WEBVTT 00:00:00.000 --> 00:00:02.680 I want you to imagine a scenario. It's a Tuesday 00:00:02.680 --> 00:00:05.940 morning. You are sitting at your desk. Maybe 00:00:05.940 --> 00:00:07.799 you have your coffee in hand. Naturally. You 00:00:07.799 --> 00:00:10.400 lean back in your chair. You fold your hands 00:00:10.400 --> 00:00:13.240 in your lap. You are absolutely not touching 00:00:13.240 --> 00:00:15.900 the keyboard. Right. And you are definitely not 00:00:15.900 --> 00:00:17.640 touching the mouse. Kind of hard to work that 00:00:17.640 --> 00:00:20.859 way. Exactly. But on your screen, the cursor 00:00:20.859 --> 00:00:24.600 is moving. And it's not. It's not glitching out. 00:00:24.780 --> 00:00:27.920 It's moving with purpose. Yeah. It clicks open 00:00:27.920 --> 00:00:31.320 a web browser. It navigates to a vendor's website. 00:00:31.519 --> 00:00:34.119 It's just by itself? Completely by itself. It 00:00:34.119 --> 00:00:36.520 fills out a complex order form, checking boxes, 00:00:36.780 --> 00:00:39.060 scrolling down to find the submit button. Wow. 00:00:39.240 --> 00:00:42.259 Then it opens a spreadsheet and logs the confirmation 00:00:42.259 --> 00:00:45.299 number. That's wild. It's not a recording? Not 00:00:45.299 --> 00:00:47.719 a script you wrote line by line? No. It's an 00:00:47.719 --> 00:00:50.659 AI literally using the computer interface. Just 00:00:50.659 --> 00:00:53.880 like a human employee would. It sounds like science 00:00:53.880 --> 00:00:56.200 fiction. Or maybe a hacker taking over your machine. 00:00:56.520 --> 00:00:58.700 It really does. But this is actually the reality 00:00:58.700 --> 00:01:00.299 we're looking at today. This is the computer 00:01:00.299 --> 00:01:02.899 use feature. And it is the headline act for the 00:01:02.899 --> 00:01:05.840 model we're discussing. Plaudsonnet 4 .6. Welcome 00:01:05.840 --> 00:01:08.359 to the deep dive. Today, we are not just talking 00:01:08.359 --> 00:01:11.120 about a chat bot that writes better poetry. Right. 00:01:11.500 --> 00:01:13.719 We're looking at a fundamental shift in mechanics. 00:01:14.019 --> 00:01:17.799 We are. Our source for this is a guide. It's 00:01:17.799 --> 00:01:21.099 called Mastering Claude 3 .5 Sonnet. Though, 00:01:21.400 --> 00:01:23.640 just to be totally clear for you listening, the 00:01:23.640 --> 00:01:26.099 text focuses heavily on the upgraded version. 00:01:26.420 --> 00:01:30.019 Sonnet 4 .6. Exactly. It's a fascinating document. 00:01:30.459 --> 00:01:32.799 It frames this moment not just as an upgrade 00:01:32.799 --> 00:01:36.879 in IQ, but as a shift in utility. Exactly. The 00:01:36.879 --> 00:01:39.159 premise here is striking. We're looking at a 00:01:39.159 --> 00:01:42.640 model that claims to be twice as fast as the 00:01:42.640 --> 00:01:45.079 Heavy Opus model. Which was the previous King 00:01:45.079 --> 00:01:47.739 of the Hill. Right. Yet it costs half the price. 