WEBVTT 00:00:00.000 --> 00:00:02.379 All right, so for this deep dive, I want you 00:00:02.379 --> 00:00:04.960 to just close your eyes for a second. Picture 00:00:04.960 --> 00:00:06.780 something with me. Okay, I'm with you. We're 00:00:06.780 --> 00:00:09.599 in upstate New York, and I don't mean, you know, 00:00:09.599 --> 00:00:11.919 the Hudson Valley with all the cute bed and breakfasts. 00:00:11.960 --> 00:00:14.480 No, we're talking way upstate, near the Canadian 00:00:14.480 --> 00:00:18.019 border almost. Exactly, and we're talking 1930s, 00:00:18.019 --> 00:00:20.480 the Great Depression is just grinding everything 00:00:20.480 --> 00:00:24.300 to a halt. We're in a town called Peru. Not the 00:00:24.300 --> 00:00:26.460 country, the town. Right, the town. It's right 00:00:26.460 --> 00:00:28.940 on the edge of the Adirondacks. Super close to 00:00:28.940 --> 00:00:31.660 Lake Champlain. It's rugged. It's rural. And 00:00:31.660 --> 00:00:34.359 it is completely disconnected from what we would 00:00:34.359 --> 00:00:37.280 think of as the modern world. Totally. So on 00:00:37.280 --> 00:00:40.259 this farm, there's no electricity, no indoor 00:00:40.259 --> 00:00:43.060 plumbing, no running water. It's a life that's 00:00:43.060 --> 00:00:45.520 dictated by the sun and the seasons, not by, 00:00:45.619 --> 00:00:48.280 you know, a train schedule or a clock on the 00:00:48.280 --> 00:00:51.619 wall. It's all hard work and harsh winters. And 00:00:51.619 --> 00:00:53.659 there's a one -room schoolhouse about a mile's 00:00:53.659 --> 00:00:56.479 walk down a dirt road. And on this farm... Lives 00:00:56.479 --> 00:00:59.219 a young girl. She's the oldest of six kids. Yeah. 00:00:59.520 --> 00:01:01.479 And she's doing the farm work. She's milking 00:01:01.479 --> 00:01:04.000 cows. She's helping with the harvest. If you 00:01:04.000 --> 00:01:05.659 saw her back then, you would think you were looking 00:01:05.659 --> 00:01:07.500 at a character straight out of Little House on 00:01:07.500 --> 00:01:09.239 the Prairie. You definitely wouldn't look at 00:01:09.239 --> 00:01:11.640 her and think, there, there goes the architect 00:01:11.640 --> 00:01:14.799 of all modern computing. Right. It's just, it 00:01:14.799 --> 00:01:17.840 feels like a complete category error. And yet, 00:01:17.920 --> 00:01:21.549 that girl. Frances Fran Allen goes from that 00:01:21.549 --> 00:01:24.510 farm to becoming the very first woman to win 00:01:24.510 --> 00:01:27.609 the A .M. Turing Award. The Nobel Prize of Computing. 00:01:27.609 --> 00:01:29.709 It's the highest honor there is in the field. 00:01:29.829 --> 00:01:32.069 She didn't just like participate in the digital 00:01:32.069 --> 00:01:36.209 revolution. She fundamentally wrote the grammar 00:01:36.209 --> 00:01:39.230 for how we talk to machines efficiently. It's 00:01:39.230 --> 00:01:42.129 one of the most. I think, unlikely trajectories 00:01:42.129 --> 00:01:43.849 in the entire history of science. That really 00:01:43.849 --> 00:01:45.769 is. And what I love about her story is that it 00:01:45.769 --> 00:01:47.590 wasn't a straight line. It wasn't, you know, 00:01:47.650 --> 00:01:50.090 child prodigy builds a radio in the garage, goes 00:01:50.090 --> 00:01:53.129 to MIT, changes the world. No, it was messy. 00:01:53.290 --> 00:01:56.109 It was. It was driven by accident, by necessity, 00:01:56.269 --> 00:01:59.030 and to be honest, by a little bit of desperation. 00:01:59.489 --> 00:02:01.450 That's the part that gets me. The desperation. 00:02:02.469 --> 00:02:05.390 She didn't show up at IBM on day one planning 00:02:05.390 --> 00:02:08.490 to invent the future or save the world. Not at 00:02:08.490 --> 00:02:10.669 all. She joined because she had bills to pay. 00:02:10.789 --> 00:02:13.969 She had student loans. The great universal motivator. 00:02:14.069 --> 00:02:16.590 Student debt has probably driven more innovation 00:02:16.590 --> 00:02:19.189 than we give it credit for. It's so true. So 00:02:19.189 --> 00:02:22.009 our mission for this deep dive is to trace that 00:02:22.009 --> 00:02:24.250 journey. We are going to walk from that farm 00:02:24.250 --> 00:02:26.930 in Peru through the high -security computer labs 00:02:26.930 --> 00:02:29.949 of the NSA and finally land on the algorithms 00:02:29.949 --> 00:02:33.330 that probably help run the software we are using 00:02:33.330 --> 00:02:35.669 to record this right now. We are talking about 00:02:35.669 --> 00:02:38.509 Fran Allen, the woman who taught computers how 00:02:38.509 --> 00:02:41.669 to be efficient. And to really get why she is 00:02:41.669 --> 00:02:43.669 so important, we have to understand what that 00:02:43.669 --> 00:02:46.349 word efficiency even meant in the world she was 00:02:46.349 --> 00:02:48.409 working in. Oh, absolutely. Because today efficiency 00:02:48.409 --> 00:02:51.090 means, what, your phone battery lasts an extra 00:02:51.090 --> 00:02:53.370 hour? Right. Back then, it meant something completely 00:02:53.370 --> 00:02:55.349 different. It was the difference between something 00:02:55.349 --> 00:02:58.030 being possible and impossible. So let's set the 00:02:58.030 --> 00:02:59.590 stage. We need to do a little time travel. It's 00:02:59.590 --> 00:03:03.110 1957. Eisenhower is president. Sputnik is about 00:03:03.110 --> 00:03:06.490 to launch and scare the absolute daylights out 00:03:06.490 --> 00:03:09.550 of America. And Fran Allen has just walked through 00:03:09.550 --> 00:03:11.550 the doors of IBM Research in Poughkeepsie, New 00:03:11.550 --> 00:03:14.710 York. What's the vibe in that building? The vibe 00:03:14.710 --> 00:03:17.189 is, well, it's distinct. You have to remember 00:03:17.189 --> 00:03:21.590 computer science as a major, as a degree. It 00:03:21.590 --> 00:03:24.370 didn't exist. There were no CS departments at 00:03:24.370 --> 00:03:26.289 universities. Who were these people who was working 00:03:26.289 --> 00:03:28.330 there? They were mathematicians, physicists, 00:03:28.969 --> 00:03:32.009 a few electrical engineers, maybe some logicians. 00:03:32.210 --> 00:03:35.129 They were people from other fields who were drawn 00:03:35.129 --> 00:03:37.550 to these new machines. And the machines themselves. 00:03:38.360 --> 00:03:40.219 They weren't sitting on desks. No, the machines 00:03:40.219 --> 00:03:42.379 were the desks. Or really, they were the entire 00:03:42.379 --> 00:03:45.780 room. These giant humming beasts. Full of vacuum 00:03:45.780 --> 00:03:48.780 tubes cooled by massive air conditioning units 00:03:48.780 --> 00:03:51.139 just to keep them from literally melting down. 00:03:51.280 --> 00:03:53.919 And the culture around them was... It was very 00:03:53.919 --> 00:03:55.759 much like a priesthood. A priesthood. I like 00:03:55.759 --> 00:03:57.979 that. Yeah. You had the hardware guys who actually 00:03:57.979 --> 00:03:59.840 built the machine, and then you had the researchers 00:03:59.840 --> 00:04:01.400 who were trying to figure out how to talk to 00:04:01.400 --> 00:04:03.300 it. And talking to it was a total nightmare. 00:04:03.520 --> 00:04:07.099 It was excruciating. In 1957, if you wanted the 00:04:07.099 --> 00:04:09.120 computer to do something simple like add two 00:04:09.120 --> 00:04:11.360 numbers, you had to speak its native language. 00:04:11.659 --> 00:04:13.740 Assembly code. Assembly code. You had to tell 00:04:13.740 --> 00:04:16.519 it, go to this exact memory address, move the 00:04:16.519 --> 00:04:18.379 data from there into this specific register. 