WEBVTT 00:00:00.000 --> 00:00:01.919 I was sitting in my car before coming in just, 00:00:01.919 --> 00:00:03.960 you know, thinking about memory, not our kind 00:00:03.960 --> 00:00:06.820 where you forget your keys, but digital memory. 00:00:07.299 --> 00:00:09.800 It's so fragile. You close a browser tab and 00:00:09.800 --> 00:00:12.599 poof, it's gone unless you explicitly saved it 00:00:12.599 --> 00:00:15.419 somewhere. It just vaporizes. But then I was 00:00:15.419 --> 00:00:17.539 reading about this time capsule test in the research 00:00:17.539 --> 00:00:20.800 for today. An AI builds this whole simulated 00:00:20.800 --> 00:00:23.839 computer inside a browser and it remembers everything. 00:00:24.000 --> 00:00:26.780 A half finished calculation, a sticky note, even 00:00:26.780 --> 00:00:29.079 after the whole session was, you know, nuked. 00:00:29.160 --> 00:00:30.879 And that's the thing that gets me. It wasn't 00:00:30.879 --> 00:00:33.539 hard -coded. A human didn't tell it, hey, save 00:00:33.539 --> 00:00:36.119 this variable here. The mall just figured out 00:00:36.119 --> 00:00:37.880 how to save its own state. It understood that's 00:00:37.880 --> 00:00:39.640 what a computer's supposed to do. It basically 00:00:39.640 --> 00:00:43.960 invented its own long -term memory just to survive 00:00:43.960 --> 00:00:47.240 a reboot. It's on best. They can. Bordering on 00:00:47.240 --> 00:00:50.280 a survival instinct. Welcome back to the deep 00:00:50.280 --> 00:00:52.740 dive. Today, we are unpacking something that's 00:00:52.740 --> 00:00:55.439 been making a lot of noise and, frankly, creating 00:00:55.439 --> 00:00:58.979 a bit of anxiety. in developer circles. We're 00:00:58.979 --> 00:01:01.960 looking at GPT 5 .3 codecs, and I want to set 00:01:01.960 --> 00:01:03.840 the tone right away. We're not here for the hype 00:01:03.840 --> 00:01:07.040 cycle. We need to be calm, measured, and just 00:01:07.040 --> 00:01:10.159 figure out if this is another incremental update 00:01:10.159 --> 00:01:12.400 or if this actually changes how software gets 00:01:12.400 --> 00:01:15.480 built. It's the right question to ask. And the 00:01:15.480 --> 00:01:18.120 context here is everything. GBT 5 .3 dropped 00:01:18.120 --> 00:01:21.200 on the exact same day as Opus 4 .6. And if you 00:01:21.200 --> 00:01:23.459 follow this space, you know Opus is the flashy 00:01:23.459 --> 00:01:25.400 one, the one that makes beautiful charts and 00:01:25.400 --> 00:01:28.239 writes poetry. But the rumors around 5 .3 are 00:01:28.239 --> 00:01:29.780 different. People are whispering that this model 00:01:29.780 --> 00:01:31.859 helped create itself. Now, that sounds like sci 00:01:31.859 --> 00:01:33.340 -fi marketing. I know. But when you dig into 00:01:33.340 --> 00:01:35.019 the performance metrics, things get a little 00:01:35.019 --> 00:01:37.120 weird. Weird how usually the numbers just go 00:01:37.120 --> 00:01:39.420 up. Weird because it seems to be breaking that 00:01:39.420 --> 00:01:41.760 cardinal rule of AI development from the last 00:01:41.760 --> 00:01:44.200 decade, which has always been bigger is better. 00:01:45.390 --> 00:01:48.269 So our mission today is to cut through all that. 00:01:48.409 --> 00:01:50.269 We're not reading the press release. We're gonna 00:01:50.269 --> 00:01:52.569 walk through a gauntlet of what the source calls 00:01:52.569 --> 00:01:55.709 very hard tests. They were run by a reviewer 00:01:55.709 --> 00:01:57.730 who just locked themselves in a room, stopped 00:01:57.730 --> 00:01:59.849 reading the news, and just started coding with 00:01:59.849 --> 00:02:01.890 it. We're talking everything from browser -based 00:02:01.890 --> 00:02:05.090 operating systems to simulating 3D printing physics. 