WEBVTT 00:00:00.000 --> 00:00:04.599 A fully finished 23 second video intro beat generated 00:00:04.599 --> 00:00:07.639 from a single slash goal prompt. Wow. There was 00:00:07.639 --> 00:00:10.099 no step by step babysitting involved at all. 00:00:10.199 --> 00:00:12.480 It took one hour and 15 minutes to complete. 00:00:12.660 --> 00:00:14.960 Yeah, that is wild. And it didn't even max out 00:00:14.960 --> 00:00:17.679 the one million context window. Right. And it 00:00:17.679 --> 00:00:19.460 used three hundred fifty seven thousand tokens 00:00:19.460 --> 00:00:22.579 to do it. This completely shifts how we have 00:00:22.579 --> 00:00:25.019 to look at open source capabilities today. Welcome 00:00:25.019 --> 00:00:27.399 to the deep dive. We are very glad you are here 00:00:27.399 --> 00:00:30.460 with us. Absolutely. Today, our mission is to 00:00:30.460 --> 00:00:34.200 unpack a really rigorous hands -on guide. We 00:00:34.200 --> 00:00:38.460 are testing the open source GLM 5 .2 model, and 00:00:38.460 --> 00:00:40.460 we're doing this directly inside Cloud Code. 00:00:40.640 --> 00:00:42.719 We are going to walk through some actual real 00:00:42.719 --> 00:00:45.780 world tests here. We pit GLM 5 .2 against the 00:00:45.780 --> 00:00:48.579 heavyweight OPUS 4 .8. Right. We will outline 00:00:48.579 --> 00:00:51.119 a fascinating 80 -20 workflow split for your 00:00:51.119 --> 00:00:53.780 daily tasks. We will also break down the surprisingly 00:00:53.780 --> 00:00:56.359 simple non -technical setup. Finally, we will 00:00:56.359 --> 00:00:58.939 reveal why this specific shift makes open source 00:00:58.939 --> 00:01:02.200 AI impossible to ignore. Beat. But to really 00:01:02.200 --> 00:01:03.899 understand why this matters, we have to look 00:01:03.899 --> 00:01:06.590 at the data. Yeah, what actually happened when 00:01:06.590 --> 00:01:08.730 these two models went head -to -head in the wild? 00:01:08.930 --> 00:01:10.730 That is the most important question. There were 00:01:10.730 --> 00:01:14.129 several highly practical tests run here. Right. 00:01:14.129 --> 00:01:17.150 We wanted to see if GLM 5 .2 could genuinely 00:01:17.150 --> 00:01:20.189 compete on real work. Or did it just feel good 00:01:20.189 --> 00:01:22.629 because it was the cheaper option? Well, the 00:01:22.629 --> 00:01:25.969 first test was a one -shot web design task. The 00:01:25.969 --> 00:01:29.250 exact same prompt was used for both models. They 00:01:29.250 --> 00:01:31.290 both had to build a fully functional landing 00:01:31.290 --> 00:01:34.459 page. GLM finished the job in 3 minutes and 59 00:01:34.459 --> 00:01:38.060 seconds. Wow. And Opus took 14 minutes and 59 00:01:38.060 --> 00:01:40.560 seconds to finish. That is an absolutely massive 00:01:40.560 --> 00:01:43.420 difference in time. And GLM was around five times 00:01:43.420 --> 00:01:45.959 cheaper to run. Yeah. But the really fascinating 00:01:45.959 --> 00:01:48.700 part to me is the actual quality. Exactly. These 00:01:48.700 --> 00:01:51.859 were not clunky early 2000s wireframes. Right. 