00:04:18.819 --> 00:04:21.360 Now go to this other address, get that data, 00:04:21.459 --> 00:04:23.759 add it to the first register. and then put the 00:04:23.759 --> 00:04:25.939 result in this third location. It's just one 00:04:25.939 --> 00:04:28.759 tiny step above writing in raw zeros and ones. 00:04:29.160 --> 00:04:32.079 Barely. It's cryptic. It's incredibly tedious. 00:04:32.139 --> 00:04:34.319 And if you make a single typo, the whole program 00:04:34.319 --> 00:04:38.680 just crashes. No error messages. But, and I guess 00:04:38.680 --> 00:04:40.779 this is the important part, it gave you absolute 00:04:40.779 --> 00:04:43.639 control. Total control. And the programmers of 00:04:43.639 --> 00:04:45.579 that era, let's call them the assembly cult. 00:04:45.800 --> 00:04:48.740 The wizards. The wizards, yeah. They prided themselves 00:04:48.740 --> 00:04:51.639 on their ability to squeeze every single ounce 00:04:51.639 --> 00:04:54.019 of performance out of that hardware. They knew 00:04:54.019 --> 00:04:56.060 exactly how many milliseconds each instruction 00:04:56.060 --> 00:04:59.139 took. They were true artisans. And right into 00:04:59.139 --> 00:05:01.339 the middle of this world of wizards, a new thing 00:05:01.339 --> 00:05:05.410 appears. Fortran. Right. Fortran, the first high 00:05:05.410 --> 00:05:08.170 -level programming language, had only been announced 00:05:08.170 --> 00:05:10.529 maybe two months before Fran Allen even started 00:05:10.529 --> 00:05:13.629 at IBM. And the idea was completely radical for 00:05:13.629 --> 00:05:15.730 the time. Totally. Instead of writing all that 00:05:15.730 --> 00:05:17.670 cryptic assembly code, you could just write something 00:05:17.670 --> 00:05:19.889 that looks like a math equation. You could write 00:05:19.889 --> 00:05:23.069 C equals A plus B. And then a special program, 00:05:23.370 --> 00:05:26.230 a compiler, would do the hard work of translating 00:05:26.230 --> 00:05:28.569 that into the machine code for you. That was 00:05:28.569 --> 00:05:31.790 the promise. That sounds amazing. Infinitely 00:05:31.790 --> 00:05:34.149 better. Why would anyone possibly object to that? 00:05:34.350 --> 00:05:37.189 Because the wizards didn't trust the translator. 00:05:37.529 --> 00:05:39.810 They looked at Fortran and the idea of a compiler 00:05:39.810 --> 00:05:43.110 and they said, there is no way a machine can 00:05:43.110 --> 00:05:45.410 write code that is as good as the code I can 00:05:45.410 --> 00:05:47.949 write by hand. They thought it would be messy. 00:05:48.129 --> 00:05:49.829 They believed it would produce what they called 00:05:49.829 --> 00:05:54.649 spaghetti code. Bloated, slow, inefficient instructions 00:05:54.649 --> 00:05:57.589 that wasted the machine's precious time. I get 00:05:57.589 --> 00:05:59.470 it. It's like a master chef watching someone 00:05:59.470 --> 00:06:01.769 use a microwave. That's a great analogy. They're 00:06:01.769 --> 00:06:04.550 thinking, sure, it's faster, but the result is 00:06:04.550 --> 00:06:08.500 going to be garbage. Exactly. And in 1957, garbage 00:06:08.500 --> 00:06:11.480 code wasn't just, you know, an aesthetic problem. 00:06:11.620 --> 00:06:14.480 It was incredibly expensive. Computer time was 00:06:14.480 --> 00:06:17.339 rented out for hundreds, sometimes thousands 00:06:17.339 --> 00:06:20.980 of dollars an hour. Memory was measured in kilobytes. 00:06:21.180 --> 00:06:23.660 So every wasted instruction, every wasted byte 00:06:23.660 --> 00:06:26.480 of memory was literally burning money. It was 00:06:26.480 --> 00:06:29.720 burning real money. So the skepticism from the 00:06:29.720 --> 00:06:32.579 establishment was immense. And that is the world 00:06:32.579 --> 00:06:34.990 that Fran Allen walks into. So let's back. Back 00:06:34.990 --> 00:06:36.889 up just a little bit. How does she even get there? 00:06:36.970 --> 00:06:39.209 We said she was a teacher. She was very much 00:06:39.209 --> 00:06:41.410 on the education track. She graduated from the 00:06:41.410 --> 00:06:43.550 New York State College for Teachers, which is 00:06:43.550 --> 00:06:47.509 now SUNY Albany, back in 1954. And she goes right 00:06:47.509 --> 00:06:50.029 back to her hometown. Right back to Peru, New 00:06:50.029 --> 00:06:52.490 York. stood in front of a classroom teaching 00:06:52.490 --> 00:06:54.829 algebra and trigonometry to high schoolers. I 00:06:54.829 --> 00:06:56.550 can just picture her, you know, chalk dust on 00:06:56.550 --> 00:06:58.529 her hands, trying to convince a bunch of teenagers 00:06:58.529 --> 00:07:00.949 to care about cosines. And by all accounts, she 00:07:00.949 --> 00:07:02.970 loved it. She was good at it. She actually wanted 00:07:02.970 --> 00:07:05.149 that to be her career. But the state of New York 00:07:05.149 --> 00:07:08.810 had a requirement. Ah, bureaucracy. Always. To 00:07:08.810 --> 00:07:10.550 get your permanent teaching certificate, you 00:07:10.550 --> 00:07:12.889 had to have a master's degree. The old credential 00:07:12.889 --> 00:07:15.850 treadmill. It gets us all. So she enrolls at 00:07:15.850 --> 00:07:17.569 the University of Michigan, gets her master's 00:07:17.569 --> 00:07:21.230 in mathematics in 1957. But now she has a new 00:07:21.230 --> 00:07:23.829 problem. Money. She's been in grad school for 00:07:23.829 --> 00:07:26.389 a couple of years, not earning anything, and 00:07:26.389 --> 00:07:29.209 she's taken out student loans. She's in debt. 00:07:29.410 --> 00:07:32.089 Deep in the red. And that's when IBM recruiters 00:07:32.089 --> 00:07:34.850 show up on campus. They're hiring pretty much 00:07:34.850 --> 00:07:37.310 anyone with a strong math background. The pay 00:07:37.310 --> 00:07:39.790 is really good. So she makes a practical decision. 00:07:40.009 --> 00:07:43.199 A purely practical one. She thinks, OK. I'll 00:07:43.199 --> 00:07:45.560 go work for this company in Poughkeepsie for, 00:07:45.660 --> 00:07:47.939 I don't know, a year, maybe two. I'll pay off 00:07:47.939 --> 00:07:50.180 this debt and then I'm right back to the chalkboard 00:07:50.180 --> 00:07:52.899 in Peru. And that two -year plan ends up becoming 00:07:52.899 --> 00:07:55.819 a 45 -year career. It's the classic accidental 00:07:55.819 --> 00:07:58.379 career. But this is where her background as a 00:07:58.379 --> 00:08:01.980 teacher becomes her superpower. IBM hires her. 00:08:02.120 --> 00:08:04.759 She knows absolutely nothing about computers. 00:08:04.980 --> 00:08:07.180 So what do they have her do? Naturally, they 00:08:07.180 --> 00:08:09.740 assign her to teach the new employees how to 00:08:09.740 --> 00:08:13.180 use Fortran. Wait, hold on. They assign the person 00:08:13.180 --> 00:08:16.339 who doesn't know the new language to teach the 00:08:16.339 --> 00:08:18.699 new language. That feels very corporate, doesn't 00:08:18.699 --> 00:08:20.540 it? It really does. But you have to remember, 00:08:20.720 --> 00:08:23.040 it was the Wild West back then. They probably 00:08:23.040 --> 00:08:24.779 thought, well, she has a teaching degree. She's 00:08:24.779 --> 00:08:27.620 a teacher. Great. Okay, but think about the position 00:08:27.620 --> 00:08:30.639 that puts her in. She has to stand up in front 00:08:30.639 --> 00:08:34.100 of a room full of incredibly smart people. these 00:08:34.100 --> 00:08:37.139 mathematicians and physicists, and explain this 00:08:37.139 --> 00:08:40.120 brand new, highly controversial tool. And she 00:08:40.120 --> 00:08:42.340 can't bluff. She has to actually know her stuff. 