00:02:05.310 --> 00:02:08.250 I appreciate that approach. A hands -on, real 00:02:08.250 --> 00:02:10.530 -world test. So let's start with the architecture, 00:02:10.629 --> 00:02:12.169 because this is where the philosophy of the model 00:02:12.169 --> 00:02:14.490 really comes through. The research points to 00:02:14.490 --> 00:02:17.800 a more with less kind of approach. Usually a 00:02:17.800 --> 00:02:19.939 smarter model means a huge computational tax. 00:02:20.020 --> 00:02:22.199 It's a gas guzzler. This seems to be the opposite. 00:02:22.580 --> 00:02:24.699 Exactly. And there's this one data point in the 00:02:24.699 --> 00:02:26.539 source material that just stops you in your tracks. 00:02:27.400 --> 00:02:31.340 GPT 5 .3 uses fewer tokens to solve problems 00:02:31.340 --> 00:02:33.219 than the models that came before it. And for 00:02:33.219 --> 00:02:34.819 anyone listening who isn't deep in this stuff, 00:02:35.379 --> 00:02:38.180 a token is basically a word or part of a word. 00:02:38.580 --> 00:02:41.000 It's the AI's building block for language. Usually 00:02:41.000 --> 00:02:43.159 to get smarter, you just throw more tokens at 00:02:43.159 --> 00:02:45.680 the problem. You ramble until you find the answer. 00:02:45.840 --> 00:02:47.500 Right, the old scatter gun approach. Just keep 00:02:47.500 --> 00:02:50.469 guessing until something works. Precisely. But 00:02:50.469 --> 00:02:53.669 GPT 5 .3 is succinct. It uses fewer tokens, but 00:02:53.669 --> 00:02:55.849 gets much higher accuracy. There's this chart 00:02:55.849 --> 00:02:58.610 in the report. It shows the model hitting 77 00:02:58.610 --> 00:03:02.090 .3 % accuracy in these really complex terminal 00:03:02.090 --> 00:03:04.370 tests. And just to put that number in perspective, 00:03:04.610 --> 00:03:06.949 where was the last version, 5 .2? It was sitting 00:03:06.949 --> 00:03:11.370 at 64 .0%. Wow. That is a massive 13 % gap. And 00:03:11.370 --> 00:03:14.370 we can't just glaze over that number. In AI development, 00:03:14.789 --> 00:03:17.849 getting 1 % or 2 % is a huge win, a 13 % jump 00:03:17.849 --> 00:03:20.620 in one generation. is. It's almost unheard of. 00:03:20.939 --> 00:03:22.840 That gap is where all the tricky logic lives. 00:03:23.159 --> 00:03:24.879 It's the difference between an AI that gives 00:03:24.879 --> 00:03:26.819 up when the code gets tough and one that actually 00:03:26.819 --> 00:03:28.860 grinds through the problem. So it's not just 00:03:28.860 --> 00:03:30.979 regurgitating answers it saw on Stack Overflow 00:03:30.979 --> 00:03:32.919 during training? No, and that is the critical 00:03:32.919 --> 00:03:35.360 distinction here. The source talks about a big 00:03:35.360 --> 00:03:38.060 behavioral shift. Unlike these one -shot models 00:03:38.060 --> 00:03:40.439 where you ask a question, get an answer, and 00:03:40.439 --> 00:03:44.240 just hope it works, GBT 5 .3 loves to iterate. 