00:01:47.900 --> 00:01:50.420 Half the price. Half. And if you believe the 00:01:50.420 --> 00:01:52.939 benchmarks, it's significantly smarter. That 00:01:52.939 --> 00:01:55.700 is the holy grail of software, isn't it? Better, 00:01:55.980 --> 00:01:58.099 faster, cheaper. Usually in engineering, you're 00:01:58.099 --> 00:02:00.099 told to pick two. If it's fast and cheap, it's 00:02:00.099 --> 00:02:03.099 usually garbage. Yeah, usually. And that is why 00:02:03.099 --> 00:02:05.719 we need to unpack this carefully. Claims like 00:02:05.719 --> 00:02:08.039 that usually come with an asterisk. So we have 00:02:08.039 --> 00:02:10.560 a roadmap for today's deep dive. Let's do it. 00:02:11.379 --> 00:02:13.930 First, we need to understand the brain. Specifically, 00:02:14.129 --> 00:02:17.150 the context window and this spooky new ability 00:02:17.150 --> 00:02:19.750 to control computers. Then we have to look at 00:02:19.750 --> 00:02:22.449 the economics. For businesses, that is the bottom 00:02:22.449 --> 00:02:24.810 line. Then we will walk through the stress tests. 00:02:25.009 --> 00:02:27.830 Everything from fake operating systems to 3D 00:02:27.830 --> 00:02:30.550 survival games. And finally, we'll look at agent 00:02:30.550 --> 00:02:33.870 mode. Where the AI takes the wheel on browser 00:02:33.870 --> 00:02:35.909 automation. Exactly. So let's start with the 00:02:35.909 --> 00:02:38.650 engine. What is going on under the hood here? 00:02:38.830 --> 00:02:41.569 What makes Sonnet 4 .6 distinct? Yeah. Well, 00:02:41.569 --> 00:02:43.469 the first thing that jumps out is the sheer size 00:02:43.469 --> 00:02:45.969 of the memory space. OK. This model comes with 00:02:45.969 --> 00:02:49.090 a 1 million token context window. 1 million. 00:02:49.129 --> 00:02:51.509 We hear these numbers thrown around a lot. 32k, 00:02:51.629 --> 00:02:54.530 128k, now a million. Right. But let's ground 00:02:54.530 --> 00:02:57.569 that. What is a token, basically? A token is 00:02:57.569 --> 00:02:59.789 roughly a piece of a word. So what does a million 00:02:59.789 --> 00:03:02.710 tokens actually feel like to you as a user? Think 00:03:02.710 --> 00:03:05.409 of it this way. It's like giving the AI a very 00:03:05.409 --> 00:03:10.060 thick or a massive disorganized folder of code. 00:03:10.740 --> 00:03:13.699 In previous generations, the AI was like a sieve. 00:03:13.780 --> 00:03:16.560 It's there, it might read the first few chapters, 00:03:16.860 --> 00:03:18.919 but by the time you asked a question at chapter 00:03:18.919 --> 00:03:20.860 10, it already forgot the character names from 00:03:20.860 --> 00:03:24.080 chapter one. Exactly. The goldfish memory problem. 00:03:24.180 --> 00:03:26.259 You have to keep reminding it. Like, hey, remember 00:03:26.259 --> 00:03:28.680 that rule I gave you 20 minutes ago? Precisely. 00:03:29.199 --> 00:03:32.939 With a one million token window, that sieve becomes 00:03:32.939 --> 00:03:36.280 a vault. A vault. It holds that entire thick 00:03:36.280 --> 00:03:39.460 book in its head at once. Wow. It remembers the 00:03:39.460 --> 00:03:41.900 footnote on page three while working on page 00:03:41.900 --> 00:03:45.560 500. That is a huge quality of life improvement. 00:03:45.639 --> 00:03:47.500 It holds the state of the conversation perfectly. 00:03:47.659 --> 00:03:50.280 But the guide suggests the bigger shish isn't 00:03:50.280 --> 00:03:54.539 just memory, it's agency. This computer use capability. 00:03:54.960 --> 00:03:57.419 Right. Let's talk about that. This feels like 00:03:57.419 --> 00:04:00.960 a real departure. Until now, AI has been a text 00:04:00.960 --> 00:04:03.650 generator. You type, it types back. But Sonnet 00:04:03.650 --> 00:04:06.930 4 .6 is different. The guide describes this as 00:04:06.930 --> 00:04:10.050 the ability to handle boring clicking tasks. 