00:08:42.600 --> 00:08:45.019 So she has to learn it and learn it fast. Faster 00:08:45.019 --> 00:08:47.980 than fast. She described it as learning the concepts 00:08:47.980 --> 00:08:51.100 days, sometimes just hours, before she had to 00:08:51.100 --> 00:08:53.840 give the lecture. But how? There were no textbooks 00:08:53.840 --> 00:08:57.139 yet. Fortran was too new. Where did she go for 00:08:57.139 --> 00:09:00.159 information? This is the pivotal moment. This 00:09:00.159 --> 00:09:02.519 is the act that defines her entire intellectual 00:09:02.519 --> 00:09:06.159 legacy. With no manual and no grammar book, she 00:09:06.159 --> 00:09:08.039 went to the only source of truth that existed. 00:09:08.379 --> 00:09:10.820 What was that? She got her hands on the source 00:09:10.820 --> 00:09:14.059 code of the Fortran compiler itself. Whoa. Okay, 00:09:14.100 --> 00:09:15.820 unpack what that means. What does that physically 00:09:15.820 --> 00:09:18.679 look like in 1957? It would have been stacks 00:09:18.679 --> 00:09:21.860 and stacks of paper. Huge printouts of assembly 00:09:21.860 --> 00:09:25.220 code. She was reading the raw instructions that 00:09:25.220 --> 00:09:27.220 the creators of Fortran, John Backus, and his 00:09:27.220 --> 00:09:29.679 team had written to build the translator. So 00:09:29.679 --> 00:09:31.899 it's like trying to understand how a car works. 00:09:32.379 --> 00:09:35.100 Not by reading the owner's manual, but by taking 00:09:35.100 --> 00:09:37.899 the entire engine apart piece by piece. That 00:09:37.899 --> 00:09:41.120 is a perfect analogy, and it is incredibly difficult. 00:09:41.200 --> 00:09:43.659 Reading someone else's code is hard today with 00:09:43.659 --> 00:09:46.779 all our modern tools. Reading uncommented assembly 00:09:46.779 --> 00:09:49.580 code from 1957 must have been like trying to 00:09:49.580 --> 00:09:52.419 decipher ancient hieroglyphics. But doing that 00:09:52.419 --> 00:09:54.279 must have given her an insight that almost no 00:09:54.279 --> 00:09:57.110 one else had. A profound insight. Most people 00:09:57.110 --> 00:09:58.990 were just looking at the surface of Fortran, 00:09:59.070 --> 00:10:01.490 you know, the syntax, the commands. Fran Allen 00:10:01.490 --> 00:10:03.350 was looking underneath it. She was seeing the 00:10:03.350 --> 00:10:05.669 machinery. She saw the machinery. She saw how 00:10:05.669 --> 00:10:08.049 the compiler took a simple command like C plus 00:10:08.049 --> 00:10:11.090 B and twisted and turned it into machine logic. 00:10:11.309 --> 00:10:12.970 She saw the places where it was inefficient. 00:10:13.230 --> 00:10:15.889 She saw the brilliant little shortcuts the original 00:10:15.889 --> 00:10:18.830 programmers had used. She saw all the messy parts. 00:10:19.070 --> 00:10:21.750 She saw the matrix? She really did. And in that 00:10:21.750 --> 00:10:23.809 process, she realized something fundamental. 00:10:24.429 --> 00:10:26.809 Tyler wasn't just a simple translator. It was 00:10:26.809 --> 00:10:30.289 a bridge between the human world and the machine 00:10:30.289 --> 00:10:32.269 world. And if you could make that bridge better. 00:10:32.450 --> 00:10:34.850 If you could optimize how that translation happened, 00:10:35.049 --> 00:10:37.610 you could make the hardware do things its own 00:10:37.610 --> 00:10:39.970 designers never even dreamed of. You could unlock 00:10:39.970 --> 00:10:43.289 its true potential. That moment, reading the 00:10:43.289 --> 00:10:46.639 source code. that's the spark that's where the 00:10:46.639 --> 00:10:49.639 math teacher truly becomes the computer scientist 00:10:49.639 --> 00:10:52.320 absolutely she wasn't just conveying information 00:10:52.320 --> 00:10:54.860 anymore she was analyzing the very structure 00:10:54.860 --> 00:10:57.159 of information itself and she got good at it 00:10:57.529 --> 00:10:59.870 Really, really good. Which is why I'm guessing 00:10:59.870 --> 00:11:01.909 she didn't stay in that teaching role for very 00:11:01.909 --> 00:11:04.370 long. No. Pretty quickly, IBM realized what they 00:11:04.370 --> 00:11:06.210 had and they pulled her off the teaching gig 00:11:06.210 --> 00:11:08.929 and threw her into something much, much darker. 00:11:09.230 --> 00:11:12.370 The spy games. The spy games. So we've jumped 00:11:12.370 --> 00:11:15.049 forward a little bit. It's 1959. The Cold War 00:11:15.049 --> 00:11:18.289 is, well, it's freezing. The tension with the 00:11:18.289 --> 00:11:20.490 Soviet Union is at an all -time high. And the 00:11:20.490 --> 00:11:22.350 U .S. government needs a serious technological 00:11:22.350 --> 00:11:26.850 edge. Enter a new project. Project Harvest. Harvest. 00:11:26.970 --> 00:11:29.789 It sounds so pastoral and friendly, doesn't it? 00:11:29.830 --> 00:11:31.830 Like she's going back to her roots on the farm. 00:11:31.970 --> 00:11:33.590 I have a feeling they weren't counting bushels 00:11:33.590 --> 00:11:36.830 of wheat with this one. No. The client for Project 00:11:36.830 --> 00:11:39.289 Harvest was the National Security Agency, the 00:11:39.289 --> 00:11:42.470 NSA. And the goal? The goal was cryptanalysis, 00:11:42.590 --> 00:11:45.809 code breaking. The NSA was intercepting huge 00:11:45.809 --> 00:11:48.970 volumes of Soviet communications, and they needed 00:11:48.970 --> 00:11:51.210 to decrypt them as close to real time as possible. 00:11:51.429 --> 00:11:55.049 Real time, in 1959. yeah with that technology 00:11:55.049 --> 00:11:57.809 that was the dream or as close as they could 00:11:57.809 --> 00:11:59.870 get the amount of data they were dealing with 00:11:59.870 --> 00:12:02.590 was massive far beyond what any existing computer 00:12:02.590 --> 00:12:05.200 could handle they needed a new machine So IBM 00:12:05.200 --> 00:12:07.399 is building the hardware. IBM was building a 00:12:07.399 --> 00:12:09.519 supercomputer called Stretch, which was already 00:12:09.519 --> 00:12:12.259 a huge leap forward. But for the NSA, they built 00:12:12.259 --> 00:12:14.899 a highly specialized, customized version of Stretch. 00:12:15.100 --> 00:12:17.179 And that was called Harvest. What made Harvest 00:12:17.179 --> 00:12:19.860 so special? It was a monster. It had unique, 00:12:20.080 --> 00:12:22.940 specialized hardware designed specifically for 00:12:22.940 --> 00:12:24.899 streaming massive amounts of data and looking 00:12:24.899 --> 00:12:27.460 for patterns. It was essentially built to scan 00:12:27.460 --> 00:12:30.379 for keywords and recurring sequences in encrypted 00:12:30.379 --> 00:12:33.899 text at incredible speeds. But a powerful piece. 00:12:34.000 --> 00:12:36.539 of hardware is just an expensive paperweight 00:12:36.539 --> 00:12:38.200 if you don't have the software to control it. 00:12:38.399 --> 00:12:41.679 Exactly. And that's where Fran comes in. They 00:12:41.679 --> 00:12:43.899 needed a new programming language that could 00:12:43.899 --> 00:12:46.299 harness the power of this unique machine. The 00:12:46.299 --> 00:12:48.019 language was called Alpha, right? That's right. 