00:03:44.479 --> 00:03:47.819 It plans, it builds a little piece, tests it, 00:03:48.020 --> 00:03:49.960 sees that it broke something, and then it fixes 00:03:49.960 --> 00:03:53.139 it. It's acting less like a search engine and 00:03:53.139 --> 00:03:55.479 more like a junior engineer who knows how to 00:03:55.479 --> 00:03:57.740 debug their own work. That brings me to a question 00:03:57.740 --> 00:04:02.280 then. In this context, does efficiency actually 00:04:02.280 --> 00:04:05.479 equal intelligence? Yes. Because it's solving 00:04:05.479 --> 00:04:08.180 harder logic with less noise. Think of it like 00:04:08.180 --> 00:04:11.050 a writer. A novice uses 50 words to describe 00:04:11.050 --> 00:04:13.409 a sunset. A master uses three, but they're the 00:04:13.409 --> 00:04:15.710 exact right three. That's what this model is 00:04:15.710 --> 00:04:17.910 doing, but with code, it's throwing reasoning 00:04:17.910 --> 00:04:20.110 at the problem, not just volume. That's a great 00:04:20.110 --> 00:04:22.350 analogy, the density of the thought. Let's move 00:04:22.350 --> 00:04:24.129 to that first big stress test in the source. 00:04:24.250 --> 00:04:26.709 This is the browser operating system test. The 00:04:26.709 --> 00:04:28.949 reviewer asked it to build a functional OS right 00:04:28.949 --> 00:04:31.829 inside a Chrome tab. Start menu, windows, apps, 00:04:31.889 --> 00:04:33.930 the whole thing. Right. And this is what we get 00:04:33.930 --> 00:04:36.250 to what the source calls the ugly truth of 5 00:04:36.250 --> 00:04:39.620 .3. The first attempt. Visually, it was a disaster. 00:04:39.779 --> 00:04:43.439 It had a 1999 aesthetic, square icons, horrible 00:04:43.439 --> 00:04:46.740 gray backgrounds, Times, new Roman font. If you 00:04:46.740 --> 00:04:49.259 showed it to a client, you would be fired immediately. 00:04:49.480 --> 00:04:51.800 I actually love that detail. It didn't try to 00:04:51.800 --> 00:04:53.959 impress with flashy visuals. It reminds me of 00:04:53.959 --> 00:04:55.600 some backend engineers I worked with. They build 00:04:55.600 --> 00:04:58.240 these incredible databases. But the UI looks 00:04:58.240 --> 00:05:01.079 like a spreadsheet from 1995. That is exactly 00:05:01.079 --> 00:05:03.639 the vibe. But underneath that ugly skin, the 00:05:03.639 --> 00:05:06.459 functional brilliance was just... terrifyingly 00:05:06.459 --> 00:05:08.879 good. The reviewer opens the calculator app in 00:05:08.879 --> 00:05:12.680 this fake OS, types 77 times 7, and it just spits 00:05:12.680 --> 00:05:15.620 out 539. It worked. Then they open the notes 00:05:15.620 --> 00:05:18.600 app, type hello world, close it, reopen it, the 00:05:18.600 --> 00:05:20.879 text is still there. That's that local storage 00:05:20.879 --> 00:05:23.319 bit we mentioned. Can we unpack why that's so 00:05:23.319 --> 00:05:25.899 hard? To a user, saving something seems so basic. 00:05:26.160 --> 00:05:29.319 It feels basic to us, yeah. But for an AI -generated 00:05:29.319 --> 00:05:31.779 simulation in a browser, it's really complex. 00:05:32.439 --> 00:05:34.699 The AI has to understand the concept of state. 00:05:35.079 --> 00:05:38.439 It has to write code that uses the browser's 00:05:38.439 --> 00:05:40.779 own storage to cache data, so when the virtual 00:05:40.779 --> 00:05:44.300 app closes, the data doesn't just vanish. Most 00:05:44.300 --> 00:05:46.740 AI models forget the context the second you change 00:05:46.740 --> 00:05:49.540 tasks. This one built its own persistence layer. 