00:01:52.480 --> 00:01:55.019 Both results were highly polished and visually 00:01:55.019 --> 00:01:58.140 impressive. The pages had movement built into 00:01:58.140 --> 00:02:00.519 the design. They had properly structured sections 00:02:00.519 --> 00:02:02.959 and very clear calls to action. This was not 00:02:02.959 --> 00:02:04.959 some flimsy budget result you would have to rewrite 00:02:04.959 --> 00:02:08.099 anyway. No. It was a completely viable starting 00:02:08.099 --> 00:02:10.840 point for a real project. Looking at them side 00:02:10.840 --> 00:02:13.280 by side, you would have a hard time justifying 00:02:13.280 --> 00:02:15.819 that massive price tag for Opus. When the final 00:02:15.819 --> 00:02:18.520 result is that close, the price starts to matter 00:02:18.520 --> 00:02:21.400 heavily. It really does. But we didn't stop at 00:02:21.400 --> 00:02:24.919 visual design. Next up was hard coding. So GLM 00:02:24.919 --> 00:02:28.259 crushes basic web layout, which is great. But 00:02:28.259 --> 00:02:30.879 visual spacing is one thing. Logical reasoning 00:02:30.879 --> 00:02:33.639 is another entirely. Absolutely. That naturally 00:02:33.639 --> 00:02:37.860 makes you wonder what happened. Less forgiving. 00:02:38.039 --> 00:02:40.800 It was a complex coding assignment was evaluated 00:02:40.800 --> 00:02:43.699 using codecs. Using codecs kept the evaluation 00:02:43.699 --> 00:02:47.219 entirely neutral and fair. Exactly. And GLM 5 00:02:47.219 --> 00:02:50.159 .2 actually did very well here. It handled the 00:02:50.159 --> 00:02:52.530 bulk of the task perfectly. But. Opus caught 00:02:52.530 --> 00:02:55.569 one very subtle edge case that GLM missed entirely. 00:02:57.610 --> 00:02:59.849 tricky database values, it caught the subtle 00:02:59.849 --> 00:03:03.009 difference between true and one, or one and one 00:03:03.009 --> 00:03:04.810 point now. Oh, wow. If you have ever stared at 00:03:04.810 --> 00:03:06.729 a broken database at two in the morning, you 00:03:06.729 --> 00:03:09.169 know exactly why that matters. Yeah, bugs always 00:03:09.169 --> 00:03:11.490 live in those tiny details. It feels like a very 00:03:11.490 --> 00:03:14.050 natural division of labor. GLM is kind of like 00:03:14.050 --> 00:03:16.569 the fast junior developer. It writes the bulk 00:03:16.569 --> 00:03:20.110 of the code very quickly. And Opus is the senior 00:03:20.110 --> 00:03:22.590 architect catching the tricky database bugs before 00:03:22.590 --> 00:03:25.500 they launch. That is a perfect analogy for how 00:03:25.500 --> 00:03:29.379 this works. GLN is fantastic for fast implementation 00:03:29.379 --> 00:03:32.780 and initial scaffolding. Right. And OPUS is necessary 00:03:32.780 --> 00:03:36.039 for careful reasoning and tricky edge case handling. 00:03:36.280 --> 00:03:39.139 Exactly. I have to be honest here. I still wrestle 00:03:39.139 --> 00:03:42.629 with prompt drift myself. beat. You know, you 00:03:42.629 --> 00:03:45.530 give an AI a long list of instructions. And by 00:03:45.530 --> 00:03:48.289 step three, it completely forgets step one. So 00:03:48.289 --> 00:03:50.110 watching an open source model hold its focus 00:03:50.110 --> 00:03:52.370 over complex instructions is deeply impressive 00:03:52.370 --> 00:03:54.409 to me. It really is an incredible leap forward. 00:03:54.469 --> 00:03:56.689 Right. But there's a specific speed quirk we 00:03:56.689 --> 00:03:58.969 definitely should mention. Right. GLM is not 00:03:58.969 --> 00:04:01.569 always the faster model. Yeah. On a highly creative 00:04:01.569 --> 00:04:05.830 HTML task, GLM took 35 minutes. Opus finished 00:04:05.830 --> 00:04:08.610 that exact same task in only 11 minutes. The 00:04:08.610 --> 00:04:10.870 rule here seems to revolve entirely around reasoning. 