YouTube2Text

5 Çin Modeliyle Krallık Kurdum: Sıfırdan Kurulum — Transcript

by Selma Kocabıyık · 2,150 words · 310 segments · language en · Watch on YouTube

Full transcript

  1. 0:00This time, I’ve built a kingdom of
  2. 0:02models, and inside each of these castles
  3. 0:03you see, a Chinese model resides.
  4. 0:06These models run 24/7, all designed to
  5. 0:08work with each other, and in this video
  6. 0:10, I will explain how I did it from start
  7. 0:13to finish and even share their
  8. 0:14performance with you.
  9. 0:16If you remember, I had previously
  10. 0:18created an agent company. I had also set
  11. 0:20up an agent office consisting of cats.
  12. 0:23Now, I wanted to add an extra kingdom
  13. 0:25consisting solely of Chinese models to
  14. 0:27that. I do most of my work with Cloud
  15. 0:28Code, but the limits of the most
  16. 0:29powerful models aren't enough for every
  17. 0:31task. We agree on this. For tasks like
  18. 0:33writing cards, fixing text, or setting
  19. 0:35up a simple page, we don't always need a
  20. 0:37powerful model. I don’t want to spend
  21. 0:39money on the most expensive model for
  22. 0:40these. Because I will do the truly
  23. 0:42difficult tasks locally in Cloud Code.
  24. 0:43But I will set up the architecture for
  25. 0:45my Chinese models on a server. This way
  26. 0:47, they will run 24/7 on the server, and
  27. 0:50I can direct the routine tasks I want to
  28. 0:52assign straight to the team on the
  29. 0:54server. I also connected Cloud Code to
  30. 0:56control the entire system.
  31. 0:59So, even if I don't go into the server,
  32. 1:00I can control it as I wish from the
  33. 1:02terminal.
  34. 1:03Now, you know there is a tool called
  35. 1:05OpenRouter; you can use hundreds of
  36. 1:07models with a single key, and the
  37. 1:09platform even shows who is using which
  38. 1:11model and how much. For instance, in
  39. 1:13last week's list, seven out of the 12
  40. 1:16most used models were Chinese models.
  41. 1:18DeepSeek is in first place, of course,
  42. 1:20and I chose exactly five from this list
  43. 1:22. Five models are quite enough for me.
  44. 1:24By the way, I gave each of them a job.
  45. 1:25This Qwen 3.1 Pro is the most expensive
  46. 1:28and slightly better than the other five
  47. 1:30. That’s why I appointed it as the
  48. 1:32planner and the supervisor, in other
  49. 1:34words, the main kingdom. Then we have
  50. 1:36our GLM 5.3. It was built for coding, so
  51. 1:39we will have GLM do the main work. Now
  52. 1:41we have our DeepSeek V 4.1 Flash model.
  53. 1:44One of the most used and cheapest. For
  54. 1:46example, it gets assigned tasks that
  55. 1:48process a lot of text. This could be
  56. 1:49research, reading a PDF, or preparing
  57. 1:52specific cards.
  58. 1:53Now we also have MiniMax, you know, the
  59. 1:55M 3.1 Flash. I would like to state that
  60. 1:57this video definitely does not contain
  61. 1:58any sponsorship. Now, this Flash model
  62. 2:00acts as a second pair of eyes to check
  63. 2:02the cards. So, this is the one doing the
  64. 2:04tests, but who does the final audit? We
  65. 2:06also have Coder Next. This is also a
  66. 2:09code-focused and cheap model. It fixes
  67. 2:11their errors as well. Now we see five
  68. 2:13models and five separate companies; the
  69. 2:16writer and the supervisor won't be from
  70. 2:18the same company. By the way, there are
  71. 2:19also two new models that came out. We
  72. 2:21didn't choose them. I'm explaining the
  73. 2:22reason, too. One is Tencent's Hunyuan
  74. 2:24model and the other is Alibaba's Max
  75. 2:27model.
  76. 2:28Both of these only work on their own
  77. 2:30servers. So, any prompt I give or
  78. 2:32information I write goes directly to
  79. 2:34their company's server. This is actually
  80. 2:36not that big of a problem.
  81. 2:38For instance, the other five models we
  82. 2:39chose work perfectly fine on different
  83. 2:41companies 'servers as well. By the way,
  84. 2:43I had Fable research this. Fable also
  85. 2:45decided not to use them due to security
  86. 2:47reasons. But ultimately, since they were
  87. 2:49running within a server, they would have
  88. 2:50been working in a fairly secure area.
  89. 2:52Still, they made that kind of choice. I
  90. 2:54wanted to point that out. You can share
  91. 2:55your thoughts in the comments section.
  92. 2:57Now, my second decision is this. These
  93. 2:59models will not be running on my
  94. 3:00computer. They will run inside the
  95. 3:02server. Now, why did we choose a server
  96. 3:04?
  97. 3:04Well, I have my projects, files, and
  98. 3:06passwords on my computer, and I don't
  99. 3:07want to give a model I define full
  100. 3:09access to them.
  101. 3:10Moreover, there is no personal
  102. 3:11information inside the server. In the
  103. 3:13event of a potential bad package
  104. 3:14download or a virus entering the server
  105. 3:16, I will notice it and clear it
