GPT 5.6 SOL y NUEVO KIMI 3 ¡La IA AGÉNTICA ya es REAL! — Transcript
Full transcript
- 0:00As usual, every time I go on vacation,
- 0:02the world of artificial intelligence
- 0:04doesn't stop, in fact, it accelerates.
- 0:07And as you know, in recent weeks we've
- 0:09had a lot of news, and today we're
- 0:10going to analyze the most important
- 0:12ones. Specifically, the release of a
- 0:16new constellation of models, Sun, Earth
- 0:18and Moon by Open AI, or for example, by
- 0:21Antropic, the unlocking of Fable 5 that
- 0:23we could finally use at least until
- 0:25July 7. Among other news, we've also
- 0:27seen the release of the new voice model
- 0:30of Chat GPT, quite interesting, or for
- 0:32example, by Antropic the extension of
- 0:36the Fable 5 model from July 7 to July
- 0:3912. We've also seen how more pressure
- 0:42has been added by competitors such as
- 0:44Exeyi with its new Grock 4.5 model or
- 0:47the Spark 1.1 model by Meta. Or for
- 0:51example, we've also seen by Antropic
- 0:53the extension of the Fable 5 model
- 0:55until July 19. One of the biggest
- 0:57bombshells in open source models so far
- 1:00this year, with the release of the new
- 1:02Kimi 3 model, which represents for the
- 1:04first time an open source model with
- 1:072.8 trillion parameters, a huge model
- 1:09that surpasses what the cutting-edge
- 1:11models of a few months ago were and is
- 1:13now competing very closely with the new
- 1:16flagship models from Open AI and
- 1:18Antropic. Or news from Antropic,
- 1:21confirming the availability of Fable 5
- 1:24from July 20th in subscriptions.
- 1:26They're driving us crazy. And well,
- 1:28then news from Google that, oh, no,
- 1:31Google hasn't released any model, it
- 1:33hasn't released Yemini 3.5 Pro. Well,
- 1:36as you can see, a few weeks full of
- 1:37very current events, lots of new
- 1:39features that we're going to be
- 1:40analyzing today through the reflections
- 1:42, conclusions that I've been able to
- 1:44build after testing these models during
- 1:46the last week. And I'm going to show
- 1:48you some use cases that have blown my
- 1:50mind and that demonstrate the maturity
- 1:52that these models have already achieved
- 1:54and that I believe mark a before and
- 1:55after in the way of working with
- 1:57artificial intelligence. If you can't
- 1:59go on vacation, don't worry. You're
- 2:01welcome, guys. And taking advantage of
- 2:03the fact that we're talking about
- 2:05travel, I'll take this opportunity to
- 2:06tell you that, well, as you can imagine
- 2:08, a large part of my work has to do
- 2:09with working while I'm traveling. If
- 2:11I'm traveling for a presentation, for a
- 2:13talk, or, for example, if I have to
- 2:15work with an artificial intelligence,
- 2:18with a tool that is not available in my
- 2:20geographical area, in my country, then
- 2:22one of the tools that I use the most is
- 2:24a VPN like Norbn. Typical situation.
- 2:27Imagine that I want to try out the
- 2:28latest artificial intelligence
- 2:30available and suddenly I realize that
- 2:32I'm in Europe and that I don't have
- 2:34access to it because it has been
- 2:35limited. No problem, with Nord VPN I
- 2:37can change my geographic location, the
- 2:40location from where I'm connecting to
- 2:42the internet and then I can use it
- 2:44without any problem as if I were in
- 2:46another country. Or for example, I'm at
- 2:49an airport, in a hotel, in a coffee
- 2:51shop and I'm connected to a public
- 2:52Wi-Fi network and I have to work with
- 2:54my accounts, bank accounts, private
- 2:56information, which as you know is not
- 2:58the most secure from a privacy point of
- 3:00view. Well, that's where I activate
- 3:03Nord VPN and with one click my entire
- 3:04internet connection becomes encrypted.