00:04:11.050 --> 00:04:13.349 Exactly. The more reasoning required by the prompt, 00:04:13.569 --> 00:04:16.569 the slower GLM feels in practice. Execution -heavy 00:04:16.569 --> 00:04:19.750 tasks are incredibly fast. Planning and creative 00:04:19.750 --> 00:04:22.629 taste just take much longer. They both succeeded 00:04:22.629 --> 00:04:26.170 on the creative task, though. GLM built an interactive 00:04:26.170 --> 00:04:28.449 page called the Anatomy of Attention. Right. 00:04:28.529 --> 00:04:30.689 It featured moving background elements and token 00:04:30.689 --> 00:04:34.750 visuals. and Opus built the life of a Death Star. 00:04:35.050 --> 00:04:37.430 It was a beautifully structured timeline -style 00:04:37.430 --> 00:04:40.149 page. Both were excellent one -shot results. 00:04:40.689 --> 00:04:43.930 The final test here was deep research. This utilized 00:04:43.930 --> 00:04:46.910 a storm -style workflow. For anyone unfamiliar, 00:04:47.069 --> 00:04:49.250 that just means multiple sub -agents work together 00:04:49.250 --> 00:04:51.720 using different personas. Right. They compiled 00:04:51.720 --> 00:04:55.420 a very rich, highly detailed HTML report. It 00:04:55.420 --> 00:04:58.120 combined different expert lenses and actively 00:04:58.120 --> 00:05:01.139 challenged its own assumptions. And GLM managed 00:05:01.139 --> 00:05:04.339 this complex agent workflow beautifully. The 00:05:04.339 --> 00:05:06.800 significantly lower cost changes the math entirely 00:05:06.800 --> 00:05:09.240 for us here. Running multiple autonomous agents 00:05:09.240 --> 00:05:11.759 suddenly becomes justifiable for daily work. 00:05:11.959 --> 00:05:13.579 Yeah, you can test 10 different angles without 00:05:13.579 --> 00:05:16.100 burning through your cash. But wait, if GLM struggles 00:05:16.100 --> 00:05:17.879 with heavy reasoning, doesn't that make it too 00:05:17.879 --> 00:05:20.790 risky for professional work? Well, not necessarily. 00:05:21.410 --> 00:05:23.629 It is really about matching the model to the 00:05:23.629 --> 00:05:25.589 risk profile. You have to match the specific 00:05:25.589 --> 00:05:28.269 tool to the specific task. Right. Scaffolding 00:05:28.269 --> 00:05:30.870 a basic website layout carries very low risk. 00:05:31.170 --> 00:05:33.949 Exactly. But migrating a production database 00:05:33.949 --> 00:05:37.029 carries extremely high risk. You always use the 00:05:37.029 --> 00:05:38.829 heavy reasoning model for the high risk work. 00:05:38.949 --> 00:05:41.149 So you don't compare final answers, you compare 00:05:41.149 --> 00:05:43.569 your acceptable risk for the task. Precisely. 00:05:43.910 --> 00:05:46.470 Which leads us to a much broader mindset shift 00:05:46.470 --> 00:05:49.319 we need to discuss. Yeah. The real skill today 00:05:49.319 --> 00:05:53.459 is knowing exactly when to use GLM 5 .2. AI work 00:05:53.459 --> 00:05:55.980 is not a simple contest anymore. It is not about 00:05:55.980 --> 00:05:59.519 the best model wins mentality. No, it is a multi 00:05:59.519 --> 00:06:03.120 -step, highly iterative process. Most work involves 00:06:03.120 --> 00:06:05.339 researching, drafting, and continuous testing. 00:06:05.699 --> 00:06:07.839 Then you edit, make decisions, and finally ship 00:06:07.839 --> 00:06:10.500 the product. Each of those steps demands a completely 00:06:10.500 --> 00:06:12.819 different level of intelligence. Exactly. Let's 00:06:12.819 --> 00:06:15.379 break down the 80 % work first. This is the natural 00:06:15.379 --> 00:06:19.060 domain of GLM 5 .2. It easily handles first drafts 00:06:19.060 --> 00:06:21.480 and gathering initial research. It does basic 00:06:21.480 --> 00:06:24.000 web design and cleans up your messy notes. Right. 