00:12:48.600 --> 00:12:51.919 And Alan was assigned to help design Alpha. But, 00:12:51.960 --> 00:12:53.940 and this is the crucial part, she was also put 00:12:53.940 --> 00:12:55.759 in charge of managing the compiler optimization 00:12:55.759 --> 00:12:59.179 team for the entire project. Okay, so this is 00:12:59.179 --> 00:13:01.480 where the stakes for optimization get a lot higher. 00:13:01.919 --> 00:13:03.899 Oh, they go through the roof. It's no longer 00:13:03.899 --> 00:13:06.879 about saving IBM a few hundred dollars in computer 00:13:06.879 --> 00:13:09.759 time. No. Now if your compiler generates inefficient 00:13:09.759 --> 00:13:12.139 code, it means the intelligence report comes 00:13:12.139 --> 00:13:14.340 in three days too late and you miss a potential 00:13:14.340 --> 00:13:17.440 missile launch. Precisely. Speed was the only 00:13:17.440 --> 00:13:19.779 metric that mattered. Fran Allen was leading 00:13:19.779 --> 00:13:22.379 a team that had to ensure the software they wrote 00:13:22.379 --> 00:13:25.519 didn't waste a single clock cycle of that powerful, 00:13:25.779 --> 00:13:28.500 one of a kind hardware. She's sitting right at 00:13:28.500 --> 00:13:30.740 the intersection of top secret government intelligence 00:13:30.740 --> 00:13:33.980 and the absolute bleeding edge of technology. 00:13:34.419 --> 00:13:37.299 And this was all happening behind a wall of secrecy. 00:13:37.379 --> 00:13:39.759 Most of the team members didn't even know the 00:13:39.759 --> 00:13:42.039 full scope of what they were working on until 00:13:42.039 --> 00:13:44.940 details started leaking to the press years later. 00:13:45.080 --> 00:13:46.850 I have to pause. on this for a second. She's 00:13:46.850 --> 00:13:50.269 a woman in 1959 leading one of the most critical 00:13:50.269 --> 00:13:53.110 software teams for a national security project. 00:13:53.389 --> 00:13:55.690 It's remarkable. We talk a lot about the women 00:13:55.690 --> 00:13:58.330 of hidden figures at NASA who were the computers. 00:13:58.950 --> 00:14:01.309 But here's Fran Allen, not just doing the math, 00:14:01.409 --> 00:14:03.889 but designing the entire compiler architecture 00:14:03.889 --> 00:14:07.750 for the NSA. Was she the only woman on that team? 00:14:07.990 --> 00:14:10.149 In those high -level architectural meetings, 00:14:10.429 --> 00:14:13.980 often yes. She wasn't the only woman at IBM Research, 00:14:14.120 --> 00:14:15.720 and we'll get into that. But in her specific 00:14:15.720 --> 00:14:19.039 role at that level of leadership, she was a pioneer. 00:14:19.320 --> 00:14:21.759 And the system they built, Harvest, it was used 00:14:21.759 --> 00:14:25.759 until 1976. A 17 -year run. That is an eternity 00:14:25.759 --> 00:14:27.899 in the world of computing. That's a testament 00:14:27.899 --> 00:14:29.720 to how well they designed it. So she's cut her 00:14:29.720 --> 00:14:31.940 teeth on Fortran. She's been battle -tested by 00:14:31.940 --> 00:14:34.500 the NSA. Now we get to the core of her legacy. 00:14:34.919 --> 00:14:37.200 The optimizing compiler. This is the deep dive. 00:14:37.399 --> 00:14:39.120 We've used that phrase a few times, but I really 00:14:39.120 --> 00:14:41.960 want to break it down because I feel like optimization 00:14:41.960 --> 00:14:45.440 is a word we all throw around very loosely. You 00:14:45.440 --> 00:14:47.779 know, I need to optimize my morning routine. 00:14:47.879 --> 00:14:50.659 Right. Or I've optimized my sandwich making process. 00:14:50.840 --> 00:14:53.899 Exactly. But in computer science, this has a 00:14:53.899 --> 00:14:57.000 very specific, rigorous mathematical meaning. 00:14:57.340 --> 00:14:59.600 It does. And to understand it, we have to go 00:14:59.600 --> 00:15:01.659 back to that skepticism from the assembly cult. 00:15:01.840 --> 00:15:04.379 The wizards. The wizards. Remember, they truly 00:15:04.379 --> 00:15:07.039 believed that a machine could never write code 00:15:07.039 --> 00:15:10.159 as efficient as a human could. Fran Allen set 00:15:10.159 --> 00:15:12.639 out to prove that not only could a machine do 00:15:12.639 --> 00:15:15.059 it, but that a machine could actually do it better 00:15:15.059 --> 00:15:17.759 than a human. Because a machine can analyze patterns 00:15:17.759 --> 00:15:21.419 at a scale that a human brain just can't. Precisely. 00:15:21.419 --> 00:15:23.879 And for this work, she teamed up with another 00:15:23.879 --> 00:15:26.820 giant of computing, John Koch. He's a fascinating 00:15:26.820 --> 00:15:29.100 figure in his own right. Oh, a brilliant, eccentric 00:15:29.100 --> 00:15:32.399 character. He and Fran became this dynamic duo 00:15:32.399 --> 00:15:36.220 at IBM's Thomas J. Watson Research Center. Starting 00:15:36.220 --> 00:15:38.779 in the mid -1960s, they began publishing a series 00:15:38.779 --> 00:15:41.220 of papers that essentially invented the entire 00:15:41.220 --> 00:15:44.299 science of compiler optimization. Okay, so let's 00:15:44.299 --> 00:15:47.059 get into the how. How do you teach a computer 00:15:47.059 --> 00:15:50.360 to look at code and make it faster? Let's walk 00:15:50.360 --> 00:15:52.399 through some of the big techniques they cataloged. 00:15:52.419 --> 00:15:55.539 The first one I saw mentioned was procedure inlining. 00:15:56.029 --> 00:15:58.190 What is that? Okay, let's use an office analogy 00:15:58.190 --> 00:16:01.129 for this. Imagine you're a manager. You're the 00:16:01.129 --> 00:16:03.830 main program. Got it. And part of your job involves 00:16:03.830 --> 00:16:06.570 signing a document. But the signed document procedure 00:16:06.570 --> 00:16:08.950 is handled by your assistant who works in an 00:16:08.950 --> 00:16:11.750 office down the hall. Okay, so the standard non 00:16:11.750 --> 00:16:14.649 -optimized way to do this is I get a document. 00:16:15.049 --> 00:16:16.730 I have to get up from my desk, walk down the 00:16:16.730 --> 00:16:18.570 hall, hand it to my assistant, wait for them 00:16:18.570 --> 00:16:20.549 to sign it, take it back and walk back to my 00:16:20.549 --> 00:16:22.450 desk. Right. And all of that walking back and 00:16:22.450 --> 00:16:25.149 forth, that's overhead. In a computer, the processor 00:16:25.149 --> 00:16:27.690 has to jump to a different part of memory to 00:16:27.690 --> 00:16:30.429 find a subroutine and then jump back. That takes 00:16:30.429 --> 00:16:33.529 time. It might save memory space because you 00:16:33.529 --> 00:16:36.190 only write the signing instructions once, but 00:16:36.190 --> 00:16:39.149 it costs you in execution time. So what is procedure 00:16:39.149 --> 00:16:41.789 inlining then? Procedure inlining is when the 00:16:41.789 --> 00:16:44.330 compiler looks at the situation and says, you 00:16:44.330 --> 00:16:46.460 know what? This trip down the hall is happening 00:16:46.460 --> 00:16:48.659 a lot, and it's slowing things down. I'm going 00:16:48.659 --> 00:16:50.279 to teach the manager how to sign the document 00:16:50.279 --> 00:16:52.659 themselves. So it just copies the instructions? 