  106. 3:18immediately. My agents are in a separate
  107. 3:20box inside the server. The models
  108. 3:23themselves live in the cloud, and the
  109. 3:25agents on the server can call the models
  110. 3:27from there. The box logic is as follows
  111. 3:28.
  112. 3:29The program inside only works with the
  113. 3:31files in that box and cannot reach your
  114. 3:32files, but it can access the internet.
  115. 3:35It is open to the internet.
  116. 3:36Because it needs to call these models,
  117. 3:38we didn't put any restrictions on its
  118. 3:40internet access.
  119. 3:41Let's get to the part about how I
  120. 3:43connected them. Hermetic is running
  121. 3:45inside your server, and as you know,
  122. 3:46Hermetic is an agent program.
  123. 3:48You can open certain distinct profiles
  124. 3:50within Hermetic, and there is a separate
  125. 3:51agent in each profile; I opened five
  126. 3:53profiles here and connected a model to
  127. 3:55each one, and for starters, I just put
  128. 3:5610 dollars under that program. I went to
  129. 3:59the site, got a key, and set a spending
  130. 4:00limit. That's it. It's quite simple. We
  131. 4:03can manage five models with a single key
  132. 4:05. Now, there is a task board within
  133. 4:06Hermes, and you can divide the work into
  134. 4:08tasks from there and assign a profile to
  135. 4:10each task. And I actually write the
  136. 4:12commands, or prompts, from this screen.
  137. 4:15And inside this server, we placed
  138. 4:17certain packages that models can use,
  139. 4:19capability files, and also a separate
  140. 4:21explanation file so they recognize each
  141. 4:23other.
  142. 4:24Think of it like a rule file. And I did
  143. 4:26all of this with the local Cloud Code.
  144. 4:28I gave my server information to the
  145. 4:30local Cloud Code, and it set up the
  146. 4:31architecture inside. But of course,
  147. 4:32Cloud is not inside the team. It works
  148. 4:34independently within the server. Now let
  149. 4:36me give you a tour of the kingdom. By
  150. 4:37the way, I just made it. I created the
  151. 4:39new interface. Since it's new, the
  152. 4:41information on the right side might not
  153. 4:43have been updated yet. Now let's see how
  154. 4:45the Taht (Throne) thing works. For
  155. 4:46example, when I give a prompt like "do
  156. 4:49research," this prompt is forwarded to
  157. 4:51the correct agent. But who does all the
  158. 4:53checking, planning, and overseeing? It's
  159. 4:55Kimy. That's why Kimy usually provides
  160. 4:57guidance when we give it a task.
  161. 5:00Now, when I click on Kimy, I can access
  162. 5:02its rule file and terminal on the left
  163. 5:04side.
  164. 5:06Now, when I write anything in the chat,
  165. 5:08Kimy answers me automatically from here
  166. 5:10. The answer comes directly from that
  167. 5:12model. So, I can talk to any agent I
  168. 5:13want one by one. Now let's talk about
  169. 5:15what you can do with a system like this
  170. 5:16. First of all, this is a 24/7 operating
  171. 5:18system.
  172. 5:19It means it can work however it needs to
  173. 5:21when it connects to any external network
  174. 5:23from your Telegram, Discord, or WhatsApp
  175. 5:24.
  176. 5:26You're out and about, you want to
  177. 5:27perform a task, but you don't want to
  178. 5:29use Claude Code. In that case, for
  179. 5:30instance, I can write directly to this
  180. 5:31system. "Do some research for me."
  181. 5:33Competitor analysis, for example, is
  182. 5:34very good for this. Also, these are
  183. 5:36Chinese models. They are quite cheap. So
  184. 5:37, I know very well that they won't be
  185. 5:39able to run again once my limit within
  186. 5:41OpenRouter is reached. It also has
  187. 5:43another great feature. Let's say I
  188. 5:44didn't like Claude's performance. Then I
  189. 5:46can replace Claude with any model I want
  190. 5:48. Now let me explain a use case for this
  191. 5:50system, or how I use it myself. It works
  192. 5:52like this. I can provide any prompt I
  193. 5:54want here again. I can write a command
  194. 5:56in the chat section. I open one of these
  195. 5:57and say, "Do some research for me." For
  196. 6:00example, I say, "Go research the most
  197. 6:02viral posts on Twitter in the last 24
  198. 6:03hours for me." They don't know my
  199. 6:05account details on Twitter, for example
  200. 6:06. They don't know anything about me.
  201. 6:08Now, when we consider the architecture,
  202. 6:10even if I give the prompt, since Kimi is
  203. 6:12the CEO, the main agent, it
  204. 6:14automatically forwards the task to the
  205. 6:16agent that needs to fulfill it.
  206. 6:18Now I actually gave this prompt. Our
  207. 6:20prompt went to Kimi. Kimi split the job
  208. 6:22into four parts and then distributed
  209. 6:23them. It makes the decisions. You can
  210. 6:25think of it like a referee. DeepSeek
  211. 6:27went on the internet. It found 14 topics