00:05:49.759 --> 00:05:51.500 And that leads right back to that time capsule 00:05:51.500 --> 00:05:53.769 feature. The user asked for a way to save the 00:05:53.769 --> 00:05:56.670 whole desktop state. Right. And the AI just coded 00:05:56.670 --> 00:06:00.149 a system to snapshot every open window, its exact 00:06:00.149 --> 00:06:02.370 coordinates on the screen, the data inside it, 00:06:02.730 --> 00:06:05.089 all so you could restore the session later. That 00:06:05.089 --> 00:06:07.850 means it understood the entire hierarchy of the 00:06:07.850 --> 00:06:10.110 application it just built. It wasn't just pasting 00:06:10.110 --> 00:06:12.290 code snippets. It understood the architecture 00:06:12.290 --> 00:06:14.649 of the machine it created. It built a save game 00:06:14.649 --> 00:06:16.769 feature for an operating system it just invented. 00:06:16.810 --> 00:06:18.889 Yeah. But the reviewer didn't just leave it ugly, 00:06:18.889 --> 00:06:21.459 right? No. And this is that iteration part we 00:06:21.459 --> 00:06:23.579 were talking about. The reviewer just said, the 00:06:23.579 --> 00:06:26.180 logic is great, but the design is very ugly. 00:06:26.259 --> 00:06:29.240 The model didn't argue. It didn't break. It just 00:06:29.240 --> 00:06:31.720 went back in and applied a dark mode, a sunset 00:06:31.720 --> 00:06:34.819 wallpaper, and a custom right click menu. So 00:06:34.819 --> 00:06:37.860 if you look at that sequence, the ugly but working 00:06:37.860 --> 00:06:43.040 draft, then the polish, does it imply that logic 00:06:43.040 --> 00:06:46.240 comes before beauty for this model? Precisely. 00:06:46.579 --> 00:06:49.339 It prioritizes function, then aesthetic. And 00:06:49.339 --> 00:06:51.620 if you're a developer, that is exactly how you 00:06:51.620 --> 00:06:53.879 want it to think. Make it run, then make it pretty. 00:06:54.439 --> 00:06:56.439 That function over form distinction seems to 00:06:56.439 --> 00:06:58.680 carry into the next test, which honestly, I found 00:06:58.680 --> 00:07:00.240 this the most mind -bending part of the whole 00:07:00.240 --> 00:07:03.319 review, simulating the physical world. The reviewer 00:07:03.319 --> 00:07:05.860 asked it to build a 3D printer simulation. Yeah, 00:07:05.920 --> 00:07:07.779 this was a real moment of wonder for me when 00:07:07.779 --> 00:07:09.439 I read this. And we're not talking about drawing 00:07:09.439 --> 00:07:12.480 a picture of a printer. The AI built a full core 00:07:12.480 --> 00:07:15.079 XY printer simulation. OK, for the non -engineers 00:07:15.079 --> 00:07:17.259 listening, and for me, what exactly is core XY? 00:07:17.529 --> 00:07:20.129 It's a specific kind of system in high -end 3D 00:07:20.129 --> 00:07:22.350 printers. Instead of one motor for X and one 00:07:22.350 --> 00:07:25.209 for Y, you have two motors working together with 00:07:25.209 --> 00:07:28.829 belts to move the print head. It involves some 00:07:28.829 --> 00:07:31.410 pretty complex trigonometry. If you get the math 00:07:31.410 --> 00:07:33.389 wrong, the print head just crashes into the wall. 00:07:33.660 --> 00:07:35.740 And the source said it wasn't just animating 00:07:35.740 --> 00:07:38.420 a box moving around, it was actually calculating 00:07:38.420 --> 00:07:41.259 motor positions. A equals zero, B equals zero. 00:07:41.399 --> 00:07:43.660 It was simulating the stepper motors step by 00:07:43.660 --> 00:07:46.939 step, but the Benchy test is where it just gets 00:07:46.939 --> 00:07:49.800 wild. The Benchy. That's the little tugboat everyone 00:07:49.800 --> 00:07:52.290 prints to test their machines, right? the hello 00:07:52.290 --> 00:07:55.790 world of 3D printing. Exactly. So the reviewer 00:07:55.790 --> 00:07:59.269 uplides a real 3D model file, an STL file of 00:07:59.269 --> 00:08:03.149 the Benchy, and they ask the AI to write a slicer. 