00:06:24.060 --> 00:06:26.360 It generates your initial options for a project. 00:06:27.300 --> 00:06:29.899 Cost matters immensely here because you are doing 00:06:29.899 --> 00:06:32.300 constant iterative testing. You are exploring 00:06:32.300 --> 00:06:34.540 different ideas. You don't want maximum pricing 00:06:34.540 --> 00:06:37.019 when you are just sketching a rough draft. Exactly. 00:06:37.459 --> 00:06:39.459 Then we have the remaining 10 to 20 percent work. 00:06:39.819 --> 00:06:42.740 This is the domain of OPUS 4 .8. This is the 00:06:42.740 --> 00:06:45.720 heavy thinking. Final reasoning, edge case review, 00:06:45.939 --> 00:06:48.800 and high -risk coding tasks. The system context 00:06:48.800 --> 00:06:51.420 is crucial here, too. The harness really matters. 00:06:51.579 --> 00:06:53.839 Right. For clarity, a harness is just a digital 00:06:53.839 --> 00:06:56.839 workspace where the AI can use tools. Cloud Code 00:06:56.839 --> 00:06:59.040 provides that powerful harness for us. It lets 00:06:59.040 --> 00:07:01.939 the model read local files and actively run terminal 00:07:01.939 --> 00:07:05.680 commands. So GLM is not just a cheap OPUS replacement. 00:07:05.980 --> 00:07:08.540 It is a cheaper worker operating inside the exact 00:07:08.540 --> 00:07:11.680 same system. Mm -hmm. Beat. But aren't we just 00:07:11.680 --> 00:07:13.660 complicating things by juggling multiple workers 00:07:13.660 --> 00:07:16.019 for a single project? It might seem that way 00:07:16.019 --> 00:07:18.629 at first glance. But using one expensive model 00:07:18.629 --> 00:07:21.629 for basic sorting tasks burns cash needlessly. 00:07:22.269 --> 00:07:24.029 Right. Splitting the work optimizes both your 00:07:24.029 --> 00:07:26.810 budget and the applied brain power. You get significantly 00:07:26.810 --> 00:07:29.149 better efficiency across the board. Let the cheap 00:07:29.149 --> 00:07:31.730 model gather the lumber. Let the expensive model 00:07:31.730 --> 00:07:34.129 build the house. Exactly. It is about working 00:07:34.129 --> 00:07:36.709 much smarter within the environment you already 00:07:36.709 --> 00:07:39.459 use every day. Placeholder for sponsor reads 00:07:39.459 --> 00:07:42.680 skip promotional text from newsletter, use provided 00:07:42.680 --> 00:07:45.620 sponsor copy. Hearing about those massive price 00:07:45.620 --> 00:07:47.980 differences makes me want to try this immediately. 00:07:48.379 --> 00:07:50.540 Yeah. But whenever we talk about open source 00:07:50.540 --> 00:07:52.920 routing, it usually involves spinning up servers 00:07:52.920 --> 00:07:56.319 or Docker containers. Right. Is this actually 00:07:56.319 --> 00:08:00.480 feasible for a normal user to set up? It absolutely 00:08:00.480 --> 00:08:03.459 is. You are not learning a brand new tool here. 00:08:03.459 --> 00:08:05.740 You are just routing the model call to a different 00:08:05.740 --> 00:08:08.379 location. You simply route it to z .ai instead 00:08:08.379 --> 00:08:10.639 of Anthropix server. Exactly. Step one is going 00:08:10.639 --> 00:08:14.439 to the z .ai API console online. You can actually 00:08:14.439 --> 00:08:16.620 test the model out right there first. They have 00:08:16.620 --> 00:08:19.600 3D generation tools and small mini games available. 00:08:19.819 --> 00:08:22.019 It gives you a great feel for how the model responds. 