YouTube2Text

GPT 5.6 SOL y NUEVO KIMI 3 ¡La IA AGÉNTICA ya es REAL! — Transcript

by Dot CSV Lab · 6,039 words · 861 segments · language en · Watch on YouTube

Full transcript

  1. 0:00As usual, every time I go on vacation,
  2. 0:02the world of artificial intelligence
  3. 0:04doesn't stop, in fact, it accelerates.
  4. 0:07And as you know, in recent weeks we've
  5. 0:09had a lot of news, and today we're
  6. 0:10going to analyze the most important
  7. 0:12ones. Specifically, the release of a
  8. 0:16new constellation of models, Sun, Earth
  9. 0:18and Moon by Open AI, or for example, by
  10. 0:21Antropic, the unlocking of Fable 5 that
  11. 0:23we could finally use at least until
  12. 0:25July 7. Among other news, we've also
  13. 0:27seen the release of the new voice model
  14. 0:30of Chat GPT, quite interesting, or for
  15. 0:32example, by Antropic the extension of
  16. 0:36the Fable 5 model from July 7 to July
  17. 0:3912. We've also seen how more pressure
  18. 0:42has been added by competitors such as
  19. 0:44Exeyi with its new Grock 4.5 model or
  20. 0:47the Spark 1.1 model by Meta. Or for
  21. 0:51example, we've also seen by Antropic
  22. 0:53the extension of the Fable 5 model
  23. 0:55until July 19. One of the biggest
  24. 0:57bombshells in open source models so far
  25. 1:00this year, with the release of the new
  26. 1:02Kimi 3 model, which represents for the
  27. 1:04first time an open source model with
  28. 1:072.8 trillion parameters, a huge model
  29. 1:09that surpasses what the cutting-edge
  30. 1:11models of a few months ago were and is
  31. 1:13now competing very closely with the new
  32. 1:16flagship models from Open AI and
  33. 1:18Antropic. Or news from Antropic,
  34. 1:21confirming the availability of Fable 5
  35. 1:24from July 20th in subscriptions.
  36. 1:26They're driving us crazy. And well,
  37. 1:28then news from Google that, oh, no,
  38. 1:31Google hasn't released any model, it
  39. 1:33hasn't released Yemini 3.5 Pro. Well,
  40. 1:36as you can see, a few weeks full of
  41. 1:37very current events, lots of new
  42. 1:39features that we're going to be
  43. 1:40analyzing today through the reflections
  44. 1:42, conclusions that I've been able to
  45. 1:44build after testing these models during
  46. 1:46the last week. And I'm going to show
  47. 1:48you some use cases that have blown my
  48. 1:50mind and that demonstrate the maturity
  49. 1:52that these models have already achieved
  50. 1:54and that I believe mark a before and
  51. 1:55after in the way of working with
  52. 1:57artificial intelligence. If you can't
  53. 1:59go on vacation, don't worry. You're
  54. 2:01welcome, guys. And taking advantage of
  55. 2:03the fact that we're talking about
  56. 2:05travel, I'll take this opportunity to
  57. 2:06tell you that, well, as you can imagine
  58. 2:08, a large part of my work has to do
  59. 2:09with working while I'm traveling. If
  60. 2:11I'm traveling for a presentation, for a
  61. 2:13talk, or, for example, if I have to
  62. 2:15work with an artificial intelligence,
  63. 2:18with a tool that is not available in my
  64. 2:20geographical area, in my country, then
  65. 2:22one of the tools that I use the most is
  66. 2:24a VPN like Norbn. Typical situation.
  67. 2:27Imagine that I want to try out the
  68. 2:28latest artificial intelligence
  69. 2:30available and suddenly I realize that
  70. 2:32I'm in Europe and that I don't have
  71. 2:34access to it because it has been
  72. 2:35limited. No problem, with Nord VPN I
  73. 2:37can change my geographic location, the
  74. 2:40location from where I'm connecting to
  75. 2:42the internet and then I can use it
  76. 2:44without any problem as if I were in
  77. 2:46another country. Or for example, I'm at
  78. 2:49an airport, in a hotel, in a coffee
  79. 2:51shop and I'm connected to a public
  80. 2:52Wi-Fi network and I have to work with
  81. 2:54my accounts, bank accounts, private
  82. 2:56information, which as you know is not
  83. 2:58the most secure from a privacy point of
  84. 3:00view. Well, that's where I activate
  85. 3:03Nord VPN and with one click my entire
  86. 3:04internet connection becomes encrypted.
