Fable 5 + GPT 5.6 Sol = CHEAT CODE — Transcript

Full transcript

1. 0:00GPT 5.6 aka Soul is coming out tomorrow
2. 0:04and the big question on everyone's mind
3. 0:05is does this new model beat Claude
4. 0:08Fable? Well, I think that's the wrong
5. 0:10question to ask because instead of
6. 0:12trying to figure out which one of these
7. 0:13models is better, we should be asking
8. 0:16how can we use these two powerful models
9. 0:18together? And in today's video, I'm
10. 0:20going to be giving you a skill that does
11. 0:21exactly that. This skill includes a
12. 0:24supercharged plan mode based on Matt
13. 0:25PCO's Grill Me. We then have an
14. 0:27adversarial planning session where
15. 0:29Claude and Codex go headto-head till
16. 0:30they come to a conclusion. Then once
17. 0:33we're ready, we take that Fabled driven
18. 0:36plan and hand it off to Codeex, which
19. 0:38starting tomorrow will include Soul 5.6.
20. 0:42After Codeex goes to work, then we have
21. 0:44Fable review the entire process. All in
22. 0:47all, we're not just getting the best of
23. 0:48both of these models, we're also saving
24. 0:50tokens in the aggregate versus having
25. 0:52Fable do everything. So, I'm going to
26. 0:54show you how this works. We're going to
27. 0:55do a quick demo and then I'll be giving
28. 0:57you the skill. So why should we even
29. 0:58care about creating some sort of skill
30. 1:00where Fable does most of the planning
31. 1:01and then we pass it off to Soul 5.6?
32. 1:04Well, first reason is Soul 5.6 is wildly
33. 1:08powerful, at least according to the
34. 1:10benchmarks. Now, grain of salt, this is
35. 1:12coming from OpenAI, but when we look at
36. 1:14Soul 5.6 Ultra and just standard 5.6 six
37. 1:18soul. We see numbers on Terminal Bench
38. 1:202.1 that put it ahead of Claude Mythos,
39. 1:23let alone Fable 5. The second reason is
40. 1:25token efficiency. There is a real reason
41. 1:28why you see so much content around, hey,
42. 1:30how can we reduce Fable's usage and the
43. 1:33types of things you see are like advisor
44. 1:35mode. You know, essentially having Fable
45. 1:37plan and have Opus execute. Well, why
46. 1:39have Opus execute if with at the same
47. 1:41price I could have 5.6 do it or 5.5 do
48. 1:45it? Point is, we're doing that same
49. 1:47construct but with a better model and
50. 1:50arguably a cheaper model than Opus. When
51. 1:52we look at 5.6, it is more token
52. 1:55efficient than 5.5 which was more token
53. 1:57efficient than Opus 4.6. And we can see
54. 2:00that in the data. What we're looking at
55. 2:01right here is GPT 5.5 on extra high. Its
56. 2:05pass rate on this benchmark was 23% at
57. 2:09$1.24. When I look at 5.6, six, it's 25%
58. 2:15score, so higher score at 56. So way
59. 2:18cheaper and therefore way more
60. 2:20efficient. And when we look at direct
61. 2:21comparisons of 5.5 versus Opus 4.8,
62. 2:24there's really no contest. Higher pass
63. 2:27rates, lower cost. So we're essentially
64. 2:30taking that same idea and just ramping
65. 2:32it up with this 5.6 improvement. So how
66. 2:34does this skill actually work? Well, I
67. 2:35actually have a couple skills for you.