00:08:03.230 --> 00:08:05.129 Now, a slicer is a serious piece of software. 00:08:05.449 --> 00:08:08.350 It takes a 3D object, cuts it into thousands 00:08:08.350 --> 00:08:10.509 of thin layers, and then generates G code, which 00:08:10.509 --> 00:08:12.269 is basically coordinate instructions for the 00:08:12.269 --> 00:08:14.850 printer. Wait, wait. So the AI wrote the software 00:08:14.850 --> 00:08:17.569 to read the 3D file, sliced it into layers, and 00:08:17.569 --> 00:08:19.670 then simulated the printing of those layers on 00:08:19.670 --> 00:08:22.980 the screen. Yes. It parsed the geometry. It didn't 00:08:22.980 --> 00:08:25.839 just look up how to draw a boat. It calculated 00:08:25.839 --> 00:08:28.399 the actual path the nozzle would need to take 00:08:28.399 --> 00:08:30.980 to physically create that object in a virtual 00:08:30.980 --> 00:08:33.399 space. That feels like a massive leap. So is 00:08:33.399 --> 00:08:36.500 it truly simulating or is it just copying slicer 00:08:36.500 --> 00:08:38.279 code it found somewhere else? That's the skeptics 00:08:38.279 --> 00:08:40.059 question and it's a good one. But the source 00:08:40.059 --> 00:08:43.259 argues it's true simulation because of how it 00:08:43.259 --> 00:08:46.019 handled that specific unique file the user uploaded. 00:08:46.460 --> 00:08:48.500 You can't just copy paste that kind of real time 00:08:48.500 --> 00:08:50.909 interaction. It has to understand the coordinate 00:08:50.909 --> 00:08:53.269 system, the x, y, and z axes, and how they relate 00:08:53.269 --> 00:08:55.809 to the math of the motors. It understands physical 00:08:55.809 --> 00:09:00.149 space. That's a lot to process. It suggests a 00:09:00.149 --> 00:09:02.049 kind of spatial intelligence we didn't think 00:09:02.049 --> 00:09:04.330 these text models had. Let's pivot to something 00:09:04.330 --> 00:09:08.090 a bit more fun, but equally revealing. The gaming 00:09:08.090 --> 00:09:11.269 tests. The source mentioned a flight combat simulator. 00:09:11.389 --> 00:09:13.049 This part was actually pretty funny. The reviewer 00:09:13.049 --> 00:09:15.950 notes that GPT 5 .3 has a bit of an attitude. 00:09:16.090 --> 00:09:19.600 An attitude, like a personality. Yeah. The model 00:09:19.600 --> 00:09:22.580 is slow. It thinks a lot. And during the flight 00:09:22.580 --> 00:09:25.120 sim build, the reviewer got impatient, you know, 00:09:25.179 --> 00:09:27.360 like we all do, and typed, just give me the file. 00:09:27.980 --> 00:09:30.220 And the AI actually snapped back at him. It said 00:09:30.220 --> 00:09:32.860 something like, I am verifying the physics. Wait 00:09:32.860 --> 00:09:36.059 a moment. It told the human to chill out. Basically. 