  87. 3:07Plus, it's one of those applications
  88. 3:09that you install once and have there
  89. 3:10working in the background very
  90. 3:12comfortably, without interfering with
  91. 3:13your work. If you want to try it, Nord
  92. 3:17VPN has given me a discount for you,
  93. 3:19simply click on the following link and
  94. 3:22get Nord VPN for 2 years with 4 extra
  95. 3:24months and a 30-day money-back
  96. 3:26guarantee if you need it. I'll leave
  97. 3:29the link to this offer in the
  98. 3:31description box and in the pinned
  99. 3:33comment below. And now, let's continue
  100. 3:36with the video and we're going to start
  101. 3:39talking about GPT 5.6, which, as you
  102. 3:41know, is the new family of Openi AI
  103. 3:43models that completely breaks with the
  104. 3:45nomenclature that we've been used to in
  105. 3:47the past, because now Open AI has
  106. 3:49wanted to copy the strategy, and quite
  107. 3:51sensibly, that Antropic was following,
  108. 3:54which on the one hand differentiates
  109. 3:56the generation of the model as 5.3, 5.4
  110. 3:58, 5.5, 5, which would be the different
  111. 4:00levels of capabilities that the models
  112. 4:02can offer. Well, we know how Antropic
  113. 4:06has historically had Haikusonet and
  114. 4:08Opus for the different levels of
  115. 4:09capabilities. And I think that here
  116. 4:12OpenI was a bit envious of seeing how
  117. 4:14Antropic was able to make a splash with
  118. 4:16the release of myths, inaugurating a
  119. 4:18new level above what for all of us was
  120. 4:20the maximum level, which was the Opus
  121. 4:22level, because OpenI actually also had
  122. 4:24its scale of models in the past. They
  123. 4:27had mini models, normal models, Max
  124. 4:29models, but of course, when you want to
  125. 4:31expand your capacity and create a new
  126. 4:33tier like Antropic has done, what else
  127. 4:35is there than the Max? There is no Max,
  128. 4:37there is nothing more. And that's why I
  129. 4:40think Openi AI has been envious of this
  130. 4:42coup with the release of myths that
  131. 4:44Antropic made, and they have changed
  132. 4:45the nomenclature towards a type of
  133. 4:47scale that we can all understand, which
  134. 4:49in this case is related to stellar
  135. 4:50objects, where we have the Luna model
  136. 4:52as the mini model: the smallest, most
  137. 4:54efficient, fastest and least
  138. 4:55intelligent. Then the Terra model,
  139. 4:58which would be the intermediate model,
  140. 5:00which strikes a good balance between
  141. 5:02capabilities, speed and price. And then
  142. 5:05the star that heats up the most, the
  143. 5:07GPT Sol model, which we will understand
  144. 5:09from now on as the most intelligent
  145. 5:11model within the GPT families. And this
  146. 5:14will now allow Open AI to do two things
  147. 5:16. To continue evolving model lines in a
  148. 5:18decoupled way, since we could see a
  149. 5:20version 5.7 of the Sol model, while
  150. 5:22Earth and Moon are left behind, as well
  151. 5:24as Santropic with its models. And this
  152. 5:27also enables them, looking forward, to
  153. 5:29inaugurate a new scale for a more
  154. 5:31powerful future family of models, which
  155. 5:33might be called Galaxies, perhaps
  156. 5:35something like a GPT Andromeda. Hey,
  157. 5:38Samman, take note of this one, it's not
  158. 5:39bad. From here on, I'm not going to
  159. 5:42focus so much on telling you, as is
  160. 5:44usual, what the benefits of this new
  161. 5:46GPT 5.6, six, are, because I think
  162. 5:48we've all had time to try it out and I
  163. 5:51think there have rarely been so many
  164. 5:53consensus when it comes to declaring a
  165. 5:55model as good, a very good model where
  166. 5:57most people agree that it is not a
  167. 5:59superior model to Fable 5, Antropic's
  168. 6:01flagship model, which we could say is
  169. 6:04the most advanced artificial
  170. 6:05intelligence model today; but which,
  171. 6:08perhaps without being the most powerful
  172. 6:10model, strikes a balance that is quite
  173. 6:12successful on the part of Open AI in
  174. 6:14terms of the efficiency and
  175. 6:16intelligence of the model. And we can
  176. 6:19see this very well with the typical
  177. 6:21cost curves where we see that the Open
  178. 6:23AI models do indeed have different
  179. 6:24capacities depending on the type of
  180. 6:26reasoning capacity that we establish.
  181. 6:29Well, yes, in this case we also have
  182. 6:31different levels of reasoning for each
  183. 6:33of the different models. We also have
  184. 6:35an ultra mode. The truth is that we
  185. 6:38have quite a few selections that, again
  186. 6:40, make it a bit complex to know which
  187. 6:42model we have to work with in each case
  188. 6:45; but, well, when we compare the cost
  189. 6:47curves, what we realize is that the
  190. 6:49Open AI models, look here, the cost
  191. 6:51curves, have a capacity level close to
  192. 6:53or higher than, for example, Opus 4.8 8
  193. 6:55, or close in this case to the Fable 5
  194. 6:57model, but at a cost, efficiency, and
  195. 7:01also at a much faster execution speed
  196. 7:04than that offered by the Antropic
  197. 7:06models. And this is something that is
  198. 7:09clearly seen in the study carried out
  199. 7:11by Artificial Analysis with its typical