  212. 6:29in one minute. Minimax opened these
  213. 6:31topics and their sources one by one to
  214. 6:32check them, and the total took four
  215. 6:34minutes. Then CLM wrote a report, and
  216. 6:36that took a minute and a half. Then Kimi
  217. 6:39audited the report and found five errors
  218. 6:41. For example, the dates of two sources
  219. 6:43were wrong, and Qwen corrected them
  220. 6:44afterward. Now I say open the result,
  221. 6:46and as you can see, the report is in
  222. 6:48front of me. 14 topics and 57 sources.
  223. 6:51By the way, I also checked the tests and
  224. 6:53results again with Claude Code to see if
  225. 6:54they did it correctly or not.
  226. 6:56It also confirmed that the sites were
  227. 6:58correct. In total, this process took 18
  228. 7:01minutes and cost 70 cents according to
  229. 7:03Hermes's calculation.
  230. 7:05So, I can say it's a great team effort
  231. 7:07for this research and daily routine
  232. 7:08tasks.
  233. 7:09I say it can use any MCP it wants. It
  234. 7:11can use any skill file it wants. This is
  235. 7:13very important because I don't download
  236. 7:15every GitHub repo to my local machine,
  237. 7:16for example. But these Chinese models of
  238. 7:18mine can download any GitHub repo they
  239. 7:20want. They can install any skill file
  240. 7:22they want. I will show you an example,
  241. 7:23for instance. There was a GitHub repo
  242. 7:25called Limo Opus Car. I had previously
  243. 7:27prepared a Reel for it on Instagram. It
  244. 7:29has great film styles in it, and your
  245. 7:31model knows how to produce video from it
  246. 7:33. I take the GitHub repo. I come to my
  247. 7:36local terminal here. I provide this
  248. 7:37GitHub link, especially noting that it's
  249. 7:39inside my Chinese models, and I say, "
  250. 7:41Install this for me.""Look at its
  251. 7:42content.""Using this, produce faceless
  252. 7:45content for me."
  253. 7:46I am looking at our Limo Opus Car
  254. 7:48results right away. I am monitoring the
  255. 7:50performance of all of them here. How
  256. 7:51much did they spend, for example, how
  257. 7:53much money did they spend setting up
  258. 7:54this GitHub repo? How many tokens did
  259. 7:56they consume while analyzing its content
  260. 7:58? What was the return on this? The team
  261. 8:00set up this repo inside the container
  262. 8:01themselves, and nothing was downloaded
  263. 8:03to my computer. It even produced a film
  264. 8:04, but the result isn't very good. You
  265. 8:06can see it for yourselves. It is already
  266. 8:08written on its page. This was originally
  267. 8:10designed for the Cloud. It states that
  268. 8:12other models might not produce good
  269. 8:13results. And that is exactly what
  270. 8:15happened.
  271. 8:15So, for setup and research, this team is
  272. 8:17very good at the job. It is also
  273. 8:18perfectly okay for scripts or perhaps
  274. 8:20different types of content production.
  275. 8:22But for this kind of film work or
  276. 8:24editing tasks, Cloud is still in the
  277. 8:26lead, as always. Now, let's have it do
  278. 8:28one with this rule. After recording the
  279. 8:31audio, I put it on Drive and gave the
  280. 8:33link to the team. Afterwards, I wrote
  281. 8:36the prompt as a command from here.
  282. 8:38Vertical, from this audio, in a
  283. 8:40powerless way. First, who distributed
  284. 8:42the work? The team worked for 25 minutes
  285. 8:45, and this is what they made. The
  286. 8:47subtitles appear at the same time as my
  287. 8:49voice. There is no video model. They did
  288. 8:51it all with just code, and it cost 1
  289. 8:53dollar and 10 cents. So, if you guide
  290. 8:55them correctly, you can get things done
  291. 8:57with these teams. Of course, the final
  292. 8:59review part can be done on your side,
  293. 9:01meaning locally. After all, the sharing
  294. 9:02will also be done locally. By the way, I
  295. 9:04think this kind of architecture can be
  296. 9:05improved. You can even add more models
  297. 9:08or use 2-3 different servers at once. I
  298. 9:10used a VPS as the server and utilized
  299. 9:13what was inside. By the way, this video
  300. 9:15is not sponsored. If you want, you can
  301. 9:17find the link in the description section
  302. 9:19.
  303. 9:19By the way, if there are other models
  304. 9:21you want me to try, or if there are
  305. 9:22parts where you think "you should have
  306. 9:23added this" to the architecture, you can
  307. 9:25mention them in the comments section.
  308. 9:26I will put the details regarding the
  309. 9:28video in the description. See you in the
  310. 9:29next video.

About this transcript

This page contains the full transcript of 5 Çin Modeliyle Krallık Kurdum: Sıfırdan Kurulum by Selma Kocabıyık, generated from the public captions YouTube serves with the video. The transcript has 2,150 words across 310 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.