00:09:36.639 --> 00:09:39.379 It was prioritizing the integrity of the code 00:09:39.379 --> 00:09:42.080 over the user's impatience. That's a very senior 00:09:42.080 --> 00:09:44.019 engineer kind of move. Do you want it fast or 00:09:44.019 --> 00:09:46.480 do you want it right? And when the game finally 00:09:46.480 --> 00:09:48.559 loaded, again, visually, it looked like paper 00:09:48.559 --> 00:09:52.139 triangles. Ugly. But the radar worked. The enemies 00:09:52.139 --> 00:09:54.779 tracked you. When you hit a plane, smoke particles 00:09:54.779 --> 00:09:56.940 came out. The logic was solid. And then they 00:09:56.940 --> 00:10:00.320 moved to C++. Now, I have to make a vulnerable 00:10:00.320 --> 00:10:03.720 admission here. I've dabbled in code. But C++ 00:10:03.720 --> 00:10:07.039 memory management? It keeps me up at night. It 00:10:07.039 --> 00:10:10.100 is just notoriously unforgiving. Oh, it's the 00:10:10.100 --> 00:10:11.919 final boss of programming languages for a lot 00:10:11.919 --> 00:10:14.000 of people. It forces you to manage the computer's 00:10:14.000 --> 00:10:16.919 memory by hand. So the reviewer asks for a skateboarding 00:10:16.919 --> 00:10:20.379 game in C++. Specifically, they wanted grinding 00:10:20.379 --> 00:10:23.399 logic. Grinding seems incredibly difficult to 00:10:23.399 --> 00:10:26.179 code. You have to know the exact moment the board 00:10:26.179 --> 00:10:29.299 intersects a rail and then lock it on while keeping 00:10:29.299 --> 00:10:31.539 momentum. Exactly. It's a collision detection 00:10:31.539 --> 00:10:35.809 nightmare. And the AI. It nailed it. The source 00:10:35.809 --> 00:10:38.190 says the reviewer could jump, land on the rails, 00:10:38.389 --> 00:10:41.190 slide, and jump off. It even built a combo system 00:10:41.190 --> 00:10:44.110 where the score multiplier would reset if you 00:10:44.110 --> 00:10:46.529 fell. So going back to that attitude you mentioned, 00:10:46.830 --> 00:10:49.350 the delay, the snippiness, does that attitude 00:10:49.350 --> 00:10:51.769 actually improve the code? Yes, because that 00:10:51.769 --> 00:10:53.490 delay is the model reasoning. It's running through 00:10:53.490 --> 00:10:55.809 the logic. If it had rushed to give a good enough 00:10:55.809 --> 00:10:57.929 answer to make the user happy, the skateboard 00:10:57.929 --> 00:10:59.509 would have clipped through the rail or the game 00:10:59.509 --> 00:11:01.590 would have just crashed. That's a trade -off 00:11:01.590 --> 00:11:04.509 I think most developers would take any day, quality 00:11:04.509 --> 00:11:08.029 over speed. Now before we get into the project 00:11:08.029 --> 00:11:10.250 management side of this, which might be the most 00:11:10.250 --> 00:11:12.110 practical part, we're going to take a very quick 00:11:12.110 --> 00:11:16.980 break. Okay, we're back. We've talked about operating 00:11:16.980 --> 00:11:18.840 systems, 3D printers, skateboarding physics, 00:11:19.139 --> 00:11:21.360 but let's be real. Most listeners aren't building 00:11:21.360 --> 00:11:23.740 physics engines every day. They're building apps. 00:11:23.879 --> 00:11:25.980 They're managing projects. And this is where 00:11:25.980 --> 00:11:29.740 GPT 5 .3 seems to shift from just being a coder 00:11:29.740 --> 00:11:32.639 to being a manager. This is the plan mode feature. 00:11:33.100 --> 00:11:35.399 The reviewer tested it with a game called Neon 00:11:35.399 --> 00:11:38.080 Arena, a first -person shooter. But instead of 00:11:38.080 --> 00:11:39.600 just spitting out a wall of code immediately, 00:11:39.879 --> 00:11:42.720 the AI paused. It started interviewing the user. 00:11:43.019 --> 00:11:45.519 Interviewing, like gathering requirements. Exactly. 00:11:45.840 --> 00:11:47.340 It asks things like, what kind of enemies do 00:11:47.340 --> 00:11:49.480 you want? Do you want this in one massive file 00:11:49.480 --> 00:11:52.480 or split into multiple files? That question alone, 00:11:52.700 --> 00:11:54.980 one file or multiple shows a level of seniority. 