  200. 7:13evaluation index with different
  201. 7:15benchmarks, where we see that yes,
  202. 7:17Fable 5 is the most intelligent model,
  203. 7:19there is no doubt; it is the one in
  204. 7:21first position, but very close to GPT
  205. 7:225.6 6, in this case by a difference of
  206. 7:24one percentage point. But the thing is,
  207. 7:28with almost similar performance, when
  208. 7:30we study the cost per intelligence,
  209. 7:31what it costs to evaluate these models,
  210. 7:33look how outrageous. Look at the
  211. 7:36difference we have between the Sol
  212. 7:38model, which in this case has as a task
  213. 7:40, compared to the Fable 5 model, which
  214. 7:43goes up to almost $ 2.75 in terms of
  215. 7:45execution price and also in terms of
  216. 7:47the tediousness, uh, when making
  217. 7:49inferences; the agility it demonstrates
  218. 7:51when giving us answers, both in cost
  219. 7:53and the number of tokens it has to
  220. 7:55generate at the start. Okay, we can see
  221. 7:58that the difference is quite striking
  222. 8:01and this is what for me marks the big
  223. 8:03difference between what would be GPT
  224. 8:055.6 Sol versus Fable 5. Yes, Fable 5 is
  225. 8:08an excellent model, but I think that
  226. 8:10the point we are at now is a point
  227. 8:13where agentic AI that works for us on
  228. 8:15our computers starts to work. I think
  229. 8:19we have crossed a threshold where this
  230. 8:20is no longer a proof of concept, but
  231. 8:22rather these models are starting to be
  232. 8:24quite, quite intelligent to be useful
  233. 8:26in an agentic workflow, where we can
  234. 8:28start to orient them to work with
  235. 8:29different programs on our computers, to
  236. 8:31generate files, to work with our own
  237. 8:33local context on our computers. In
  238. 8:37other words, they are starting to be
  239. 8:38tools that we want to work with
  240. 8:40intensively and where in this case, the
  241. 8:42availability of use, the cost, and the
  242. 8:44agility of the models start to make a
  243. 8:45difference. And that is where, for me,
  244. 8:49Open AI has hit the nail on the head
  245. 8:51and Open AI knows it; and that's why
  246. 8:53they've also accompanied the release of
  247. 8:56GPT 5.6 with an update to their Chat
  248. 8:58GPT application to introduce the
  249. 9:00concept of Chat GPT Work, which is
  250. 9:02nothing more than inviting the general
  251. 9:05public to start using agentic AI as it
  252. 9:07should be, as it's done in 2026 and as
  253. 9:09many of us have been using for months
  254. 9:12for programming. You'd be surprised how
  255. 9:15many people still only understand Chat
  256. 9:17GPT as that interface on the web page
  257. 9:19that you're going to ask questions
  258. 9:20about and that they have no idea of all
  259. 9:22the agentic potential that artificial
  260. 9:24intelligence has right now. And this is
  261. 9:27the big market that these big labs want
  262. 9:30to attack. In fact, this is something
  263. 9:32that I've thought about a lot: that
  264. 9:33people, the general public, still have
  265. 9:35no idea how these tools have evolved.
  266. 9:37And I think that in the coming weeks
  267. 9:39I'm going to prepare a free tutorial
  268. 9:42here for the channel, to add many more
  269. 9:44people to the genetic era of these
  270. 9:46tools. And that's why the application
  271. 9:49that used to be Codex has now evolved
  272. 9:51and is the Chat GPT application, the
  273. 9:53super application concept, and where we
  274. 9:55have the tab here to choose whether we
  275. 9:57want to continue working as developers
  276. 9:59with Codex or directly with Chat GPT to
  277. 10:01carry out these more general tasks,
  278. 10:03more day-to-day tasks on our computer.
  279. 10:07And well, this was also changed a
  280. 10:09couple of days ago, which is the toggle
  281. 10:11that was on the website but not in the
  282. 10:12application and they put it there
  283. 10:14because I think people weren't finding
  284. 10:16how to chat with their usual Chat GPT.
  285. 10:18So, here you had the distinction
  286. 10:20between chatting or working, which is
  287. 10:22working with agentic AI. And, as you
  288. 10:25can see, here we have the new family of
  289. 10:27models; models that, as I've already
  290. 10:29insisted, are quite agile, quite
  291. 10:31capable and that have pleasantly
  292. 10:32surprised me, because I haven't come
  293. 10:34across any task in which I've felt that
  294. 10:36the artificial intelligence has wasted
  295. 10:38my time during the tests I've done this
  296. 10:39week. This happened with previous
  297. 10:43models, where you would give it a task
  298. 10:45and it would seem that it was moving
  299. 10:46towards a solution, but then when you
  300. 10:48realized it, it would iterate, iterate,
  301. 10:50iterate and please you, but in no case
  302. 10:52did it give you a result. Well, with
  303. 10:54GPT 5.6, it is the first time that I
  304. 10:57notice that almost 100%of the tasks
  305. 10:59that I have asked it to do have been
  306. 11:01solved quite easily. Easily with the
  307. 11:04Sol model, of course, Sol is easily
  308. 11:06done. And, for example, one of the
  309. 11:08first videos that I saw of the model
  310. 11:09working with other programs, which in
  311. 11:14the end is what I mean, it is the