68. 2:37In a vacuum, we have the codeex build
69. 2:40skill. This is the idea that you created
70. 2:41a plan with Fable and Codex is just
71. 2:43going to go ahead and build that
72. 2:44particular feature or particular
73. 2:46product. I also have included an updated
74. 2:48grillme codeex. Now, I've done a video
75. 2:50on this skill before and what we've done
76. 2:52is we've added on this idea of GPT 5.6
77. 2:57actually going out there and building
78. 2:58things for us. And so, when we look at
79. 3:01the more comprehensive GMI codeex, which
80. 3:03is the big skill, it occurs in four
81. 3:06stages. The idea is you have some sort
82. 3:09of project, some sort of feature you
83. 3:10want to start and you kick it off with
84. 3:12grill me codeex and the first thing that
85. 3:13happens is an interview. This interview
86. 3:16is literally the grill me skill from
87. 3:18Matt PCO. So it is a plan mode on
88. 3:20steroids. It goes way way deeper than
89. 3:22Claude Code normally would. And we do
90. 3:24this with Fable. All right. So Fable's
91. 3:27driving the ship here. Secondly, we have
92. 3:29adversarial planning. So Fable's come up
93. 3:32with a plan. We then take that Fable
94. 3:34plan and we push it over to Codeex. Now,
95. 3:38in today's video, that's going to be
96. 3:395.5, but tomorrow that will be 5.6. And
97. 3:42Fable and Codex go back and forth for a
98. 3:45maximum of five iterations where Fable
99. 3:47says, "Hey, here's the plan." Codex
100. 3:48says, "Okay, looks good. Accept X, Y,
101. 3:50and Z." Then Fable says, "Uh, I agree, I
102. 3:53disagree." And they go back and forth
103. 3:54till they reach a consensus. Now once
104. 3:56they reach that consensus and this is
105. 3:58where the upgrades happen is we now push
106. 4:01the actual build to codeex to 5.5 today
107. 4:04and 5.6 tomorrow. I think this is way
108. 4:07better than passing things off to Opus
109. 4:10or to Sonnet or using advisor mode
110. 4:12inside of cloud code because these GPT
111. 4:14models are just better than those
112. 4:16smaller anthropic models and they are
113. 4:18cheaper. So, it really is a scenario
114. 4:21where unless you just are super anti-GPT
115. 4:23and anti-codex, it's hard to argue
116. 4:25otherwise, especially if we get to a
117. 4:28place where Fable is like kind of off
118. 4:29the market. And lastly, once Codeex
119. 4:31finishes the build, Fable is going to
120. 4:33come in and it's going to review what it
121. 4:35did and it's going to go through a
122. 4:36maximum of two sort of iterations where,
123. 4:39let's say, Fable thinks Codex did
124. 4:41something wrong. It's going to say,
125. 4:42"Hey, Codex, you did that wrong. Fix
126. 4:43it." It's going to do that twice. If by
127. 4:45the third time it's not complete, well
128. 4:47then Fable will clean it up itself. So
129. 4:49this is the process by which I think we
130. 4:52get the best of OpenAI and Anthropic.
131. 4:55Now before we hop into the demo, a quick
132. 4:57word from today's sponsor, me. So I just
133. 4:59released my Cloud Code master class
134. 5:00inside of Chase AI plus and it is the
135. 5:02number one way to go from zero to AI
136. 5:04dev, especially if you don't come from a
137. 5:05technical background. I update this
138. 5:07every single week. We focus on real use
139. 5:10cases and it also includes a codeex
140. 5:12masterass as well. So, if you want to
141. 5:15get a little bit more serious about AI
142. 5:16and you have no idea where to begin,
143. 5:18this is the place for you. There will be
144. 5:20a link in the pin comment. Now,
145. 5:21installing and using the skill is pretty
146. 5:23straightforward. I will put a link to
147. 5:24the GitHub in the description. Now, to
148. 5:26use this, we're just going to do
149. 5:27for/grill codeex. And we just give it
150. 5:30our prompt what it is we're trying to
151. 5:31build. So, we're trying to build trip
152. 5:33atlas, which is a stylized cinematic
153. 5:35trip planner web app. And I go into a
154. 5:37little bit more details about what I
155. 5:38want it to be, right? I want it to look
156. 5:39kind of cool. I can put in the different
157. 5:41places I'm going, all that. And once I
158. 5:43do this, what's going to happen is it's
159. 5:44going to kick off the grill me section
160. 5:46of the plan, which if you're familiar
161. 5:47with Matt PCO's work, it essentially is
162. 5:50just a plan mode on steroids. It's going
163. 5:52to ask me like 8 n 10 different
164. 5:54questions. I go a lot deeper than your
165. 5:56standard plan mode stuff. So, it's
166. 5:57asking me what is this for? We're going
167. 5:59to say this is for a real personal tool,
168. 6:01not just a video demo. And for each of
169. 6:04these, it also gives its
170. 6:05recommendations. So, if you're confused
171. 6:07about what I should choose and why,
172. 6:08that's all spelled out for you. Now it's
173. 6:10asking about geocoding and it's going to
174. 6:12continue to go down these series of
175. 6:13questions until it's happy with what
176. 6:15we're creating. Now I'm going to skip
177. 6:17through the rest of the questions
178. 6:18because you can imagine what the next
179. 6:20seven or eight questions will look like
180. 6:21and we'll move into the adversarial
181. 6:24planning stage. So we can see here it's
182. 6:25written the plan and it also creates a
183. 6:27markdown file where it logs all the back
184. 6:29and forth between codeex and cloud code.