00:04:10.569 --> 00:04:13.889 Boring clicking tasks. Yeah. It can read complex 00:04:13.889 --> 00:04:16.689 spreadsheets, visually scan them, fill out long 00:04:16.689 --> 00:04:18.709 forms without getting tired. It acts almost like 00:04:18.709 --> 00:04:21.689 a real person sitting at the terminal. When you 00:04:21.689 --> 00:04:24.310 say visually scanned, does it actually see the 00:04:24.310 --> 00:04:27.410 screen? Or is it looking at the code behind the 00:04:27.410 --> 00:04:30.839 website? That is a key distinction. It is actually 00:04:30.839 --> 00:04:33.980 analyzing screenshots of the interface in real 00:04:33.980 --> 00:04:36.180 time. Really? Yeah. It looks at the buttons, 00:04:36.379 --> 00:04:39.100 the layout, the text fields, just like your eyes 00:04:39.100 --> 00:04:41.420 do. The source mentions it's great at step -by 00:04:41.420 --> 00:04:43.180 -step development. Is that just about writing 00:04:43.180 --> 00:04:46.620 code? It's much broader. The model builds a project, 00:04:46.920 --> 00:04:49.240 inspects what it made, and then fixes its own 00:04:49.240 --> 00:04:52.000 errors. As it goes. Right. It's recursive. It 00:04:52.000 --> 00:04:53.639 doesn't just shoot out an answer and hope for 00:04:53.639 --> 00:04:56.620 the best. It manages the task from start to finish. 00:04:56.980 --> 00:04:59.779 Exactly. Does this capacity for step -by -step 00:04:59.779 --> 00:05:02.720 correction fundamentally change how we interact 00:05:02.720 --> 00:05:05.620 with it? Yes. It moves from just answering questions 00:05:05.620 --> 00:05:08.439 to managing entire workflows start to finish. 00:05:08.699 --> 00:05:10.699 So we manage workflows now instead of just asking 00:05:10.699 --> 00:05:13.199 questions. Exactly. You become the architect. 00:05:13.579 --> 00:05:17.019 The AI becomes the builder. But usually, When 00:05:17.019 --> 00:05:20.540 you get an agent that can replace a junior employee's 00:05:20.540 --> 00:05:23.100 clicking capability, you expect a premium price 00:05:23.100 --> 00:05:25.680 tag. And that is where the market is getting 00:05:25.680 --> 00:05:28.779 aggressive. The guide breaks down the economics 00:05:28.779 --> 00:05:31.420 very clearly. Let's hear the numbers. For Sonnet 00:05:31.420 --> 00:05:36.540 4 .6, the input cost is $3 for every 1 million 00:05:36.540 --> 00:05:41.480 tokens. OK. The output cost is $15 per 1 million 00:05:41.480 --> 00:05:44.560 tokens. Contextualize that for me. How does that 00:05:44.560 --> 00:05:47.519 stack up against the older versions? It is offering 00:05:47.519 --> 00:05:50.680 top -level smarts at a middle -level price. It 00:05:50.680 --> 00:05:53.279 is half the price of the heavy Opus model, but 00:05:53.279 --> 00:05:55.579 it runs two times faster. And speed matters. 00:05:55.920 --> 00:05:58.839 Waiting 10 seconds versus 20 seconds is the difference 00:05:58.839 --> 00:06:01.720 between staying in flow and getting totally distracted. 00:06:01.899 --> 00:06:04.060 It really is. Yeah. But let's look at the benchmarks. 00:06:04.800 --> 00:06:08.300 Specifically, SW Bench. SW Bench. Explain that 00:06:08.300 --> 00:06:10.939 in plain English for us. It is a test. checking 00:06:10.939 --> 00:06:13.620 how well AI fixes code like a software engineer. 00:06:13.920 --> 00:06:16.040 OK. So it's not just writing a simple function. 