  312. 11:16strong point, was this video here,
  313. 11:18where I saw how GPT 5.6 was, well,
  314. 11:20creating this model of a computer with
  315. 11:22a pretty spectacular final rendering in
  316. 11:25the 3D software Blender. It is a video
  317. 11:29that can be a bit misleading when you
  318. 11:30realize that the video actually starts
  319. 11:32with the model already loading assets
  320. 11:34from this laptop that has already been
  321. 11:36created; That is to say, it is not
  322. 11:38modeling in three dimensions, uh, what
  323. 11:40we are really seeing at the end, but
  324. 11:42simply loading the assets, generating
  325. 11:44the rendering configuration and
  326. 11:45reaching a result like what we see here
  327. 11:47, well, for me it is quite useful as a
  328. 11:49person who does not know how to use
  329. 11:51this software. So I wanted to try it
  330. 11:53out. And look what I mean when I say
  331. 11:56that the model works. The test simply
  332. 11:58consisted of me installing Blender on
  333. 12:00my own computer and telling it, "I just
  334. 12:02installed Blender on my computer, but I
  335. 12:04don't know how to use it. Even so, I've
  336. 12:07seen that there are MCPs so that you
  337. 12:09can use it directly for me. Configure
  338. 12:11it and do a usage test. I'm not going
  339. 12:13to give you any more information, I'm
  340. 12:14not going to tell you how to configure
  341. 12:16it, figure it out. And in a matter of 4
  342. 12:18minutes, look, the model manages to
  343. 12:21configure its own Blender MCP. It
  344. 12:24starts communicating, performs
  345. 12:25different tests in an automated way,
  346. 12:27closing this loop to check that its
  347. 12:29result is correct; it makes its own
  348. 12:31render where it says MCP OK, that is,
  349. 12:33it is now connected to Blender and is
  350. 12:35working. And voila, in 4 minutes, the
  351. 12:38truth is that we have everything
  352. 12:40connected. And the next test that I
  353. 12:42wanted to try was to upload a photo,
  354. 12:44this photo here, and ask it to recreate
  355. 12:46this scene and room in Blender. This is
  356. 12:49a task that, obviously, the model has
  357. 12:51to be very fluent in Blender to do. to
  358. 12:54do it professionally, but I simply
  359. 12:56wanted to see how advanced this vision
  360. 12:58capability could be to see the image
  361. 13:00and translate this into operations
  362. 13:02within Blender, which resulted, after,
  363. 13:04look, 4 or 5 minutes of work, in this
  364. 13:07scene here that I found very funny.
  365. 13:10There are people who have been
  366. 13:11commenting on the shape of the
  367. 13:12microphone, guys, please, uh, behave.
  368. 13:14But it really is quite surprising. Look
  369. 13:16, don't you think? I don't even know
  370. 13:18how to move around Blender. How can I,
  371. 13:19uh, rotate the camera here? I'm trying
  372. 13:21to, okay, do a rotation like this. Look
  373. 13:23, I don't, I don't know Blender and yet
  374. 13:26the artificial intelligence has already
  375. 13:28given me this here as a base, okay?
  376. 13:31Which as I say, is not a result that is
  377. 13:33even remotely useful for a professional
  378. 13:35working with 3D, but it begins to give
  379. 13:37us clues as to where these models are
  380. 13:39going to evolve in the coming years.
  381. 13:43Quite a fan of my face, honestly, but
  382. 13:45look at the little details, uh, of the
  383. 13:47lights, the audio waveform, Not bad,
  384. 13:49right? It's not too bad or too good,
  385. 13:53but since I want to show you that this
  386. 13:55can actually start to have a real
  387. 13:57professional utility, I'm going to take
  388. 14:00you to a domain that I'm much more
  389. 14:02familiar with, which has to do with the
  390. 14:04use of After Effects. After Effects is
  391. 14:08the software that I've always used to
  392. 14:10create many of the animations that you
  393. 14:12see on this and my other channels, and
  394. 14:14that, well, requires a lot of work to
  395. 14:16make a short 5 or 10 second animation,
  396. 14:18especially when it comes to placing a
  397. 14:20lot of elements, making designs, trying
  398. 14:22to create that first sketch of what you
  399. 14:24want to animate and, then, well, yes,
  400. 14:26getting into the animation process.
  401. 14:28It's a lot of time. And so I wanted to
  402. 14:30try it out and I said," Hey, I've seen
  403. 14:32that there are ways to connect
  404. 14:34artificial intelligence to programs. "
  405. 14:36For example, a Blender MCP for After
  406. 14:38Effects, well I didn't know if it
  407. 14:40existed, it didn't exist, so I'm going
  408. 14:43to directly ask GPT 5.6 if it can make
  409. 14:45one with After Effects so that I can
  410. 14:47then control it autonomously. I'm not
  411. 14:51even asking it to configure an MCP, but
  412. 14:53rather I'm asking Chat GPT to build it,
  413. 14:55to implement it. And believe me, I
  414. 14:58didn't think this was going to work so
  415. 15:01well. 6 minutes 38 seconds, the model
  416. 15:04working and suddenly it tells me," Yes,
  417. 15:06the After Effects MCP is already
  418. 15:07implemented, compiled and connected to
  419. 15:09the project. It includes 17 tools that
  420. 15:12allow us to do this type of thing now
  421. 15:15using artificial intelligence itself.