00:08:22.180 --> 00:08:24.040 Yeah. Then you choose your preferred billing 00:08:24.040 --> 00:08:26.740 method. You can pay per token or choose a set 00:08:26.740 --> 00:08:29.720 monthly plan. Those monthly plans run roughly 00:08:29.720 --> 00:08:34.860 $16, $64, or $144. There's some really great 00:08:34.860 --> 00:08:37.360 practical advice on this point. Keep your Claude 00:08:37.360 --> 00:08:40.600 plan active and just add z .ai on the side. Yeah, 00:08:40.600 --> 00:08:42.279 you don't have to choose just one platform. You 00:08:42.279 --> 00:08:44.340 use both of them together. Next, you generate 00:08:44.340 --> 00:08:47.580 a secure API key inside the console. This brings 00:08:47.580 --> 00:08:50.740 us to editing a specific local file. It is called 00:08:50.740 --> 00:08:55.559 settings .local .json. You find this file sitting 00:08:55.559 --> 00:08:58.179 inside the .clod folder on your machine. What 00:08:58.179 --> 00:09:00.440 is brilliant here is that you aren't installing 00:09:00.440 --> 00:09:03.759 heavy new software. Clod code is already looking 00:09:03.759 --> 00:09:06.399 for a brain at a specific web address. Right. 00:09:06.539 --> 00:09:08.980 All you were doing in that settings file is changing 00:09:08.980 --> 00:09:11.179 the address book. You pointed away from Anthropic 00:09:11.179 --> 00:09:14.549 servers. You route the base URL directly to z 00:09:14.549 --> 00:09:18.049 .ai instead. You leave the Anthropic API key 00:09:18.049 --> 00:09:21.409 completely blank, and you put your new z .ai 00:09:21.409 --> 00:09:24.110 key in as the auth token. Finally, you set the 00:09:24.110 --> 00:09:26.950 specific model name you want to use. It is surprisingly 00:09:26.950 --> 00:09:29.190 simple. But there is a genius trick we need to 00:09:29.190 --> 00:09:31.389 highlight here. You create two entirely separate 00:09:31.389 --> 00:09:33.429 folders on your local machine. Yeah, this is 00:09:33.429 --> 00:09:36.070 incredibly clever. One folder is named slash 00:09:36.070 --> 00:09:39.389 gm. The other folder is named slash opus. You 00:09:39.389 --> 00:09:41.909 put the custom routing configuration file only 00:09:41.909 --> 00:09:45.830 inside the slash glm folder. Wait, so just by 00:09:45.830 --> 00:09:48.470 changing directories in your terminal, you instantly 00:09:48.470 --> 00:09:51.269 switch the brain powering your workspace, zero 00:09:51.269 --> 00:09:54.169 friction. That is brilliant. You open clod code 00:09:54.169 --> 00:09:56.330 in the glm folder for your rough drafts. You 00:09:56.330 --> 00:09:58.610 open it in the opus folder for your final code 00:09:58.610 --> 00:10:01.529 reviews. Exactly. But I have to ask a security 00:10:01.529 --> 00:10:04.769 question here. If I put my API key in a local 00:10:04.769 --> 00:10:07.570 JSON file, isn't there a risk I accidentally 00:10:07.570 --> 00:10:09.990 share it? Oh, there is absolutely a major risk 00:10:09.990 --> 00:10:12.149 there. That is why you must be extremely careful 00:10:12.149 --> 00:10:15.149 with this file. Right. Treat that API key exactly 00:10:15.149 --> 00:10:17.929 like a banking password. Keep it completely out 00:10:17.929 --> 00:10:21.289 of public repos or shared team screenshots. Treat 00:10:21.289 --> 00:10:23.389 the key like a password. Never push your local 00:10:23.389 --> 00:10:25.950 settings to the public. Exactly. Keep it secure 00:10:25.950 --> 00:10:28.649 and the entire workflow remains brilliant and 00:10:28.649 --> 00:10:30.750 safe. Now, we really need to look at the bigger 00:10:30.750 --> 00:10:34.710 picture here. Why does GLM 5 .2 make open source 00:10:34.710 --> 00:10:37.799 