00:06:16.139 --> 00:06:18.500 No, it's much messier. Yeah. The test gives the 00:06:18.500 --> 00:06:21.100 AI a real -world software repository and a bug 00:06:21.100 --> 00:06:23.079 description. And it has to fix it. It has to 00:06:23.079 --> 00:06:24.680 navigate the files, figure out what's broken, 00:06:24.819 --> 00:06:26.639 write the fix, and not break anything else. A 00:06:26.639 --> 00:06:29.319 test of troubleshooting. Exactly. In this test, 00:06:29.699 --> 00:06:33.660 Sonnet 4 .6 scores an 80 .2%. That sounds high. 00:06:33.860 --> 00:06:36.720 It is very high. It beats its big brother, the 00:06:36.720 --> 00:06:39.579 Opus 4 model. So we are seeing an inversion where 00:06:39.579 --> 00:06:41.899 the cheaper model is actually the more capable 00:06:41.899 --> 00:06:44.740 one. Exactly. You no longer pay a premium for 00:06:44.740 --> 00:06:47.600 intelligence, you pay for efficiency. So the 00:06:47.600 --> 00:06:50.139 math is half the price, twice the speed, and 00:06:50.139 --> 00:06:52.839 better performance. Impossible to ignore. Let's 00:06:52.839 --> 00:06:55.500 move to the visual tests. Numbers are one thing, 00:06:55.639 --> 00:06:58.439 but seeing what it builds is another. The guide 00:06:58.439 --> 00:07:00.959 walks through a few specific test cases. One 00:07:00.959 --> 00:07:03.360 of them was building a website for a local business, 00:07:03.800 --> 00:07:06.899 a locksmith in New York. This was a test of adherence 00:07:06.899 --> 00:07:09.480 to constraints. Right. The prompt was very specific. 00:07:09.920 --> 00:07:12.439 A white background. Dark blue buttons. To build 00:07:12.439 --> 00:07:15.399 trust. Exactly. And the key detail. An emergency 00:07:15.399 --> 00:07:18.680 service price of exactly $99. Sounds simple, 00:07:18.720 --> 00:07:21.519 but AI loves to hallucinate prices. Or they get 00:07:21.519 --> 00:07:23.839 creative and make the button purple. Right. But 00:07:23.839 --> 00:07:26.639 the result was spot on. Yeah. Professional fonts. 00:07:27.120 --> 00:07:30.120 Correct layout. Exact adherence to the price 00:07:30.120 --> 00:07:32.360 instruction. It looks like a human made it. It 00:07:32.360 --> 00:07:34.939 did. But then they pushed it harder. They asked 00:07:34.939 --> 00:07:37.959 it to simulate a Mac OS computer screen. Using 00:07:37.959 --> 00:07:41.000 only web code. Yes. Building a system inside 00:07:41.000 --> 00:07:43.899 a system. The first try was just okay. It was 00:07:43.899 --> 00:07:46.980 static. A painting of a computer. Right. But 00:07:46.980 --> 00:07:49.889 the tester asked it to improve. The second try 00:07:49.889 --> 00:07:52.910 was amazing. What changed? It added drag and 00:07:52.910 --> 00:07:55.730 drop folders. Wait, drag and drop? Yes. That 00:07:55.730 --> 00:07:59.029 requires complex logic. Tracking mouse position, 00:07:59.350 --> 00:08:01.550 updating elements instantly. It worked. You could 00:08:01.550 --> 00:08:03.810 change backgrounds. It built a music player with 00:08:03.810 --> 00:08:06.069 a moving time bar. I have to admit something 00:08:06.069 --> 00:08:08.389 here, Beat. I still wrestle with prompt rift 00:08:08.389 --> 00:08:11.100 myself. Oh, we all do. You ask an AI to fix one 00:08:11.100 --> 00:08:13.060 thing, it breaks three others. It forgets the 00:08:13.060 --> 00:08:15.720 button color. Yeah. Hearing it maintains the 00:08:15.720 --> 00:08:17.939 state of a music player while dragging a