GPT-6 Astra Tested: I Built a Carousel Generator From Scratch — Transcript

Full transcript

1. 0:00Today, I want to show you an app I've
2. 0:01been building with the latest GBT6 Astra
3. 0:03model from OpenAI and also talk about
4. 0:05why I think software still beats agents
5. 0:07for a lot of day-to-day work. AI agents
6. 0:09are incredibly powerful, but when it
7. 0:10comes to work where you need speed,
8. 0:12consistency, and quality, I still think
9. 0:13custom software can beat prompting
10. 0:15agents in a terminal. So, in [music]
11. 0:17this video, I'm going to show you a
12. 0:17custom app I've been building to solve a
13. 0:19painful problem in my own business using
14. 0:21GPD6, talk about what it's like to use
15. 0:23the [music] new Astra model day-to-day,
16. 0:25and also show you my favorite feature of
17. 0:27Codeex right now. If you stick around to
18. 0:28the end, I'll give away all of the
19. 0:29resources, and I'll link you to the
20. 0:31source code for this app if you want to
21. 0:32use it yourself. So, [music] let's get
22. 0:33into it. So, the painful problem for me
23. 0:35is starting out on social media. So, I
24. 0:37just started posting on Instagram, but
25. 0:39honestly, I have no followers at all.
26. 0:41And just like starting from nothing is
27. 0:43pretty overwhelming. And then the same
28. 0:44thing, I need to produce content on
29. 0:46LinkedIn. I have more followers over
30. 0:48here. I have 55,000 roughly, but I'm not
31. 0:50posting consistently anymore. I've
32. 0:52struggled to find the time. So, this is
33. 0:54the custom app I've been building. It's
34. 0:55called Vertica, and it helps me produce.
35. 0:57social media carousels essentially. So
36. 0:59the way this works is pretty
37. 1:00straightforward as you'd expect. You can
38. 1:02go ahead and create a new carousel. So
39. 1:04we can create a headline. So AI things
40. 1:07you should know. And then we can change
41. 1:10things like the layout. We can move
42. 1:12things around. We can center align the
43. 1:13text. All of these things are kind of
44. 1:15predefined. You can obviously toggle off
45. 1:17the headers. You can choose different
46. 1:19page layouts as well. So you can choose
47. 1:21body layouts or you can choose like CTAs
48. 1:24and things like that. And then you can
49. 1:26also change the design as well. So you
50. 1:28can change like the background colors.
51. 1:30You can upload images as well. So you
52. 1:32can choose images for the background.
53. 1:34You can do things like changing the
54. 1:36opacity here as well. You can also
55. 1:38export these as well. So you can click
56. 1:41up here and then you can download video
57. 1:43slides, JPEGs. If you want to upload
58. 1:45these to Instagram, you can also
59. 1:47download PDF documents. These are the
60. 1:49things that if you're using AI agents
61. 1:51can be really slow and painful to do.
62. 1:53But because we're just using custom
63. 1:54software and custom code to do this,
64. 1:56it's really quick and easy and
65. 1:57incredibly fast to do this. You can also
66. 2:00upload videos as well. So you could have
67. 2:01video backgrounds as well if you want
68. 2:03to. So overall, this just makes it
69. 2:05really quick and easy to start building
70. 2:07out carousels. And like I said, it gives
71. 2:09you the consistency that you wouldn't
72. 2:10necessarily get if you're using agent
73. 2:12skills. You can also obviously keep a
74. 2:15library of all of the things you're
75. 2:16working on. So this again is just really
76. 2:18handy. You've just got a much nicer user
77. 2:20interface. You've got your media library
78. 2:21if you want to remix different images
79. 2:23across different carousels. And it's
80. 2:25just a nicer user interface than kind of
81. 2:27just working with agents in a terminal
82. 2:29overall. So this is an example of a
83. 2:32video carousel. You can see here we've
84. 2:33got this background video generated
85. 2:35using Higsfield. And then we've got just
86. 2:38the usual slides, but we've also got the
87. 2:40ability to do things like these diagrams
88. 2:42as well. So the nice thing about HTML,
89. 2:44you can insert SVG diagrams. So this is
90. 2:47a carousel about the latest GBT6 model
91. 2:50from OpenAI. And we've pulled in some of
92. 2:52these benchmarks and some of these
93. 2:53graphics just as SVG diagrams. And you
94. 2:56can use AI agents to help you generate
95. 2:58these. So this is an example of where
96. 3:00you can combine the like the manual
97. 3:03software workflow with AI agents where
98. 3:05agents can generate parts of this for
99. 3:07you. They can also generate the entire
100. 3:09carousel as well. So what we're going to
101. 3:10do now is create a carousel using this
102. 3:13approach. So we've got this carousel
103. 3:15here. how to deliver $10,000 AI projects
104. 3:18and then a series of tips. So, what
105. 3:19we're going to do is ask codeex to use
106. 3:21browser use to go and upload these to
107. 3:23Vertica. So, can you go ahead and create
108. 3:28a new carousel using the Vertica
109. 3:32production app? So, you can see that
110. 3:34Codex is now driving the browser and you
111. 3:37can see here it's now creating a new
112. 3:38carousel. It's figuring out the best way
113. 3:40to do this. So, it's figured out that it
114. 3:42can use this JSON configuration. I have
115. 3:44a skill to help it do this, but Codeex
116. 3:47is just incredible at like interacting
117. 3:48with your browser. Doing this kind of
118. 3:50work with AI agents is just really,
119. 3:52really useful. So, you're combining the
120. 3:54best parts of software with the best
121. 3:55parts of agents in this example. So, the
122. 3:58new GP6 Astra model is pretty
123. 4:00incredible. I've been using Codeex for
124. 4:01the majority of my development work for
125. 4:02a really long time now, and I really
126. 4:04like that you can use your subscriptions
127. 4:06with any other coding harness. So, if
128. 4:08you want to use a different coding
129. 4:09harness, you can use your OpenAI models.