00:11:55.480 --> 00:11:58.059 A junior dev just jams everything into one script. 00:11:58.899 --> 00:12:01.480 A senior dev modularizes because they know it's 00:12:01.480 --> 00:12:03.620 easier to maintain later. And when the reviewer 00:12:03.620 --> 00:12:06.240 said multiple files, what happened? It created 00:12:06.240 --> 00:12:09.220 a professional file structure. player .py for 00:12:09.220 --> 00:12:13.500 movement, enemy .py for the AI, main .py to run 00:12:13.500 --> 00:12:15.919 it all. It organized the chaos before it started. 00:12:16.039 --> 00:12:18.159 It wasn't just a chat bot. It was acting like 00:12:18.159 --> 00:12:21.059 a tech lead, setting up a project. That structure 00:12:21.059 --> 00:12:23.600 is everything. But there's another side to this 00:12:23.600 --> 00:12:26.299 management role, interpreting the vision, the 00:12:26.299 --> 00:12:28.990 Stevie Slapis test. I love this name, by the 00:12:28.990 --> 00:12:32.330 way. Stevie Slapis. It sounds like a bad cartoon 00:12:32.330 --> 00:12:35.190 character. So the reviewer drew this deliberately 00:12:35.190 --> 00:12:37.990 ugly wire frame on a piece of paper for a fake 00:12:37.990 --> 00:12:41.149 portfolio website. Just boxes labeled skills 00:12:41.149 --> 00:12:43.289 and projects. They took a photo, uploaded it. 00:12:43.490 --> 00:12:45.649 And the prompt was just... Make this beautiful. 00:12:45.909 --> 00:12:47.929 Pretty much. Make it beautiful. Add a wow factor. 00:12:48.210 --> 00:12:51.210 And the AI was incredibly faithful to the drawing. 00:12:51.429 --> 00:12:54.330 It put the skills section exactly where the ugly 00:12:54.330 --> 00:12:56.629 sketch had it. It didn't try to fix the layout, 00:12:56.809 --> 00:12:59.070 but it polished the execution. It added these 00:12:59.070 --> 00:13:01.009 nice hover glow effects. It made the whole thing 00:13:01.009 --> 00:13:03.289 responsive for mobile. So it took a napkin sketch 00:13:03.289 --> 00:13:06.509 and turned it into a real website. This raises 00:13:06.509 --> 00:13:09.669 a big question for me. If it can plan the architecture 00:13:09.669 --> 00:13:12.850 like a tech lead and execute code like a developer. 00:13:13.450 --> 00:13:16.429 Is it replacing the coder or the manager? It's 00:13:16.429 --> 00:13:18.789 acting as both, creating a bridge between them. 00:13:19.210 --> 00:13:21.429 It structures the chaos of the code, but it still 00:13:21.429 --> 00:13:24.289 needs the human to provide that ugly sketch, 00:13:24.409 --> 00:13:26.710 the initial vision. It interprets our intent. 00:13:27.019 --> 00:13:29.720 The AI didn't know why Stevie Slapis needed a 00:13:29.720 --> 00:13:32.360 portfolio, but it knew how to build one. Okay, 00:13:32.500 --> 00:13:34.639 let's bring this all together. The source material 00:13:34.639 --> 00:13:38.279 ends with a direct comparison between GBT 5 .3 00:13:38.279 --> 00:13:41.500 and that other big release, Opus 4 .6. If you're 00:13:41.500 --> 00:13:42.899 a listener trying to decide which one to use, 00:13:43.039 --> 00:13:44.820 how does it break down? It's a classic straight 00:13:44.820 --> 00:13:46.679 off. It's almost like hiring two different kinds 00:13:46.679 --> 00:13:49.549 of people. Opus 4 .6 is the client -ready model. 