folder 00:08:17.939 --> 00:08:21.240 is so relieving. It shows real robustness. It 00:08:21.240 --> 00:08:23.540 handles layout well, but does it have artistic 00:08:23.540 --> 00:08:27.540 intuition? Not quite. Its SVG drawings, like 00:08:27.540 --> 00:08:31.139 butterflies and robots, are just okay. Opus is 00:08:31.139 --> 00:08:33.320 still the better artist. So it's an engineer, 00:08:33.480 --> 00:08:35.980 not a painter. Exactly. Which leads us to the 00:08:35.980 --> 00:08:39.580 gaming tests. Games are systems. Physics and 00:08:39.580 --> 00:08:42.279 logic. This is where it really flexed. First 00:08:42.279 --> 00:08:45.580 test was a marble board game. Using mouse movement 00:08:45.580 --> 00:08:49.000 to tilt a wooden board. Yes. The marble rolls 00:08:49.000 --> 00:08:52.159 faster or slower based on tilt. Black holes act 00:08:52.159 --> 00:08:55.600 as obstacles. Vector math. Understanding the 00:08:55.600 --> 00:08:58.080 connection between input and physics. And it 00:08:58.080 --> 00:09:00.740 worked instantly. The marble rolled realistically. 00:09:01.000 --> 00:09:03.539 But the next test was even bigger. A Minecraft 00:09:03.539 --> 00:09:08.070 clone. Boxelcraft. Tested using KiloCode. Huge 00:09:08.070 --> 00:09:10.870 prompt, I imagine. Massive. Health bars, food 00:09:10.870 --> 00:09:14.830 bars. Breaking and placing blocks. And deep caves 00:09:14.830 --> 00:09:17.529 under the ground. Yes. It included a proper start 00:09:17.529 --> 00:09:20.149 menu, which is rare, and it generated those deep 00:09:20.149 --> 00:09:24.049 caves. Whoa! Two -sex silence. Think about that, 00:09:24.149 --> 00:09:26.509 generating deep caves under the ground. They're 00:09:26.509 --> 00:09:28.950 crazy. The AI built a hidden part of the world 00:09:28.950 --> 00:09:31.490 just because the prompt asked for it, simulating 00:09:31.490 --> 00:09:33.570 depth we can't initially see. It really is a 00:09:33.570 --> 00:09:35.769 moment of wonder. Did the browser actually handle 00:09:35.769 --> 00:09:39.070 a full 3D survival game written by AI? Mostly 00:09:39.070 --> 00:09:41.830 yes, though the graphics were heavy. The logic 00:09:41.830 --> 00:09:44.309 held up, even if the browser struggled. So the 00:09:44.309 --> 00:09:46.629 underlying system worked flawlessly, even if 00:09:46.629 --> 00:09:48.669 rendering lagged. We're going to take a quick 00:09:48.669 --> 00:09:50.750 break for our sponsors and when we come back, 00:09:50.889 --> 00:09:53.289 we'll talk about agent mode. Sounds good. And 00:09:53.289 --> 00:09:55.570 we're back. Let's look at the agent workflow, 00:09:55.830 --> 00:09:58.649 the big data task. This is about automation, 00:09:59.389 --> 00:10:02.090 treating the AI as an autonomous agent. The prompt 00:10:02.090 --> 00:10:04.549 was a long chain of commands. Create a folder, 00:10:04.809 --> 00:10:07.629 write a Python script, open a browser. Search 00:10:07.629 --> 00:10:10.750 Google for news, save to file, build a dashboard. 00:10:10.889 --> 00:10:13.330 That is where things usually break down. The 00:10:13.330 --> 00:10:16.649 AI forgets step three. But this time, it utilized 00:10:16.649 --> 00:10:19.830 Selenium. Define Selenium for us. It's browser 00:10:19.830 --> 00:10:22.440 automation software. code that clicks buttons 00:10:22.440 --> 00:10:25.860 on websites. Usually very finicky. Very. But 00:10:25.860 --> 00:10:28.580 the AI wrote the script and executed it. Wait, 00:10:28.659 --> 00:10:31.539 it ran the code? Yes. The computer physically 00:10:31.539 --> 00:10:34.000 opened the browser by itself, typed in the search 00:10:34.000 --> 00:10:36.679 bar, found top links, populated the dashboard. 