185. 6:32And so right now we are on round one
186. 6:34where it's passing it off to GBT. And so
187. 6:36you can see them kind of going back and
188. 6:38forth here on the log. But in this case,
189. 6:41it only took them two rounds before it
190. 6:44was approved. And so we can see what the
191. 6:46two acts improved. You know, lock the
192. 6:49identity, a real person tool, kind of
193. 6:51what's going to be the actual sort of
194. 6:53stack. And then it had 12 findings in
195. 6:56the second round related to like
196. 6:57hardening the data core. Now once it's
197. 6:59completed this back and forth, you have
198. 7:01a few options. either Codex is going to
199. 7:03build it. Kind of what we've talked
200. 7:05about from the beginning. We have the
201. 7:07option just having Claude build it. So
202. 7:08for whatever reason like I don't want to
203. 7:09bring GPT in, you can keep it with Fable
204. 7:12or you can stop here. But we're going to
205. 7:14go ahead and let Codeex build this. And
206. 7:16again, we can kind of go back and forth.
207. 7:18If you think, well, GPT 5.6 is going to
208. 7:20be better than Fable or 5.5 versus Opus
209. 7:234.8. At the end of the day, the real
210. 7:26value that can't really be argued with
211. 7:27is going to be the token efficiency,
212. 7:30especially if 5.6 6 is even close to
213. 7:33what the benchmarks are claiming. So
214. 7:35Codeex has finished up its build and you
215. 7:37can see now what's happening is the
216. 7:39review stage. So now Fable is going
217. 7:41through everything Codex has built and
218. 7:43then it's going to go back to Codex and
219. 7:45say this is wrong, this was right.
220. 7:46Remember it'll do two iterations of that
221. 7:48before it's like hey I want to drive.
222. 7:50It'll take the wheel and it will start
223. 7:52writing the code itself. Now Fable is
224. 7:54done with its review. It said there were
225. 7:56a couple deviations which it felt were
226. 7:58all reasonable. goes over the files and
227. 7:59all this and now it's asking, hey, do
228. 8:01you want to commit or you want to take a
229. 8:02look at it? So, let's take a look at
230. 8:03what it actually built. And so, here's
231. 8:05what we got. So, over here we have sort
232. 8:08of a map of the world and it looks like
233. 8:10it created some custom graphics using
234. 8:12the GPT image generator. So, you're able
235. 8:15to like name the trip, you can add
236. 8:17stops. Over here on the left, you can
237. 8:19put where you're going and then sort of
238. 8:21like what you're going to do at those
239. 8:23different locations. It also has this
240. 8:25cinematic replay. And I'm just going to
241. 8:27mute this. So, let's see what happens
242. 8:28here.
243. 8:31So, it looks like I'll move over here.
244. 8:34You can see sort of this weird plane
245. 8:36hopping from spot to spot, which it
246. 8:38looks like it created as an SVG. There's
247. 8:41a little passport stamps and boom, there
248. 8:44we go. So, you know, there's a lot we
249. 8:47could do here to kind of make it look, I
250. 8:49think, better, but in general, it built
251. 8:52what we said we wanted to, right? Like
252. 8:54everything actually works here. You
253. 8:56know, if I delete things on here, delete
254. 8:58some. I can move them up, down. I can
255. 9:01change stuff. Let's say we added Tokyo.
256. 9:03All of a sudden, it actually shows how
257. 9:06far away that stop is. That's
258. 9:07interesting. If I add that to the route.
259. 9:10There we go. So, you know, actually
260. 9:13built this out. I think it'd be a good
261. 9:15not bad, I think, for the first pass.
262. 9:17And what this really was about was just
263. 9:18showing this workflow in action. And you
264. 9:21can also see down here in terms of our
265. 9:22usage, we only burned up about 130,000
266. 9:25tokens on the fable side to get this
267. 9:27whole thing done. So that's the skill in
268. 9:28action. Hopefully you get a ton of use
269. 9:29out of this one's 5.6 drops. As always,
270. 9:32let me know what you thought about this
271. 9:34video in the comments. Make sure to
272. 9:35check out Chase AI Plus and I'll see you