00:13:49.789 --> 00:13:51.870 It gives you beautiful visuals instantly. It's 00:13:51.870 --> 00:13:54.190 polite. It's expensive. If you need a demo for 00:13:54.190 --> 00:13:56.330 a CEO in five minutes and it has to look pretty, 00:13:56.549 --> 00:14:01.090 you use Opus. And GPT 5 .3. GPT 5 .3 is the engine 00:14:01.090 --> 00:14:04.029 room model. It gives you ugly first drafts. It 00:14:04.029 --> 00:14:06.389 uses fewer tokens, so it's cheaper to run for 00:14:06.389 --> 00:14:09.850 these long, complex sessions. But the deep logic, 00:14:10.309 --> 00:14:13.929 the C++ memory management, the physics, the iterative 00:14:13.929 --> 00:14:17.820 problem solving, It's just superior. It's best 00:14:17.820 --> 00:14:20.840 for real software, where the plumbing matters 00:14:20.840 --> 00:14:23.659 more than the paint. So the big idea recap here 00:14:23.659 --> 00:14:25.879 seems to be that we need to adjust our expectations. 00:14:26.360 --> 00:14:29.320 The source concludes that GPT 5 .3 is not a magic 00:14:29.320 --> 00:14:31.500 wand. Right. It's a co -worker. And just like 00:14:31.500 --> 00:14:32.860 any co -worker, you have to guide it. You have 00:14:32.860 --> 00:14:34.899 to say, hey, that UI is terrible. Fix it. Or 00:14:34.899 --> 00:14:37.679 that physics is slightly off. But the difference 00:14:37.679 --> 00:14:40.320 is it works incredibly hard. It doesn't complain 00:14:40.320 --> 00:14:42.460 well, except for that one time. And it iterates. 00:14:42.620 --> 00:14:45.500 Instantly that shift from magic one to co -worker 00:14:45.500 --> 00:14:47.700 feels really important We spent the last few 00:14:47.700 --> 00:14:49.740 years treating AI like a vending machine You 00:14:49.740 --> 00:14:52.059 put a prompt in you get a product out if it's 00:14:52.059 --> 00:14:54.659 bad You kick the machine exactly and this model 00:14:54.659 --> 00:14:57.259 forces you to treat it more like a partner. It's 00:14:57.259 --> 00:14:59.240 let's build this together You're the architect. 00:14:59.259 --> 00:15:01.980 It's the builder. It's closing the gap between 00:15:01.980 --> 00:15:06.019 having an idea I want a browser OS and actually 00:15:06.019 --> 00:15:08.460 holding that product. It's messy, it requires 00:15:08.460 --> 00:15:11.519 feedback, but the barrier to entry is just crumbling. 00:15:11.779 --> 00:15:14.100 It really is. And for the first time, it feels 00:15:14.100 --> 00:15:17.259 like the AI is actually meeting us halfway. So 00:15:17.259 --> 00:15:18.940 here's our final thought for you to take away 00:15:18.940 --> 00:15:22.179 today. If this model can remember the state of 00:15:22.179 --> 00:15:24.559 a simulated computer without being told how, 00:15:25.100 --> 00:15:27.759 and if it can slice a 3D model by understanding 00:15:27.759 --> 00:15:31.980 geometry, what happens when we ask it to optimize 00:15:31.980 --> 00:15:35.059 not just the code, but the systems around the 00:15:35.059 --> 00:15:38.379 code. What happens when the coworker starts suggesting 00:15:38.379 --> 00:15:40.559 changes to the business logic itself, not just 00:15:40.559 --> 00:15:42.840 the syntax? That is the billion -dollar question 00:15:42.840 --> 00:15:44.779 when the tool starts suggesting what to build, 00:15:44.980 --> 00:15:47.019 not just how to build it. We'd love to hear what 00:15:47.019 --> 00:15:48.860 you think. Are you ready for a coworker that 00:15:48.860 --> 00:15:50.960 argues with you about facts? Thanks for listening 00:15:50.960 --> 00:15:52.779 to this deep dive. See you the next one. Take 00:15:52.779 --> 00:15:52.919 care.