00:16:52.899 --> 00:16:54.740 It takes the instructions from the assistant's 00:16:54.740 --> 00:16:57.519 office and essentially copy -pastes them directly 00:16:57.519 --> 00:17:00.259 into the manager's workflow. Now, when the task 00:17:00.259 --> 00:17:02.320 comes up, the manager just does it right there 00:17:02.320 --> 00:17:04.960 at their desk. No walking, no overhead. Boom. 00:17:05.220 --> 00:17:07.960 Instantly faster. Much faster. But there's a 00:17:07.960 --> 00:17:10.980 trade -off, and this is key. Now the manager's 00:17:10.980 --> 00:17:13.640 instruction book is bigger. The final program 00:17:13.640 --> 00:17:16.660 takes up more memory because you've duplicated 00:17:16.660 --> 00:17:19.660 those instructions. And memory was super expensive 00:17:19.660 --> 00:17:23.059 back then. Incredibly expensive. So Alan's work 00:17:23.059 --> 00:17:25.319 was about creating the rules. She turned that 00:17:25.319 --> 00:17:27.799 tradeoff into a mathematical formula the computer 00:17:27.799 --> 00:17:31.359 could use to decide. When is it worth it? When 00:17:31.359 --> 00:17:34.140 is the speed gain from inlining more valuable 00:17:34.140 --> 00:17:36.480 than the memory cost? That makes total sense. 00:17:36.759 --> 00:17:40.450 Okay, let's do another one. Loop unrolling. This 00:17:40.450 --> 00:17:42.210 one just sounds like a yoga pose or something. 00:17:42.450 --> 00:17:45.750 It's definitely less relaxing than yoga. So computers 00:17:45.750 --> 00:17:49.890 love loops. Do this specific task 1 ,000 times. 00:17:49.990 --> 00:17:51.849 That's what they're best at. Right. Let's stick 00:17:51.849 --> 00:17:54.190 with the office. Imagine you're an intern and 00:17:54.190 --> 00:17:57.869 your job is to stamp 100 envelopes. A standard 00:17:57.869 --> 00:18:00.609 computer loop would look like this. One, pick 00:18:00.609 --> 00:18:03.789 up an envelope. Two, stamp it. Three, put it 00:18:03.789 --> 00:18:07.369 down. Four, ask yourself, am I done with all 00:18:07.369 --> 00:18:10.730 100 envelopes yet? Four, ask yourself. Five, 00:18:10.990 --> 00:18:13.950 if the answer is no, go back to step one. Okay, 00:18:14.049 --> 00:18:15.630 that seems straightforward. But look at step 00:18:15.630 --> 00:18:17.809 four. You are asking yourself that question, 00:18:18.089 --> 00:18:21.039 am I done yet? 100 separate times. Each time 00:18:21.039 --> 00:18:23.559 you ask, it's a little mental stutter. It interrupts 00:18:23.559 --> 00:18:25.599 the flow of the work. And in a computer, that 00:18:25.599 --> 00:18:27.640 question isn't free. That check takes processing 00:18:27.640 --> 00:18:30.380 power. Exactly. That conditional check and the 00:18:30.380 --> 00:18:32.119 jump back to the beginning of the loop consume 00:18:32.119 --> 00:18:34.759 CPU cycles. Loop unrolling is when the compiler 00:18:34.759 --> 00:18:36.720 looks at that and says, hey, I know for a fact 00:18:36.720 --> 00:18:38.680 you have to do at least, say, four of these. 00:18:38.720 --> 00:18:40.940 So here's a new, better plan. It batches the 00:18:40.940 --> 00:18:43.240 work. It batches the work. The new instructions 00:18:43.240 --> 00:18:46.750 are, one, stamp an envelope. Two. Stamp an envelope. 00:18:46.910 --> 00:18:49.369 Three. Stamp an envelope. Four. Stamp an envelope. 00:18:49.569 --> 00:18:53.049 Five. Now ask yourself, am I done with the whole 00:18:53.049 --> 00:18:55.269 stack yet? Ah, so you've just cut the number 00:18:55.269 --> 00:18:57.450 of times you have to ask that question by 75%. 00:18:57.450 --> 00:19:00.470 Precisely. You reduce the loop overhead. You 00:19:00.470 --> 00:19:02.829 spend more of the computer's time doing the actual 00:19:02.829 --> 00:19:05.710 work stamping and less time managing the work. 00:19:05.849 --> 00:19:08.069 It sounds almost trivial to us with our super 00:19:08.069 --> 00:19:11.039 fast computers today. But back then. Back then, 00:19:11.140 --> 00:19:14.400 on a processor from the 1960s, eliminating all 00:19:14.400 --> 00:19:17.200 those jump and compare instructions could speed 00:19:17.200 --> 00:19:20.480 up a critical part of a program by 20, 30, even 00:19:20.480 --> 00:19:23.279 40%. That's a massive gain. It's the difference 00:19:23.279 --> 00:19:25.839 between buying in bulk at a big box store versus 00:19:25.839 --> 00:19:28.240 running to the corner store for one item every 00:19:28.240 --> 00:19:30.480 single day. That's a perfect analogy. And then 00:19:30.480 --> 00:19:32.599 there's one called CodeMotion, which is my personal 00:19:32.599 --> 00:19:34.440 favorite because it really highlights how ridiculously 00:19:34.440 --> 00:19:37.569 literal and... Frankly, dumb a computer can be 00:19:37.569 --> 00:19:40.690 if you don't help it. Okay, I'm intrigued. Tell 00:19:40.690 --> 00:19:43.210 me about code motion. Alright, new analogy. Imagine 00:19:43.210 --> 00:19:45.750 you are painting a large fence and you're working 00:19:45.750 --> 00:19:48.150 your way down the panels inside a loop. Your 00:19:48.150 --> 00:19:51.450 instructions are 1. Dip your brush in the paint 00:19:51.450 --> 00:19:56.170 can. 2. Paint one fence panel. 3. Open the gate 00:19:56.170 --> 00:19:59.349 so the dog can go out. 4. Go back to step one 00:19:59.349 --> 00:20:02.839 for the next panel. Wait, but... The gate is 00:20:02.839 --> 00:20:04.839 already open after the first time. Why would 00:20:04.839 --> 00:20:06.980 I keep opening it for every single panel? Because 00:20:06.980 --> 00:20:09.440 the instruction is inside the loop. A naive computer 00:20:09.440 --> 00:20:12.019 just follows its orders blindly. It gets to step 00:20:12.019 --> 00:20:13.980 three and thinks, my instruction is open the 00:20:13.980 --> 00:20:16.539 gate. I will now perform the open gate action. 00:20:17.069 --> 00:20:19.450 It has no idea that the state of the gate hasn't 00:20:19.450 --> 00:20:21.670 changed. It just does what it's told over and 00:20:21.670 --> 00:20:24.009 over. That is just maddeningly inefficient. It 00:20:24.009 --> 00:20:26.309 is. So CodeMotion is the compiler being smart 00:20:26.309 --> 00:20:28.190 enough to analyze that loop and say, hey, wait 00:20:28.190 --> 00:20:30.309 a second. The action open the gate has nothing 00:20:30.309 --> 00:20:32.150 to do with which fence panel we're currently 00:20:32.150 --> 00:20:34.329 painting. It's an invariant. It doesn't change 00:20:34.329 --> 00:20:36.369 inside the loop. So you just move it. You move 00:20:36.369 --> 00:20:38.710 it. The compiler rewrites the program to be, 00:20:38.829 --> 00:20:42.329 one, open the gate once. Two, now start the loop. 00:20:42.490 --> 00:20:46.170 Dip brush, paint panel, repeat. It's so simple 00:20:46.170 --> 00:20:48.750 and logical when you explain it, but someone 00:20:48.750 --> 00:20:51.470 had to figure out how to teach a machine to see 00:20:51.470 --> 00:20:53.970 that logic automatically. And that was the genius 00:20:53.970 --> 00:20:57.349 of their work. In 1971, Fran Allen and John Koch 00:20:57.349 --> 00:21:01.150 published a paper called A Catalog of Optimizing 00:21:01.150 --> 00:21:05.140 Transformations. It was... It was basically the 00:21:05.140 --> 00:21:07.400 Bible for all of this. The definitive guide. 00:21:07.619 --> 00:21:10.079 It took what had always been an art form, this 00:21:10.079 --> 00:21:13.059 wizard -like intuition of, oh, I think I can 00:21:13.059 --> 00:21:15.640 tweak this little bit of assembly code here. 00:21:15.799 --> 00:21:18.700 And they turned it into a rigorous, repeatable 00:21:18.700 --> 00:21:21.720 science that could be fully automated. They proved 00:21:21.720 --> 00:21:24.539 the wizards wrong. They proved that a well -designed 00:21:24.539 --> 00:21:27.059 compiler could analyze the whole program and 00:21:27.059 --> 00:21:29.900 generate machine code that was not just as good 00:21:29.900 --> 00:21:32.200 as, but often much, much better than what the 00:21:32.200 --> 00:21:34.319 average human programmer could write. Because 00:21:34.319 --> 00:21:36.799 the compiler never gets tired. It never gets 00:21:36.799 --> 00:21:39.079 lazy and forgets to move the gate instruction. 