130. 4:11The pricing is incredible. The value for
131. 4:13the money is incredible and just the
132. 4:15models I think are the best around right
133. 4:17now. So GPT6 Astra is really really
134. 4:19great and you can kind of just see this
135. 4:22from the benchmarks, right? You can kind
136. 4:23of see already how much better that the
137. 4:27GPT6 model is compared to some of these
138. 4:29other models. And this kind of plays out
139. 4:31in my own testing of these models. I
140. 4:33don't really care about benchmarks too
141. 4:35much personally anymore cuz I find they
142. 4:37don't necessarily fully reflect reality.
143. 4:39But I just really enjoy using the GBT6
144. 4:41models. So, what's most interesting
145. 4:43about GP6 in my opinion is its computer
146. 4:45use. You can see here that the
147. 4:48benchmarks on computer use are really,
148. 4:50really impressive. The the ability to
149. 4:53control your computer and interact with
150. 4:55your browser, I think is one of the most
151. 4:56underrated features when it comes to AI
152. 4:58in general right now. I use browser use
153. 5:00all of the time. Whether that's like
154. 5:01doing QA on a website and testing things
155. 5:04in the browser or whether it's fully
156. 5:06automating parts of my workflow inside
157. 5:08the browser where there's no API
158. 5:09available. I'm using computer use and
159. 5:12the browser automations all of the time.
160. 5:15And it's one of those features that I
161. 5:16think doesn't get a lot of attention,
162. 5:17but it's really really important and
163. 5:19really really valuable. So what we're
164. 5:20going to try and do now is build a
165. 5:22carousel about GBT6 using the browser
166. 5:25use capabilities. So I'm going to go
167. 5:27over into codeex. So, I'm going to use
168. 5:29the voice dictation feature here to
169. 5:31build this out. So, you can see here
170. 5:33that now Codeex is using browser use to
171. 5:35go and check out the website. And it's
172. 5:38pulling out all of the information. It's
173. 5:40pulling out all of these benchmarks from
174. 5:41the GP6 site and then using that to
175. 5:44build out our carousel. So, you can do
176. 5:46this in two ways when it comes to
177. 5:47browser use. There's a Chrome plugin if
178. 5:49you want to drive the browser directly,
179. 5:51but you can also have Codeex run it
180. 5:53inside the Codeex desktop app. Both of
181. 5:56those are useful for different reasons.
182. 5:58I use both for different situations, but
183. 6:01here we're going to use the inapp
184. 6:03browser. So, I just logged in and now
185. 6:04Codeex is controlling the browser. You
186. 6:07can see here the icon up here. And now
187. 6:09you can see that Codex is in the browser
188. 6:12filling in the carousel
189. 6:15for us. So, it should be able to create
190. 6:17the slides. All right. All right. So,
191. 6:18now that we have the carousel built out,
192. 6:20what would be interesting to do in here
193. 6:21now is to change this static image to be
194. 6:23a video background. So, we're going to
195. 6:26try and do that now using browser use.