00:10:37.059 --> 00:10:39.519 Is the main advantage here capability, or is 00:10:39.519 --> 00:10:42.440 it reliability? Reliability. Unlike others that 00:10:42.440 --> 00:10:45.600 get lazy, this model finishes long lists of steps 00:10:45.600 --> 00:10:47.940 without stopping. It doesn't get lazy and quit 00:10:47.940 --> 00:10:50.559 halfway. Exactly. It follows through. If someone 00:10:50.559 --> 00:10:52.600 listening wants to actually use this, how do 00:10:52.600 --> 00:10:55.440 they get access? The official chat is free, but 00:10:55.440 --> 00:10:57.940 has limited messages. The dreaded rate limit. 00:10:58.039 --> 00:11:00.820 Yeah. There's the API, which is pay as you go. 00:11:01.059 --> 00:11:04.980 $3 in, $15 out. Right. Then there's LMSYS for 00:11:04.980 --> 00:11:08.259 free, side -by -side testing. And KiloCode. Open 00:11:08.259 --> 00:11:10.899 source platform offering $25 in free credits. 00:11:11.159 --> 00:11:14.519 Great for testing. OK, what is the tactical advice 00:11:14.519 --> 00:11:17.700 for prompting this thing? Three tips. First, 00:11:17.960 --> 00:11:20.879 be specific. Give us an example. Don't just say 00:11:20.879 --> 00:11:24.120 write an ad. The guide uses the AI fire newsletter. 00:11:24.240 --> 00:11:27.340 OK. Specify the hook, the audience busy office 00:11:27.340 --> 00:11:29.980 workers, and the promise to save five hours a 00:11:29.980 --> 00:11:32.879 week. Give it constraints. Second tip, iterate. 00:11:33.179 --> 00:11:35.659 Build in small pieces. Like stacking Lego blocks 00:11:35.659 --> 00:11:38.580 of beta. Log in screen first, then dashboard. 00:11:38.700 --> 00:11:41.019 It stops the AI from getting confused. And third 00:11:41.019 --> 00:11:44.000 tip. feed the memory, use that 1 million token 00:11:44.000 --> 00:11:47.080 window, paste in whole documents. Don't be shy. 00:11:47.279 --> 00:11:49.879 Exactly. Let's recap the big idea here. We are 00:11:49.879 --> 00:11:52.080 seeing a shift toward high intelligence, low 00:11:52.080 --> 00:11:55.740 price. Yes. It excels at doing coding, clicking, 00:11:56.419 --> 00:11:58.539 automating. Rather than just knowing facts. It 00:11:58.539 --> 00:12:01.620 lacks artistic flair for SVGs, but makes up for 00:12:01.620 --> 00:12:04.279 it with strict adherence to rules and massive 00:12:04.279 --> 00:12:07.000 memory. It's a highly skilled engineer. So here's 00:12:07.000 --> 00:12:09.720 my challenge to you listening. Try the marble 00:12:09.720 --> 00:12:13.230 game prompt. or use kilo code credits to build 00:12:13.230 --> 00:12:16.009 a small tool. The source mentions knowing how 00:12:16.009 --> 00:12:18.870 to use these tools reduces work stress. We are 00:12:18.870 --> 00:12:21.549 moving away from reading every line of code to 00:12:21.549 --> 00:12:23.730 orchestrating the system. It's a totally different 00:12:23.730 --> 00:12:25.950 kind of work. Focus on the what and why. Let 00:12:25.950 --> 00:12:29.110 the AI handle the how. Exactly. Thanks for diving 00:12:29.110 --> 00:12:31.289 in with us. Always a pleasure. See you in the 00:12:31.289 --> 00:12:32.009 next deep dive.