00:21:39.400 --> 00:21:40.859 And it can see the whole picture at once. It 00:21:40.859 --> 00:21:43.740 has a global view of the entire program, which 00:21:43.740 --> 00:21:45.400 was another one of Allen's huge contributions, 00:21:45.759 --> 00:21:48.880 the idea of data flow analysis. Which is what, 00:21:48.960 --> 00:21:52.119 exactly. It's about systematically tracking how 00:21:52.359 --> 00:21:54.500 Every piece of data moves through the entire 00:21:54.500 --> 00:21:58.299 program. Where is a variable created? Where is 00:21:58.299 --> 00:22:01.380 it used? Where is it changed? By mapping this 00:22:01.380 --> 00:22:03.920 out, the compiler can make incredibly smart decisions 00:22:03.920 --> 00:22:06.400 that a human looking at only one function at 00:22:06.400 --> 00:22:09.160 a time would miss. So she's basically conquered 00:22:09.160 --> 00:22:11.119 the single processor. She's written the Bible 00:22:11.119 --> 00:22:13.900 of optimization. But technology never stands 00:22:13.900 --> 00:22:16.819 still. Never. We get into the 1980s and the hardware 00:22:16.819 --> 00:22:18.819 starts to change in a fundamental way. Right. 00:22:18.940 --> 00:22:21.380 We stop trying to make one single brain faster 00:22:21.380 --> 00:22:23.819 and bigger and we start trying to use many brains 00:22:23.819 --> 00:22:26.599 at once. The shift to parallel computing. And 00:22:26.599 --> 00:22:28.579 this is where the problems get an order of magnitude 00:22:28.579 --> 00:22:30.900 more difficult. Why is that? I mean, intuitively. 00:22:31.640 --> 00:22:34.099 Two heads are better than one, right? Four processors 00:22:34.099 --> 00:22:36.700 should be four times as fast. In life, yes. In 00:22:36.700 --> 00:22:38.920 computing, it's a logistical nightmare. For its 00:22:38.920 --> 00:22:41.460 entire history up until the 80s, programming 00:22:41.460 --> 00:22:43.920 was sequential. It was a story. You do step one, 00:22:44.000 --> 00:22:45.839 then you do step two, then step three. It's a 00:22:45.839 --> 00:22:48.039 single predictable timeline. A straight line. 00:22:48.259 --> 00:22:50.799 A straight line. Parallel computing shatters 00:22:50.799 --> 00:22:53.759 that timeline. Now you have, say, four processors. 00:22:54.240 --> 00:22:56.099 You want to give each of them a piece of the 00:22:56.099 --> 00:22:59.019 work to do at the same time. But what happens 00:22:59.019 --> 00:23:01.819 if processor A needs the result of a calculation 00:23:01.819 --> 00:23:03.880 that processor B is still working on? You get 00:23:03.880 --> 00:23:06.079 a traffic jam. The whole thing has to stop and 00:23:06.079 --> 00:23:08.359 wait. Or worse, you get something called a race 00:23:08.359 --> 00:23:11.160 condition, where the final result depends on 00:23:11.160 --> 00:23:14.079 which processor gets there first. Or a deadlock, 00:23:14.119 --> 00:23:16.599 where processor A is waiting for B and B is waiting 00:23:16.599 --> 00:23:19.480 for A, and the entire system just freezes forever. 00:23:19.859 --> 00:23:24.099 So writing parallel code by hand is just Brain 00:23:24.099 --> 00:23:25.819 -breakingly difficult. It is because the human 00:23:25.819 --> 00:23:28.200 brain is really, really bad at keeping track 00:23:28.200 --> 00:23:31.460 of four or eight or a thousand simultaneous timelines. 00:23:31.799 --> 00:23:33.819 So what was Fran's solution? What was her approach 00:23:33.819 --> 00:23:36.279 to this new challenge? Well, her answer was the 00:23:36.279 --> 00:23:39.259 same as before. Automation. She believed the 00:23:39.259 --> 00:23:41.180 programmer shouldn't have to deal with this complexity. 00:23:41.400 --> 00:23:43.960 The compiler should. Let the machine manage the 00:23:43.960 --> 00:23:46.720 mess. Exactly. She led a massive project at IBM 00:23:46.720 --> 00:23:49.440 called PT -RAM, which stood for Parallel Translator. 00:23:49.680 --> 00:23:54.180 And the goal was... unbelievably audacious for 00:23:54.180 --> 00:23:56.440 the time. What was it? The goal was to take a 00:23:56.440 --> 00:23:59.000 normal sequential program written by a regular 00:23:59.000 --> 00:24:01.400 human who thinks in a straight line and have 00:24:01.400 --> 00:24:04.079 the PTR and compiler automatically figure out 00:24:04.079 --> 00:24:06.440 how to break it apart and run it safely and efficiently 00:24:06.440 --> 00:24:09.019 on a parallel machine. She wanted the software 00:24:09.019 --> 00:24:11.700 to become the world's best project manager. That's 00:24:11.700 --> 00:24:14.269 a great way to put it. And to do this, her team 00:24:14.269 --> 00:24:16.430 had to invent a new way of looking at programs. 00:24:16.809 --> 00:24:19.390 They created a concept called the program dependence 00:24:19.390 --> 00:24:21.950 graph. Okay, that sounds important. Walk us through 00:24:21.950 --> 00:24:24.450 it. Think of a complex recipe for making dinner. 00:24:24.609 --> 00:24:27.250 A normal, sequential recipe looks like a list. 00:24:27.630 --> 00:24:33.170 1. Chop onions. 2. Heat pan on stove. 3. Saute 00:24:33.170 --> 00:24:37.880 onions in the pan. 4. Toss the salad. 5. bake 00:24:37.880 --> 00:24:40.119 the chicken right a sequential cook does step 00:24:40.119 --> 00:24:42.880 one then two then three and so on the program 00:24:42.880 --> 00:24:45.380 dependence graph isn't a list it's a mat it draws 00:24:45.380 --> 00:24:47.579 lines between the tasks based on what depends 00:24:47.579 --> 00:24:50.160 on what so saute onions depends on chopped onions 00:24:50.160 --> 00:24:51.920 you can't do it without the chopped onions exactly 00:24:51.920 --> 00:24:54.539 and saute onions also depends on heat pan you 00:24:54.539 --> 00:24:56.619 need the hot pan there are dependencies but what 00:24:56.619 --> 00:24:59.839 about toss the salad that's an island It doesn't 00:24:59.839 --> 00:25:01.900 depend on the onions or the pan or the chicken. 00:25:02.059 --> 00:25:04.859 It's totally independent. Precisely. And Bake 00:25:04.859 --> 00:25:07.819 the Chicken is also mostly independent, at least 00:25:07.819 --> 00:25:10.680 at the beginning. The compiler builds this dependency 00:25:10.680 --> 00:25:13.220 map, and then it can see the opportunities for 00:25:13.220 --> 00:25:15.599 parallelism. It can see that you can have one 00:25:15.599 --> 00:25:18.140 processor chopping onions while another processor 00:25:18.140 --> 00:25:20.500 is pausing the salad at the exact same time. 00:25:20.660 --> 00:25:22.940 It untangles the spaghetti before it even gets 00:25:22.940 --> 00:25:25.740 cooked. It can assign independent tasks to different 00:25:25.740 --> 00:25:28.690 processors. This concept, the program dependence 00:25:28.690 --> 00:25:31.390 graph, is still the primary structuring method 00:25:31.390 --> 00:25:34.609 used by most parallelizing compilers today. It 00:25:34.609 --> 00:25:36.869 was a monumental breakthrough. She was basically 00:25:36.869 --> 00:25:39.690 building the air traffic control system for the 00:25:39.690 --> 00:25:42.130 information superhighway. That's a very 90s metaphor, 00:25:42.230 --> 00:25:44.069 but you know what? It actually fits perfectly. 00:25:44.309 --> 00:25:47.450 I'm bringing it back. So we've established this, 00:25:47.529 --> 00:25:51.150 this titan of technical achievement, a true giant. 