  422. 15:17In other words, we have created the
  423. 15:19bridge that will now serve as the
  424. 15:21infrastructure so that the AI itself
  425. 15:23can work with the software. Here I was
  426. 15:25still skeptical, I was still thinking,
  427. 15:27well, let's see what it has done for me
  428. 15:29, let's see what it can do now. And I
  429. 15:32told him, now to create an animation of
  430. 15:34a GPT chat-type menu, where you write
  431. 15:35the text of a prompt, enter the cursor,
  432. 15:37click and then the response is
  433. 15:39generated. And this, now yes, just
  434. 15:43working for 2 minutes, 2 minutes, the
  435. 15:45result has been an After Effects
  436. 15:46composition with all these elements
  437. 15:48that just naming all the elements, all
  438. 15:51the layers that he has put here, would
  439. 15:53have taken me more work than these 2
  440. 15:55minutes and he has created an animation
  441. 15:57like this, okay? A GPT 2023 chat-type
  442. 16:01interface. The cursor enters, a bit of
  443. 16:03a weird animation, but okay. And we see
  444. 16:06how all the elements start to be drawn
  445. 16:08and we have a pseudo-animation of what
  446. 16:09I asked him for. Unlike the Blender
  447. 16:12project, where I don't know if this
  448. 16:14works or not, here I can tell you that,
  449. 16:16at least this as a base, gets me a lot
  450. 16:19of work done. Because here the only
  451. 16:21thing I would have to do now is polish:
  452. 16:23polish these animation errors, this
  453. 16:25thing that is misplaced, change fonts,
  454. 16:27change styles, but this is already a
  455. 16:28base that gets me ahead of 80%of the
  456. 16:30work it would take me to make this
  457. 16:32animation. I really can't emphasize how
  458. 16:36brutal this is. So much so that of
  459. 16:38course, I started doing more tests and
  460. 16:40asked him, "Okay, now I want you to
  461. 16:41create another composition where a
  462. 16:43three-dimensional orbit of a gradient
  463. 16:44descent-type animation is visualized in
  464. 16:463D." And here I have to tell you that
  465. 16:48this was a hydration process. The model
  466. 16:50did not reach an initial solution that
  467. 16:53was even close to professional, but
  468. 16:54here I wanted to see, not the one-shot
  469. 16:57effect of the model, but to iterate,
  470. 16:59iterate, iterate and see if I could get
  471. 17:01a really good result. And well, I kept
  472. 17:03testing. It didn't take me that long
  473. 17:05either, but look how after five or six
  474. 17:07prompts I already have an animation
  475. 17:09that I'm starting to like quite a bit.
  476. 17:12This is already a very interesting
  477. 17:13composition that I was able to end up
  478. 17:15with something like this, okay? A
  479. 17:18composition that, for me, is already 98
  480. 17:20%of what a good animation would be, and
  481. 17:23the final touch was to give it this
  482. 17:25more digital style and to mark the
  483. 17:27color, but the complicated part of the
  484. 17:29animation was all done by the damn AI.
  485. 17:32An animation whose result looks like
  486. 17:35this. This blows my mind because this
  487. 17:39is an animation that if an editor had
  488. 17:41given it to me, it would be 10 out of
  489. 17:4310. It is the perfect animation from my
  490. 17:45point of view, my criteria of what
  491. 17:46gradient descent would be. And it is an
  492. 17:49animation that, as I say, the AI has
  493. 17:51done 80 to 90%of the work for me. And
  494. 17:53not only that, but it is the agility
  495. 17:55with which you can start moving between
  496. 17:57different ideas. Imagine that now I
  497. 17:59don't want the stroke in the gradient
  498. 18:01descent to be this descent here, but
  499. 18:03rather different ones, because the
  500. 18:04normal thing would be to redo almost
  501. 18:06the entire animation in After Effects.
  502. 18:09This would take me a lot of manual work
  503. 18:11, but here I would simply ask Chat GPT.
  504. 18:14Now I want the composition itself to
  505. 18:16have different paths that descend on
  506. 18:18different sides simultaneously with
  507. 18:20different colors. I send it, then he
  508. 18:23works. I'm going to expand your current
  509. 18:24composition with three simultaneous
  510. 18:26optimizations. I'll keep the yellow one
  511. 18:28on the right, I'll add different paths.
  512. 18:30He understands the task perfectly. Look
  513. 18:32, he's been here for 6 minutes and he's
  514. 18:34already started showing us the
  515. 18:36different paths he's just designed and
  516. 18:38the final result is what you're seeing
  517. 18:40on the screen now. How crazy, Carlos.
  518. 18:46Does this mean that GPT 5.6 has done
  519. 18:48something for you that Fable 5 couldn't
  520. 18:50? No, Fable 5 could probably have made
  521. 18:53its own MCP in After Effects. It could
  522. 18:56probably have worked with the
  523. 18:57compositions I just showed you, but the
  524. 18:59truth is that I probably would have
  525. 19:01spent the fee before getting a result
  526. 19:03like the ones we've seen. And that's
  527. 19:05the key to everything. I think right
  528. 19:08now that these agentic models have
  529. 19:10become a Swiss army knife that we can
  530. 19:12connect to Excel, Word, PDF, After
  531. 19:14Effects, Blender, finding ourselves
  532. 19:16limited in how much we can use these
  533. 19:18models is quite frustrating. And Open
  534. 19:22AI knows this and that's why they've
  535. 19:24been playing with all that efficiency,
  536. 19:26all that extra computing that they seem
  537. 19:27to have doing massive resets all this
  538. 19:29week. I think they've reset our GPT
  539. 19:33usage quotas about four or five times.