AI models impossible to ignore right now? Well, 00:10:37.980 --> 00:10:40.120 open source is finally practical for the daily 00:10:40.120 --> 00:10:42.820 professional workday. It is no longer just a 00:10:42.820 --> 00:10:45.299 theoretical weekend project for developers. The 00:10:45.299 --> 00:10:47.639 massive scale of this model is truly staggering 00:10:47.639 --> 00:10:51.799 to think about. GLM 5 .2 operates with around 00:10:51.799 --> 00:10:57.450 753 to 756 billion parameters. Yeah. Let's pause 00:10:57.450 --> 00:10:59.990 on that number for a second. That is far too 00:10:59.990 --> 00:11:02.230 massive to run locally on a normal computer. 00:11:02.509 --> 00:11:04.850 You would need serious, incredibly expensive 00:11:04.850 --> 00:11:07.470 server hardware in your house. That is exactly 00:11:07.470 --> 00:11:10.370 why API renting through providers is necessary. 00:11:10.529 --> 00:11:12.950 It is the pragmatic middle ground for users right 00:11:12.950 --> 00:11:15.009 now. You get all the massive power without the 00:11:15.009 --> 00:11:17.269 massive infrastructure headache. This brings 00:11:17.269 --> 00:11:20.610 us to the underlying cost of tokens. Tokens are 00:11:20.610 --> 00:11:23.309 just tiny puzzle pieces of text. It is kind of 00:11:23.309 --> 00:11:26.370 like stacking logo blocks of data. Exactly. The 00:11:26.370 --> 00:11:28.929 AI processes these tiny pieces to understand 00:11:28.929 --> 00:11:31.090 and generate language. When you look closely 00:11:31.090 --> 00:11:33.250 at the token pricing, the math is undeniable. 00:11:33.870 --> 00:11:37.250 Right. Opus costs $5 in and $25 out per 1 million 00:11:37.250 --> 00:11:42.070 tokens. GLM sits at roughly $1 .40 in and $4 00:11:42.070 --> 00:11:46.789 .40 out. Two secs silence. Whoa. I mean, imagine 00:11:46.789 --> 00:11:48.870 scaling to a billion queries without bankrupt. 00:11:48.840 --> 00:11:52.480 yourself. That five times price difference changes 00:11:52.480 --> 00:11:55.480 everything about how we build software. It is 00:11:55.480 --> 00:11:57.919 especially critical for complex agent workflows. 00:11:58.480 --> 00:12:01.259 These autonomous models constantly read folders, 00:12:01.759 --> 00:12:04.379 revise their own work, and call subagents. They 00:12:04.379 --> 00:12:06.639 burn tokens invisibly in the background while 00:12:06.639 --> 00:12:09.580 you drink your coffee. Exactly. Significantly 00:12:09.580 --> 00:12:12.019 lower prices encourage you to experiment freely 00:12:12.019 --> 00:12:14.759 without watching the meter. There are also performance 00:12:14.759 --> 00:12:18.350 benchmarks to consider. GLM actively beats GPT 00:12:18.350 --> 00:12:22.309 5 .5 and Opus 4 .8 on some specific software 00:12:22.309 --> 00:12:25.029 benchmarks. Right, but there is a very crucial 00:12:25.029 --> 00:12:27.210 point about those numbers. Benchmarks are just 00:12:27.210 --> 00:12:29.580 a signal. They simply tell you a model is worth 00:12:29.580 --> 00:12:32.820 testing yourself. Daily usefulness on your own 00:12:32.820 --> 00:12:35.759 specific files is the only real test. Yeah, how 00:12:35.759 --> 00:12:37.879 does it handle your unique code and your messy 00:12:37.879 --> 00:12:40.279 personal notes? There is also a major strategic 00:12:40.279 --> 00:12:42.539 advantage we should point out here. It acts as 00:12:42.539 --> 00:12:45.639 a necessary hedge. Closed platforms change their 00:12:45.639 --> 00:12:47.700 rules and their pricing structures all the time. 00:12:47.879 --> 