00:25:51.390 --> 00:25:53.890 But we can't tell her full story without talking 00:25:53.890 --> 00:25:57.039 about the environment she did all this in. IBM. 00:25:57.039 --> 00:26:00.579 The tech world. It was a boys club. It absolutely 00:26:00.579 --> 00:26:03.660 was. But here's a statistic that's going to blow 00:26:03.660 --> 00:26:07.980 your mind. In the 1970s and 80s, in the compiler 00:26:07.980 --> 00:26:10.339 research group that Fran Allen managed at IBM, 00:26:10.559 --> 00:26:12.940 do you want to take a guess at the gender breakdown? 00:26:13.099 --> 00:26:15.680 Oh, man. Based on everything I know about that 00:26:15.680 --> 00:26:19.599 era, I'm going to guess it was like 95 % men. 00:26:20.059 --> 00:26:22.119 Maybe one or two other women. It was nearly 50 00:26:22.119 --> 00:26:25.920 % women. Get out of here. 50 % in the 1970s. 00:26:26.079 --> 00:26:29.099 50%. One to one. That is. That's better than 00:26:29.099 --> 00:26:31.960 most Silicon Valley companies are today in 2026. 00:26:32.180 --> 00:26:34.500 How on earth did that happen? It happened because 00:26:34.500 --> 00:26:37.380 Fran Allen made it happen. She didn't just sit 00:26:37.380 --> 00:26:40.220 back and accept the pipeline problem as an excuse. 00:26:40.559 --> 00:26:43.960 She actively sought out, recruited, and mentored 00:26:43.960 --> 00:26:46.339 women. She created an environment where they 00:26:46.339 --> 00:26:49.019 could not just survive, but thrive. She built 00:26:49.019 --> 00:26:51.059 the culture. She built the culture. A colleague 00:26:51.059 --> 00:26:53.299 of hers, Barbara Simons, who later became president 00:26:53.299 --> 00:26:56.880 of the ACM, called Fran a strong feminist who 00:26:56.880 --> 00:26:58.920 understood that men mentorship wasn't just about 00:26:58.920 --> 00:27:01.579 being nice. It was about strategic advocacy. 00:27:01.880 --> 00:27:03.740 It's about holding the door open for the people 00:27:03.740 --> 00:27:05.819 coming behind you. And she did that explicitly. 00:27:06.119 --> 00:27:07.900 There's a great story. She took a sabbatical 00:27:07.900 --> 00:27:11.359 to teach at Stanford in 1977. And in one of her 00:27:11.359 --> 00:27:14.180 graduate classes was a student named Anita Borg. 00:27:14.400 --> 00:27:17.619 Anita Borg, as in the founder of the Grace Hopper 00:27:17.619 --> 00:27:20.279 Celebration of Women in Computing, the AnitaB 00:27:20.279 --> 00:27:24.079 .org Institute, that Anita Borg. The very same, 00:27:24.180 --> 00:27:26.619 a legendary figure in her own right. And Anita 00:27:26.619 --> 00:27:29.180 Borg later said that during her entire Ph .D. 00:27:29.180 --> 00:27:32.039 program at a major university, Fran Allen was 00:27:32.039 --> 00:27:34.440 the only female professor she ever had. The only 00:27:34.440 --> 00:27:36.839 one. Wow. Imagine the impact of that. You're 00:27:36.839 --> 00:27:38.980 sitting in a classroom, probably surrounded by 00:27:38.980 --> 00:27:40.900 men, maybe feeling like an imposter, questioning 00:27:40.900 --> 00:27:43.200 if you even belong there. And you look up at 00:27:43.200 --> 00:27:44.900 the podium and the person teaching you is the 00:27:44.900 --> 00:27:47.380 world's foremost expert on compiler optimization. 00:27:48.279 --> 00:27:50.119 And she looks like you. It just changes your 00:27:50.119 --> 00:27:52.200 whole perception of what's possible. It moves 00:27:52.200 --> 00:27:54.500 the goalposts from impossible to that could be 00:27:54.500 --> 00:27:57.799 me. It's everything. And IBM did eventually recognize 00:27:57.799 --> 00:28:00.420 her leadership. She became the first female IBM 00:28:00.420 --> 00:28:03.640 fellow in 1989. Which is a huge deal at IBM. 00:28:03.779 --> 00:28:06.400 It's their highest technical honor. It's reserved 00:28:06.400 --> 00:28:08.660 for a tiny fraction of their most brilliant researchers. 00:28:09.059 --> 00:28:11.440 She was also president of the IBM Academy of 00:28:11.440 --> 00:28:13.960 Technology. And they later created mentorship 00:28:13.960 --> 00:28:17.970 awards in her name. But the biggest award. The 00:28:17.970 --> 00:28:20.529 one that cemented her place in the Pantheon came 00:28:20.529 --> 00:28:24.410 much later. The Turing Award. The Nobel of Computing. 00:28:24.430 --> 00:28:28.170 It's 2006. Fran has been retired from IBM for 00:28:28.170 --> 00:28:30.950 a few years. And the Association for Computing 00:28:30.950 --> 00:28:33.430 Machinery makes the announcement. She becomes 00:28:33.430 --> 00:28:35.869 the first woman in the award's 40 -year history 00:28:35.869 --> 00:28:39.130 to receive it. It took 40 years. 40 years. It's 00:28:39.130 --> 00:28:41.109 a pretty sobering statistic when you think about 00:28:41.109 --> 00:28:43.369 it. And frankly, many people felt it was long 00:28:43.369 --> 00:28:45.769 overdue. Especially when you consider that her 00:28:45.769 --> 00:28:48.809 collaborator, John Koch had won the same award. 00:28:48.950 --> 00:28:51.609 In 1987. Wait a minute. They did the foundational 00:28:51.609 --> 00:28:53.869 work together. The seminal papers were credited 00:28:53.869 --> 00:28:57.569 to Allen and Koch. And he won the award 19 years 00:28:57.569 --> 00:29:01.509 before she did. 19 years earlier, yes. That stings. 00:29:01.509 --> 00:29:03.970 That really says something. It does. It reflects 00:29:03.970 --> 00:29:06.930 the biases of the era. It reflects who was often 00:29:06.930 --> 00:29:10.049 seen as the lead genius and who was perhaps perceived 00:29:10.049 --> 00:29:13.170 as the helper or the implementer, even when that 00:29:13.170 --> 00:29:16.140 wasn't the case. But to her credit. She seems 00:29:16.140 --> 00:29:18.720 to have handled it with incredible grace. She 00:29:18.720 --> 00:29:21.180 never expressed any public bitterness about it. 00:29:21.240 --> 00:29:24.099 When she finally won, she used her acceptance 00:29:24.099 --> 00:29:26.579 speech and the platform it gave her to talk about. 00:29:27.559 --> 00:29:30.220 Well, you can guess. More opportunities for women 00:29:30.220 --> 00:29:33.019 in science? Exactly. She said she hoped the award 00:29:33.019 --> 00:29:34.980 would create more opportunities for women in 00:29:34.980 --> 00:29:37.519 science, computing, and engineering. She stayed 00:29:37.519 --> 00:29:40.180 on message, always. So before we wrap this all 00:29:40.180 --> 00:29:42.750 up, I want to pivot to a part of her life that 00:29:42.750 --> 00:29:44.930 seems completely disconnected from computers. 00:29:45.109 --> 00:29:47.430 Right. We've painted this picture of her as this 00:29:47.430 --> 00:29:50.509 intense lab researcher staring at code, living 00:29:50.509 --> 00:29:53.230 in this abstract world of logic. But her weekends 00:29:53.230 --> 00:29:56.710 were different. Oh, she was not staring at a 00:29:56.710 --> 00:29:58.569 screen on the weekend. She was staring at glaciers 00:29:58.569 --> 00:30:01.650 and rock faces. Fran Allen was a serious, hardcore 00:30:01.650 --> 00:30:03.390 mountain climber. And we're not talking about... 00:30:03.519 --> 00:30:05.519 you know, casual day hikes in the local hills. 