  540. 19:35Furthermore, in some cases they are
  541. 19:37cumulative resets that you can save to
  542. 19:40apply later. They are giving us a huge
  543. 19:42amount of inference, somewhat justified
  544. 19:45by this evolution in the number of
  545. 19:47users they are having; which has had,
  546. 19:49well, with the release of GPT 5.6, a
  547. 19:52meteoric rise. And they know that they
  548. 19:54have to fight this battle, they know
  549. 19:56that at this point users have to start
  550. 19:58testing Igenetics, they have to start
  551. 20:00seeing that it already works. And they
  552. 20:03know that while Antropic is drowning
  553. 20:04due to its lack of computing, they can
  554. 20:06play at being the Daddy White who gives
  555. 20:08you more inference. But be careful,
  556. 20:12because although your competitor now
  557. 20:14seems to be showing some weakness, we
  558. 20:16know on the one hand that Antropic
  559. 20:18always comes back very strong and on
  560. 20:20the other hand your margins with open
  561. 20:22source models continue to narrow.
  562. 20:26Another big piece of news we've seen
  563. 20:28this week has been the release of the
  564. 20:31new Química 3 model from the Moonshot
  565. 20:33laboratory, a Chinese laboratory that
  566. 20:36has released the largest, open source,
  567. 20:38open weight model, as they say, the
  568. 20:41largest available to date, almost
  569. 20:43doubling what was the largest model we
  570. 20:45had until now: a 1.6 model which was
  571. 20:48the Dipsic model. In this case, the
  572. 20:51Kimica 3 model, notice how it rises to
  573. 20:53a size scale that is already starting
  574. 20:56to be equivalent to what is possibly
  575. 20:58GPT 5.6 sol or the model of Antropic
  576. 21:00Fable 5. And of course, as happens in
  577. 21:02the world of deep learning, the scale
  578. 21:04often determines your capabilities,
  579. 21:07leaving us now with a model that, look
  580. 21:09at many of the benchmarks presented
  581. 21:11here by the people at Moonshot, notice
  582. 21:13how Chemistry 3 is above, in many cases
  583. 21:15, the Opus 4.8 model of GPT 5.5. The
  584. 21:20models that were frontier models just a
  585. 21:22few weeks ago and are very close, not
  586. 21:24surpassing and I wouldn't say in the
  587. 21:26same league, but very close to what
  588. 21:28would be GPT 5.6 Sol or Fable 5, and a
  589. 21:30model that, well, at this scale,
  590. 21:31knowing how the capabilities frontiers
  591. 21:34are in the world of deep learning, we
  592. 21:36will surely see that in some tasks it
  593. 21:38gets closer to Fable 5, even surpassing
  594. 21:40it, or in other tasks to GPT 5.6 or
  595. 21:42surpassing it. That is to say, we have
  596. 21:45a new great model with all the
  597. 21:46capabilities that they can offer. With
  598. 21:48one difference that the previous ones
  599. 21:50do not have, which is that this model
  600. 21:52is openweight. This model, large
  601. 21:54organizations that do not want to be
  602. 21:55blocked because the United States
  603. 21:57suddenly cuts off access to Open AI or
  604. 21:59anthropogenic that wants to have
  605. 22:01control over data submissions, privacy,
  606. 22:03can use these models in their own
  607. 22:05organization. It will be expensive, it
  608. 22:08will require a lot of infrastructure,
  609. 22:10but having a model with these
  610. 22:11characteristics open to the public is a
  611. 22:13real earthquake. Look how, going back
  612. 22:16to Artificial Analysis Intelligence, uh
  613. 22:19, Chemistry 3 is in third position,
  614. 22:22that is, it has made a huge leap,
  615. 22:24overtaking Opus, GPT 5.6 Terra, GPT 5.5
  616. 22:27, Grock 4.5, Sonet, overtaking all
  617. 22:29these models. And notice that if we
  618. 22:31look at it in terms of cost, we would
  619. 22:33be looking at a model that is much more
  620. 22:35similar to what GPT 5.6 would be in its
  621. 22:37maximum configuration, the Sol model,
  622. 22:39than what the Fable 5 model would be,
  623. 22:41okay? As we've seen before, it's a much
  624. 22:43more expensive and cumbersome model.
  625. 22:46It's crazy. In a 2026 where Antropic,
  626. 22:48with Mos and Fable, had broken the
  627. 22:50growth trend and, suddenly, had given
  628. 22:52us a model more capable than what we
  629. 22:54all expected AI to be able to achieve
  630. 22:56this year, I was already surprised that
  631. 22:58Open AI was able to respond with
  632. 23:00something that was more or less up to
  633. 23:02par, which is the model we've analyzed
  634. 23:04today. Imagine my surprise and that of
  635. 23:08the entire community when we saw that
  636. 23:10an open source competitor had managed
  637. 23:12to offer a model that is also up to par
  638. 23:14. The difference in capabilities
  639. 23:17between private models, which is the
  640. 23:18black line we see here, which ends with
  641. 23:20Fable 5, versus the open source models,
  642. 23:22which is the blue line and which is not
  643. 23:24updated. What we're seeing here is GLM
  644. 23:265.2. Keep this in mind. GLM 5.2 from ZI
  645. 23:29, which was a model that we celebrated
  646. 23:31a few weeks ago when it came out
  647. 23:33because it was a model that, wow, had a
  648. 23:36lot of capabilities despite being an
  649. 23:38open source model. Do you see this
  650. 23:40margin here? This difference. Now look
  651. 23:42at how Chemistry 3 compares to, well,
  652. 23:45the Fable 5 models, GPT 5.6 sol, all
  653. 23:48the Frontier models, and GLM 5.2. Look
  654. 23:52at the jump in agentic programming
  655. 23:54capabilities. Deep SWE 67.5 vs. 46.2.