00:12:51.100 A vital feature can suddenly move behind an expensive 00:12:51.100 --> 00:12:53.559 paywall tomorrow. Learning how to route open 00:12:53.559 --> 00:12:55.879 source models protects your daily workflow from 00:12:55.879 --> 00:12:58.519 those sudden shifts. But think about this dependency 00:12:58.519 --> 00:13:01.360 for a moment. If open source models are this 00:13:01.360 --> 00:13:04.820 huge, won't we always be dependent on cloud providers 00:13:04.820 --> 00:13:07.830 to run them anyway? Yes. For now, that is the 00:13:07.830 --> 00:13:10.730 reality. Yes. API rental is the necessary middle 00:13:10.730 --> 00:13:13.529 ground today. Right. But it still fundamentally 00:13:13.529 --> 00:13:16.149 breaks the monopoly. You aren't relying entirely 00:13:16.149 --> 00:13:19.870 on one single closed ecosystems rules anymore. 00:13:20.059 --> 00:13:22.779 It's about having a backup plan when the closed 00:13:22.779 --> 00:13:25.940 platforms change their rules. Exactly. You maintain 00:13:25.940 --> 00:13:28.200 your professional options and your creative freedom. 00:13:28.460 --> 00:13:30.100 So let's bring all of these different pieces 00:13:30.100 --> 00:13:32.820 together for you. We are looking at a fundamental 00:13:32.820 --> 00:13:34.960 shift in how we approach knowledge work. The 00:13:34.960 --> 00:13:37.820 era of treating open source models as clunky 00:13:37.820 --> 00:13:40.850 budget alternatives is officially over. The true 00:13:40.850 --> 00:13:43.789 modern skill isn't finding the one perfect AI 00:13:43.789 --> 00:13:46.049 to do everything. You are essentially becoming 00:13:46.049 --> 00:13:48.149 an orchestrator of multiple minds. Right. You 00:13:48.149 --> 00:13:50.389 have to know exactly when to deploy a fast, cheap 00:13:50.389 --> 00:13:54.230 worker. You use GLM 5 .2 for the heavy, repetitive 00:13:54.230 --> 00:13:56.129 lifting and the rough drafting. And you must 00:13:56.129 --> 00:13:58.769 know exactly when to bring in the expensive protectionist. 00:13:59.029 --> 00:14:01.870 You call on Opus 4 .8 to close the deal and ensure 00:14:01.870 --> 00:14:04.730 total precision. It is a beautiful synergy when 00:14:04.730 --> 00:14:06.590 you set it up correctly. We want you to look 00:14:06.590 --> 00:14:10.100 at your own daily tasks today. Identify the 80 00:14:10.100 --> 00:14:12.940 % work you are currently doing manually. Where 00:14:12.940 --> 00:14:15.419 are you overpaying for AI compute right now? 00:14:15.879 --> 00:14:18.899 Where could a fast, highly capable open source 00:14:18.899 --> 00:14:21.740 model handle the drafting and sorting for you? 00:14:22.279 --> 00:14:24.720 Finding that exact balance will completely change 00:14:24.720 --> 00:14:27.740 your daily productivity. Thank you for joining 00:14:27.740 --> 00:14:30.460 us on this deep dive today. We always appreciate 00:14:30.460 --> 00:14:32.840 you spending your valuable time with us. It has 00:14:32.840 --> 00:14:35.629 been a truly fantastic exploration today. We 00:14:35.629 --> 00:14:37.809 want to leave you with one final provocative 00:14:37.809 --> 00:14:40.429 thought to mull over. If an open source model 00:14:40.429 --> 00:14:43.230 is already handling 80 % of the workflow today 00:14:43.230 --> 00:14:45.830 at a fraction of the cost, what happens to the 00:14:45.830 --> 00:14:48.309 value of closed models when open source naturally 00:14:48.309 --> 00:14:51.830 creeps up to cover 90 or 95 %? Does intelligence 00:14:51.830 --> 00:14:53.230 become essentially free?