpredictable. And even more profoundly, 00:17:35.009 --> 00:17:37.890 the simplest subclass of these theoretical machines, 00:17:38.470 --> 00:17:42.069 evolutionary finite automata, can process and 00:17:42.069 --> 00:17:45.289 accept incredibly complex arbitrary languages. 00:17:45.630 --> 00:17:48.150 Languages in the mathematical sense? Yes, including 00:17:48.150 --> 00:17:50.569 languages that are not recursively enumerable. 00:17:50.710 --> 00:17:53.140 Okay, we need to translate that. What does it 00:17:53.140 --> 00:17:56.220 mean for mathematical language to not be recursively 00:17:56.220 --> 00:17:58.799 enumerable? Well, a standard traditional computer 00:17:58.799 --> 00:18:01.220 can only solve problems that have a step -by 00:18:01.220 --> 00:18:03.539 -step algorithmic solution. If you can list the 00:18:03.539 --> 00:18:05.720 steps, the computer can solve it. Makes sense. 00:18:05.900 --> 00:18:08.480 A non -recursively enumerable problem is one 00:18:08.480 --> 00:18:11.759 where no step -by -step algorithm exists. A standard 00:18:11.759 --> 00:18:13.980 computer will freeze or run forever trying to 00:18:13.980 --> 00:18:16.369 solve it. but an evolutionary automaton can solve 00:18:16.369 --> 00:18:17.990 it. Because it doesn't rely on step -by -step 00:18:17.990 --> 00:18:20.869 logic. Exactly. It uses massive populations and 00:18:20.869 --> 00:18:23.329 random mutations to explore every possibility 00:18:23.329 --> 00:18:26.569 simultaneously. It confirms that nature is, mathematically 00:18:26.569 --> 00:18:29.450 speaking, the ultimate supercomputer. It can 00:18:29.450 --> 00:18:32.230 process information that our rigid, linear silicon 00:18:32.230 --> 00:18:34.849 machines cannot even comprehend. It's staggering. 00:18:35.150 --> 00:18:37.630 It really shifts how you look at the world around 00:18:37.630 --> 00:18:40.029 you. Because for you listening to this right 00:18:40.029 --> 00:18:42.990 now, whether you're trying to optimize a global 00:18:42.990 --> 00:18:45.170 supply chain, or you're a developer trying to 00:18:45.170 --> 00:18:47.849 train a massive new AI model, or honestly, if 00:18:47.849 --> 00:18:49.269 you're just looking out the window at a tree, 00:18:49.890 --> 00:18:52.230 you are witnessing the exact same fundamental 00:18:52.230 --> 00:18:56.529 math at work. Nature's slow, brutal optimization 00:18:56.529 --> 00:19:00.109 process is now the very thing designing our most 00:19:00.109 --> 00:19:02.529 advanced technologies. This raises an important 00:19:02.529 --> 00:19:05.470 question, though. A truly profound concept that 00:19:05.470 --> 00:19:07.990 the research points toward by including universal 00:19:07.990 --> 00:19:10.329 Darwinism in the broader implications of this 00:19:10.329 --> 00:19:12.950 field. Universal Darwinism. We're expanding beyond 00:19:12.950 --> 00:19:15.190 biology and computers entirely. Lay it on me. 00:19:15.349 --> 00:19:18.410 Well, if evolutionary finite automata can process 00:19:18.410 --> 00:19:21.170 these unimaginably complex, non -recursively 00:19:21.170 --> 00:19:23.829 innumerable languages, and if we've proven that 00:19:23.829 --> 00:19:26.309 both biological cells and silicon chips are really 00:19:26.309 --> 00:19:28.210 just different types of hardware running the 00:19:28.210 --> 00:19:30.589 exact same underlying evolutionary software, 00:19:30.650 --> 00:19:32.589 Oh man, where are you going with this? Is the 00:19:32.589 --> 00:19:35.789 entire universe itself just one giant ongoing 00:19:35.789 --> 00:19:39.529 evolutionary computation? Is reality fundamentally 00:19:39.529 --> 00:19:42.089 a massive algorithm trying to solve an unknown, 00:19:42.349 --> 00:19:44.670 undecidable problem? Oh wow, that is the kind 00:19:44.670 --> 00:19:46.559 of thought that fundamentally changes how you 00:19:46.559 --> 00:19:48.519 view your own existence? Are we just the random 00:19:48.519 --> 00:19:50.799 mutations trying to bounce the universe out of 00:19:50.799 --> 00:19:54.160 a local minimum? Incredible. Well, we will leave 00:19:54.160 --> 00:19:56.720 you to ponder that one on your own. Thank you 00:19:56.720 --> 00:19:58.700 so much for joining us on this deep dive. Keep 00:19:58.700 --> 00:20:00.660 questioning the world around you. And the next 00:20:00.660 --> 00:20:02.519 time you use a piece of cutting edge tech, just 00:20:02.519 --> 00:20:05.099 remember, it is probably running on Mother Nature's 00:20:05.099 --> 00:20:05.859 stolen playbook.