00:30:05.640 --> 00:30:08.599 No, no. We are talking about expeditions to the 00:30:08.599 --> 00:30:11.160 high Arctic. She was an active member of the 00:30:11.160 --> 00:30:13.960 American Alpine Club. She was part of an expedition 00:30:13.960 --> 00:30:16.420 to Ellesmere Island. Ellesmere Island. I mean, 00:30:16.420 --> 00:30:18.039 for people who don't know, that is the most northern 00:30:18.039 --> 00:30:20.920 part of Canada. It is way up there. That is polar 00:30:20.920 --> 00:30:23.000 bear country. It is one of the most desolate, 00:30:23.140 --> 00:30:26.579 remote, and dangerous places on the planet. And 00:30:26.579 --> 00:30:28.720 she was on a team that established a completely 00:30:28.720 --> 00:30:32.089 new route across a section of the island. That's 00:30:32.089 --> 00:30:34.009 incredible. And I feel like there's a perfect 00:30:34.009 --> 00:30:36.349 metaphor in there for her work. I think so, too. 00:30:36.470 --> 00:30:39.009 How do you see it? Well, think about what a technical 00:30:39.009 --> 00:30:41.250 climb is. You're standing at the bottom of a 00:30:41.250 --> 00:30:43.670 mountain, looking up at this chaotic, dangerous, 00:30:44.049 --> 00:30:47.730 seemingly impassable landscape. A rock face, 00:30:48.049 --> 00:30:51.730 a glacier field. And your job is to find a path 00:30:51.730 --> 00:30:54.430 through it. Not just any path, but the optimal 00:30:54.430 --> 00:30:56.569 path. the one that conserves the most energy, 00:30:56.750 --> 00:30:59.230 the one that avoids the biggest dangers, the 00:30:59.230 --> 00:31:01.329 most efficient route to the top. It's optimization. 00:31:01.690 --> 00:31:04.009 It's the exact same problem. It's exactly the 00:31:04.009 --> 00:31:06.309 same problem. Whether she was looking at a messy 00:31:06.309 --> 00:31:09.910 loop of Fortran code or a field of glacial crevasses, 00:31:10.150 --> 00:31:13.509 her mind was doing the same thing. She was identifying 00:31:13.509 --> 00:31:16.930 dependencies, finding invariance, and charting 00:31:16.930 --> 00:31:19.369 the most efficient path. She was bringing order 00:31:19.369 --> 00:31:22.779 to chaos. The great explorer of complex systems, 00:31:22.859 --> 00:31:25.579 both digital and natural. That's beautiful. She 00:31:25.579 --> 00:31:28.279 kept exploring right up until the end. She passed 00:31:28.279 --> 00:31:32.160 away in 2020 on her 88th birthday. And the outpouring 00:31:32.160 --> 00:31:34.480 of respect from the entire computing community 00:31:34.480 --> 00:31:38.359 was universal. The IE, a huge professional organization, 00:31:38.880 --> 00:31:42.220 created the Frances E. Allen Medal in her honor. 00:31:42.339 --> 00:31:43.900 And it's only the second medal in their history 00:31:43.900 --> 00:31:45.799 to be named after a woman. The other one being 00:31:45.799 --> 00:31:48.019 the Ada Lovelace Award. She's in very good company. 00:31:48.220 --> 00:31:49.839 So let's bring it all home. We've gone from the 00:31:49.839 --> 00:31:52.339 farm to the source code, from the spies to the 00:31:52.339 --> 00:31:54.819 parallel algorithms. What's the big takeaway? 00:31:54.960 --> 00:31:57.150 Why should you... listening to this right now, 00:31:57.269 --> 00:32:00.069 care about Fran Allen? You should care because 00:32:00.069 --> 00:32:02.509 the modern world you're living in is built on 00:32:02.509 --> 00:32:05.410 the foundation of her work. Every single time 00:32:05.410 --> 00:32:08.109 you open an app on your phone and it loads instantly 00:32:08.109 --> 00:32:10.990 instead of taking 10 seconds. You can thank the 00:32:10.990 --> 00:32:13.650 principles of optimization she developed. Every 00:32:13.650 --> 00:32:16.309 time you check a weather forecast and a supercomputer 00:32:16.309 --> 00:32:19.390 model running on thousands of processors has 00:32:19.390 --> 00:32:21.730 predicted the rain. That's the legacy of her 00:32:21.730 --> 00:32:24.769 work in parallel computing. She is the invisible 00:32:24.769 --> 00:32:27.349 architect. She's the reason our digital world 00:32:27.349 --> 00:32:30.269 is fast and responsive. She made computing viable. 00:32:30.859 --> 00:32:34.339 She proved that we as humans didn't need to contort 00:32:34.339 --> 00:32:37.079 our brains to speak the machine's language. We 00:32:37.079 --> 00:32:39.240 could instead teach the machine to understand 00:32:39.240 --> 00:32:42.160 ours and to do it efficiently. She democratized 00:32:42.160 --> 00:32:44.220 high performance computing. She made it possible 00:32:44.220 --> 00:32:46.359 for people who weren't assembly code wizards 00:32:46.359 --> 00:32:49.160 to write software that could solve huge problems. 00:32:49.319 --> 00:32:51.759 And on a more philosophical level, she reminds 00:32:51.759 --> 00:32:54.289 us of the profound value of looking. Under the 00:32:54.289 --> 00:32:56.769 hood. That is the single biggest lesson from 00:32:56.769 --> 00:32:59.710 her life. She was handed a tool, Fortran, and 00:32:59.710 --> 00:33:01.569 she wasn't content to just use it. She had to 00:33:01.569 --> 00:33:03.730 open it up. She had to read the source. And that 00:33:03.730 --> 00:33:05.930 leads me to one last slightly provocative thought 00:33:05.930 --> 00:33:08.470 for you to chew on. Okay. We are living in a 00:33:08.470 --> 00:33:12.250 moment right now that feels a lot like 1957 all 00:33:12.250 --> 00:33:15.829 over again. We have a powerful new tool that 00:33:15.829 --> 00:33:19.349 a lot of people are skeptical of. AI. Specifically, 00:33:19.529 --> 00:33:24.049 AI that can write code. GitHub Copilot. ChatGPT. 00:33:24.170 --> 00:33:27.009 It's the ultimate compiler in a way. Yeah. It 00:33:27.009 --> 00:33:30.109 translates plain English human intent directly 00:33:30.109 --> 00:33:33.069 into functional code. Exactly. Fran Allen spent 00:33:33.069 --> 00:33:35.910 her entire career teaching computers how to optimize 00:33:35.910 --> 00:33:38.990 code that humans wrote. Now the computers are 00:33:38.990 --> 00:33:40.609 starting to create the code from scratch. In 00:33:40.609 --> 00:33:42.390 a way, it's the ultimate fulfillment of her vision. 00:33:42.430 --> 00:33:44.230 The machine is doing all the heavy lifting. It 00:33:44.230 --> 00:33:47.180 is. And here's the question. Fran Allen became 00:33:47.180 --> 00:33:49.500 a genius precisely because she was forced to 00:33:49.500 --> 00:33:51.059 get her hands dirty. She had to read the raw 00:33:51.059 --> 00:33:53.000 assembly. She had to understand the nuts and 00:33:53.000 --> 00:33:55.859 bolts at the deepest possible level. If the AI 00:33:55.859 --> 00:33:58.759 does all of that for us, if the car dries itself 00:33:58.759 --> 00:34:01.720 perfectly every time, do we lose something critical? 00:34:01.960 --> 00:34:04.299 Are we raising a new generation of programmers 00:34:04.299 --> 00:34:07.059 who will never know how to read the source? That 00:34:07.059 --> 00:34:10.659 is the fundamental risk. Convenience can sometimes 00:34:10.659 --> 00:34:13.750 be the enemy of deep understanding. If we stop 00:34:13.750 --> 00:34:15.670 looking under the hood because the engine appears 00:34:15.670 --> 00:34:18.210 to never break down, we might completely forget 00:34:18.210 --> 00:34:20.349 how the engine works. And when it finally does 00:34:20.349 --> 00:34:24.690 break, in some new and unexpected way. We might 00:34:24.690 --> 00:34:26.610 not have a Fran Allen around who knows how to 00:34:26.610 --> 00:34:28.929 fix it. A little bit of existential dread to 00:34:28.929 --> 00:34:31.309 end our deep dive. Just a little, to keep you 00:34:31.309 --> 00:34:33.590 sharp. That's it for this one. Go read some source 00:34:33.590 --> 00:34:37.690 code, or maybe climb a mountain, or both. Thanks 00:34:37.690 --> 00:34:38.610 for listening. Take care.