  656. 23:59Program Bench 77 vs. 63.7, Frontier SWE
  657. 24:0481 vs. 67.3. Marathon SWE 42 vs. 13.
  658. 24:10It's outrageous. Still, as you know,
  659. 24:13these are benchmarks and benchmarks are
  660. 24:15the devil's work. We all know this, so
  661. 24:18we must let the community try it out
  662. 24:20and see for themselves what kind of
  663. 24:21sensations this new Chemistry 3 model
  664. 24:23transmits to all of us. Still, to close
  665. 24:27, I wanted to do a comparative test
  666. 24:29between Fable 5, GPT 5.6, 6 and
  667. 24:30Chemistry 3 in one of the tasks that,
  668. 24:32well, many of us usually care about and
  669. 24:34we know that in general GPT models have
  670. 24:36failed enormously, and that is: "
  671. 24:38Generate a frontend for me, generate
  672. 24:40something pretty, please, and show it
  673. 24:42to me on the screen without it being
  674. 24:44little boxes inside little boxes and
  675. 24:45purple colors everywhere." And the
  676. 24:48prompt that I tried was the following.
  677. 24:51Develop the frontend of what would be a
  678. 24:53theme park like Disneyland Paris or
  679. 24:54Disney World, but dedicated to the
  680. 24:56world of deep learning and artificial
  681. 24:58intelligence. We are going to mix two
  682. 25:00worlds here to see how creative the AI
  683. 25:02is. Generate all the assets needed to
  684. 25:04create a result that is rich in details
  685. 25:06and professional design. I will first
  686. 25:08show you the result that after 16
  687. 25:11minutes of work GPT 5.5 5 in high
  688. 25:13configuration has generated for me. The
  689. 25:16result is not bad, but we see some of
  690. 25:18the usual flaws of a GPT model, because
  691. 25:20as I was saying, it used to be a lot of
  692. 25:22boxes, a format that our eyes are
  693. 25:24already used to seeing and that stands
  694. 25:26out a lot, since it is made with
  695. 25:28artificial intelligence. I think there
  696. 25:31are some very interesting things. Here
  697. 25:32it has created assets for us such as
  698. 25:33the park map itself. Look how we have
  699. 25:36Neural Kingdom, Vision Valley, Language
  700. 25:37Lagoon, Robotics Groove, Ethics Garden.
  701. 25:41It's pretty cool and we can see the
  702. 25:42different areas of the map. The result
  703. 25:44is pretty good, we won't deny it, but
  704. 25:46we could do better. And this is
  705. 25:48something we have to celebrate because
  706. 25:50Open AI has been working on this very
  707. 25:51thing. Look, what we are seeing here is
  708. 25:54a visualization that encodes different
  709. 25:56frontends generated by, in this case,
  710. 25:58GPT 5.6. And notice how it is clearly
  711. 26:02seen that within this map of different
  712. 26:03styles, there are a few white gaps that
  713. 26:05people have been studying to see what
  714. 26:07they correspond to. And what they have
  715. 26:09found is that these white spaces, look
  716. 26:11at this area that is headed by blue
  717. 26:13dots, that is, frontends that GPT 5.5
  718. 26:15did generate and that GPT 5.6 or 6 Sol
  719. 26:17does not generate, otherwise you would
  720. 26:19see orange dots. Here they correspond
  721. 26:22to frontends with purple designs or
  722. 26:24designs that, for example, have boxes,
  723. 26:26fonts with a specific style. In other
  724. 26:30words, Open AI has done post-training
  725. 26:32work to remove, to eliminate these
  726. 26:34flaws that we know GPT frontends have.
  727. 26:39Thus, the result generated by the Luna
  728. 26:41model, the most basic model in the Open
  729. 26:43AI family, is a slightly more
  730. 26:44aesthetically pleasing web page. This
  731. 26:47is already starting to look like what a
  732. 26:49museum or amusement park page could be.
  733. 26:51There are certainly errors. We see that
  734. 26:54there are things that do not load,
  735. 26:55there is information that is not really
  736. 26:57generated well, but here we are working
  737. 26:58with the Luna model which, let's
  738. 27:00remember, is the least capable. But at
  739. 27:02least we have a frontend that we could
  740. 27:04say is a little better and, at least,
  741. 27:06different from what we are used to. If
  742. 27:09we now look at the result of GPT SOL,
  743. 27:11the most advanced model, this is the
  744. 27:13web page that it has created for us
  745. 27:15with assets that take advantage of the
  746. 27:17image generation capacity. Well, we see
  747. 27:20how these images have been generated to
  748. 27:22enrich our web page and we already see
  749. 27:24a very professional web page, okay? I
  750. 27:27see some things that do continue, well,
  751. 27:30getting a little closer to the idea of
  752. 27:32a web page made by artificial
  753. 27:34intelligence, but we are already
  754. 27:35starting to see results that are more
  755. 27:38attractive in terms of design. I really
  756. 27:41like the idea of Lent Space as a roller
  757. 27:43coaster of a park about artificial
  758. 27:45intelligence, quite, quite cool. And to
  759. 27:49compare Fable 5 with the same prompt,
  760. 27:51Fable 5 High, returns a web page that
  761. 27:53is this one here, which of course,
  762. 27:55since Fable 5 does not have the image
  763. 27:57generation capacity, it has had to bet
  764. 28:00a lot on SVG. It's not bad, but it's
  765. 28:03not the design of a web page as I was
  766. 28:05thinking, which in this case would be a
  767. 28:07professional web page for an amusement
  768. 28:09park and in this case it comes across
  769. 28:11as something a bit strange. Note that
  770. 28:13Fable 5 is a model that Antropic models
  771. 28:15have always been quite competent in
  772. 28:17terms of design, but I'm not seeing a
  773. 28:19result here that I can compare, for
  774. 28:21example, with what GPT 5.6 Sol has done
  775. 28:24with the same Prompt. What I am seeing
  776. 28:26is that at least in terms of content,
  777. 28:28some of the things I'm seeing are much
  778. 28:30more original, right? Inference input,
  779. 28:32training pass, family batch. Hm. Surely
  780. 28:35if you start reading things around here
  781. 28:37, you'll find some very, very funny
  782. 28:38details. For example, minimum height.