196. 6:28So, the way I'm going to do this is to
197. 6:29use a tool called hexfield.ai,
198. 6:32which is a video generator. And let's
199. 6:35see if Codex can actually do this in the
200. 6:37browser. Okay, so it looks like the
201. 6:39video is finally generated. That took a
202. 6:40little bit of time, but now you can see
203. 6:42we have this video. So what we can ask
204. 6:46Codex to do now is take that video and
205. 6:48then upload it to the carousel. So now
206. 6:50what you can see we have the finished
207. 6:51carousel. So Codex did all of this for
208. 6:54us. So again, we're just combining
209. 6:56custom software with the browser use
210. 6:59capabilities plus the agents to help us
211. 7:01fill in the carousels. So we're still in
212. 7:04control of everything. We're still able
213. 7:05to edit things by hand and go really,
214. 7:07really fast. If we now want to download
215. 7:09this and export it, again, it's very,
216. 7:10very fast. We're not relying on agents
217. 7:12to do this kind of stuff for us because
218. 7:14we have the benefits of traditional
219. 7:16software being controlled by agents
220. 7:19using browser use. I think this is like
221. 7:20a killer combination for productivity.
222. 7:23Again, I would much rather use a custom
223. 7:25tool like this to do this workflow than
224. 7:28I would to be chatting with Claude in
225. 7:29the terminal, for example. It wouldn't
226. 7:31be anywhere near as easy or as like
227. 7:33enjoyable to build out these carousels
228. 7:36in the terminal as it would be if we had
229. 7:38custom software. So if you're interested
230. 7:40in building something like this for
231. 7:42yourself, the most interesting thing to
232. 7:44think about is the actual architecture
233. 7:45of the project itself. So this is a
234. 7:48relatively straightforward application
235. 7:49that I built using GPC6.
236. 7:52And the most difficult thing to think
237. 7:54about is always the architecture when it
238. 7:56comes to these kind of software
239. 7:57projects. So if you're using this in a
240. 7:59team or if you use it as part of your
241. 8:00workflow, you probably want to deploy it
242. 8:01to the cloud. But then you have to think
243. 8:03about the architecture. So because I'm
244. 8:05using Google Cloud which has ephemeral
245. 8:07containers, I needed a way to store all
246. 8:09of the data. So I'm using object storage
247. 8:12to persist all of the metadata, the
248. 8:14files, the images, the the the video
249. 8:18files and things like that and the PDFs.
250. 8:20And then I'm using a library called
251. 8:22FFmpeg which is used to take the MP4
252. 8:26files that we generate and then overlay
253. 8:27text on top of them. So you saw the
254. 8:30video carousels. And in order to
255. 8:31generate those, we need to use this
256. 8:32ffmpeg library to actually do that. But
257. 8:35other than that, this is a relatively
258. 8:37straightforward basic like JavaScript
259. 8:39react application. So overall, my main
260. 8:41takeaway from this experiment is the
261. 8:43power of building custom software apps
262. 8:45over just building AI agents in your
263. 8:47terminal. I think having skills and
264. 8:50talking to agents in the terminal is
265. 8:51useful dayto-day, but it's also for me
266. 8:53becoming a bit of a bottleneck. I would
267. 8:55much rather use a specific software
268. 8:57application designed to solve a painful
269. 8:59problem and just use that because it's
270. 9:01way more repeatable. Software tends to
271. 9:03be a lot faster and just a lot more
272. 9:05reliable and it just kind of guides
273. 9:06[music] you through the workflow.
274. 9:07Whereas I find when I'm using AI agents,
275. 9:09I just get distracted. I find skills get
276. 9:11lost. Things are scattered all over the
277. 9:12place and I just find it's not a very
278. 9:14nice way to work. So, I would encourage
279. 9:16you to think about how to build custom
280. 9:18software apps like this inside your
281. 9:19business cuz I think it's way more
282. 9:21effective [music] to have this kind of
283. 9:22custom software that you can have your
284. 9:25agents interact with than it is to build
285. 9:27custom [music] skills and then just talk
286. 9:29to agents in the terminal all the time.
287. 9:30I just don't think it's the best
288. 9:31experience for me personally. I think
289. 9:34the best businesses going forward are
290. 9:35going to combine custom [music] software
291. 9:37with AI agents. AI agents in my opinion
292. 9:40are not enough on their own. And talking
293. 9:41to clawed code in a terminal is not a
294. 9:44good user interface for many types of
295. 9:45work. And it's so easy to get
296. 9:47distracted. It's very hard to sort of
297. 9:49make sense of what's going on when
298. 9:51you're jumping between different
299. 9:52conversations in a terminal or in an app
300. 9:54like Codeex. It's not the most
301. 9:56productive way to get work done in my
302. 9:58opinion. And I would much rather have
303. 10:00custom software that I can log into to
304. 10:02do a particular task.