  783. 28:41This bothered me a lot when I went to
  784. 28:42amusement parks as a child, the minimum
  785. 28:44height thing. 1.20 m or 120 million
  786. 28:47parameters. Okay, these are little
  787. 28:49details that I think really demonstrate
  788. 28:51the originality of these models and
  789. 28:52where I think Fall 5 has done a better
  790. 28:54job, right? Look, 1X1 convolution
  791. 28:56avenue, latent valley 10 minutes from
  792. 28:58the real world, nothing crazy. All
  793. 29:02weights reserved. There's a lot,
  794. 29:05there's a lot of humor here, eh, I like
  795. 29:07it, I like it, but it's not a
  796. 29:09professional result. And the latest
  797. 29:12contender, Kimica 3, tested from the
  798. 29:14Moonshot platform itself, has generated
  799. 29:16an agentic loop that we can see here,
  800. 29:19quite cool as it shows you the
  801. 29:20evolution of what it's building, okay?
  802. 29:24It works and works and then it
  803. 29:26generates this web page that if we look
  804. 29:28at it in full screen, it shows us
  805. 29:29Neurolandia. I don't know exactly what
  806. 29:33image generation model it has used, but
  807. 29:34anyway, this already shows more
  808. 29:36integration than Fable 5 by default.
  809. 29:39And yes, I'm seeing an aesthetic that
  810. 29:41reminds me more of what GPT 5.5 used to
  811. 29:44be for us; Well, these chips, these
  812. 29:46types of boxes, these more purple
  813. 29:48colors, so many boxes just for the sake
  814. 29:51of putting boxes. And yes, it reminds
  815. 29:54me, as I say, more of the design of
  816. 29:56previous AIs than what they offer us
  817. 29:58now, well, models like GPT 5.6. Even so
  818. 30:01, a pretty rich result with lots of
  819. 30:03little things. Here we have the mascot.
  820. 30:05Say hello to Neuro. We have information
  821. 30:08about different parades. We have an
  822. 30:10attempt at a map with SVG that's also
  823. 30:12pretty cool. Token Town, Gradient Peaks
  824. 30:15, Vector Jungle, Generative Gardens.
  825. 30:17Why isn't there a park for it? It's a
  826. 30:19pretty cool theme. The Gun mansion as a
  827. 30:23fast backend attraction, data carousel.
  828. 30:26Wow, I should have paid for it. And as
  829. 30:30I said: well, a result that, at the
  830. 30:31frontend level, doesn't appeal to me as
  831. 30:33much as the previous one, but that
  832. 30:35demonstrates the capabilities of this
  833. 30:37Kimica 3 model, which comes in like a
  834. 30:38bull in a china shop to turn the world
  835. 30:40of artificial intelligence upside down,
  836. 30:42where none of the big labs can rest on
  837. 30:44their laurels. Google, I speak for you.
  838. 30:47Where are you? What's wrong? Are you
  839. 30:49okay? You promised you would release a
  840. 30:52Gemini 3.5 Pro model in April. You told
  841. 30:55me, "In a month, Carlos, I'm going to
  842. 30:57release Gemini 3.5 Pro. What happened?
  843. 30:59Haven't you been able to keep up with
  844. 31:01your competitors?" Well, if any move
  845. 31:03were to happen in the coming weeks,
  846. 31:05guys and girls, stay tuned because
  847. 31:06we'll be reporting on it here, we'll be
  848. 31:08analyzing it sooner or later. Well, it
  849. 31:10doesn't matter. A rigorous analysis
  850. 31:12like the ones I always bring you to
  851. 31:13keep you up to date with what's
  852. 31:15happening in the world of artificial
  853. 31:16intelligence. If you liked the video,
  854. 31:18don't forget to share it. I'll leave
  855. 31:20you a couple more videos here that you
  856. 31:22can check out. Thanks again to the
  857. 31:23sponsor of the video Nord VPN. I'll
  858. 31:26leave the link to the discount I
  859. 31:27mentioned at the beginning below and
  860. 31:29I'll see you with more AI in a future
  861. 31:31video.

About this transcript

This page contains the full transcript of GPT 5.6 SOL y NUEVO KIMI 3 ¡La IA AGÉNTICA ya es REAL! by Dot CSV Lab , generated from the public captions YouTube serves with the video. The transcript has 6,039 words across 861 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.