- 3:07Plus, it's one of those applications
- 3:09that you install once and have there
- 3:10working in the background very
- 3:12comfortably, without interfering with
- 3:13your work. If you want to try it, Nord
- 3:17VPN has given me a discount for you,
- 3:19simply click on the following link and
- 3:22get Nord VPN for 2 years with 4 extra
- 3:24months and a 30-day money-back
- 3:26guarantee if you need it. I'll leave
- 3:29the link to this offer in the
- 3:31description box and in the pinned
- 3:33comment below. And now, let's continue
- 3:36with the video and we're going to start
- 3:39talking about GPT 5.6, which, as you
- 3:41know, is the new family of Openi AI
- 3:43models that completely breaks with the
- 3:45nomenclature that we've been used to in
- 3:47the past, because now Open AI has
- 3:49wanted to copy the strategy, and quite
- 3:51sensibly, that Antropic was following,
- 3:54which on the one hand differentiates
- 3:56the generation of the model as 5.3, 5.4
- 3:58, 5.5, 5, which would be the different
- 4:00levels of capabilities that the models
- 4:02can offer. Well, we know how Antropic
- 4:06has historically had Haikusonet and
- 4:08Opus for the different levels of
- 4:09capabilities. And I think that here
- 4:12OpenI was a bit envious of seeing how
- 4:14Antropic was able to make a splash with
- 4:16the release of myths, inaugurating a
- 4:18new level above what for all of us was
- 4:20the maximum level, which was the Opus
- 4:22level, because OpenI actually also had
- 4:24its scale of models in the past. They
- 4:27had mini models, normal models, Max
- 4:29models, but of course, when you want to
- 4:31expand your capacity and create a new
- 4:33tier like Antropic has done, what else
- 4:35is there than the Max? There is no Max,
- 4:37there is nothing more. And that's why I
- 4:40think Openi AI has been envious of this
- 4:42coup with the release of myths that
- 4:44Antropic made, and they have changed
- 4:45the nomenclature towards a type of
- 4:47scale that we can all understand, which
- 4:49in this case is related to stellar
- 4:50objects, where we have the Luna model
- 4:52as the mini model: the smallest, most
- 4:54efficient, fastest and least
- 4:55intelligent. Then the Terra model,
- 4:58which would be the intermediate model,
- 5:00which strikes a good balance between
- 5:02capabilities, speed and price. And then
- 5:05the star that heats up the most, the
- 5:07GPT Sol model, which we will understand
- 5:09from now on as the most intelligent
- 5:11model within the GPT families. And this
- 5:14will now allow Open AI to do two things
- 5:16. To continue evolving model lines in a
- 5:18decoupled way, since we could see a
- 5:20version 5.7 of the Sol model, while
- 5:22Earth and Moon are left behind, as well
- 5:24as Santropic with its models. And this
- 5:27also enables them, looking forward, to
- 5:29inaugurate a new scale for a more
- 5:31powerful future family of models, which
- 5:33might be called Galaxies, perhaps
- 5:35something like a GPT Andromeda. Hey,
- 5:38Samman, take note of this one, it's not
- 5:39bad. From here on, I'm not going to
- 5:42focus so much on telling you, as is
- 5:44usual, what the benefits of this new
- 5:46GPT 5.6, six, are, because I think
- 5:48we've all had time to try it out and I
- 5:51think there have rarely been so many
- 5:53consensus when it comes to declaring a
- 5:55model as good, a very good model where
- 5:57most people agree that it is not a
- 5:59superior model to Fable 5, Antropic's
- 6:01flagship model, which we could say is
- 6:04the most advanced artificial
- 6:05intelligence model today; but which,
- 6:08perhaps without being the most powerful
- 6:10model, strikes a balance that is quite
- 6:12successful on the part of Open AI in
- 6:14terms of the efficiency and
- 6:16intelligence of the model. And we can
- 6:19see this very well with the typical
- 6:21cost curves where we see that the Open
- 6:23AI models do indeed have different
- 6:24capacities depending on the type of
- 6:26reasoning capacity that we establish.
- 6:29Well, yes, in this case we also have
- 6:31different levels of reasoning for each
- 6:33of the different models. We also have
- 6:35an ultra mode. The truth is that we
- 6:38have quite a few selections that, again
- 6:40, make it a bit complex to know which
- 6:42model we have to work with in each case
- 6:45; but, well, when we compare the cost
- 6:47curves, what we realize is that the
- 6:49Open AI models, look here, the cost
- 6:51curves, have a capacity level close to
- 6:53or higher than, for example, Opus 4.8 8
- 6:55, or close in this case to the Fable 5
- 6:57model, but at a cost, efficiency, and
- 7:01also at a much faster execution speed
- 7:04than that offered by the Antropic
- 7:06models. And this is something that is
- 7:09clearly seen in the study carried out
- 7:11by Artificial Analysis with its typical
- 7:13evaluation index with different
- 7:15benchmarks, where we see that yes,
- 7:17Fable 5 is the most intelligent model,
- 7:19there is no doubt; it is the one in
- 7:21first position, but very close to GPT
- 7:225.6 6, in this case by a difference of
- 7:24one percentage point. But the thing is,
- 7:28with almost similar performance, when
- 7:30we study the cost per intelligence,
- 7:31what it costs to evaluate these models,
- 7:33look how outrageous. Look at the
- 7:36difference we have between the Sol
- 7:38model, which in this case has as a task
- 7:40, compared to the Fable 5 model, which
- 7:43goes up to almost $ 2.75 in terms of
- 7:45execution price and also in terms of
- 7:47the tediousness, uh, when making
- 7:49inferences; the agility it demonstrates
- 7:51when giving us answers, both in cost
- 7:53and the number of tokens it has to
- 7:55generate at the start. Okay, we can see
- 7:58that the difference is quite striking
- 8:01and this is what for me marks the big
- 8:03difference between what would be GPT
- 8:055.6 Sol versus Fable 5. Yes, Fable 5 is
- 8:08an excellent model, but I think that
- 8:10the point we are at now is a point
- 8:13where agentic AI that works for us on
- 8:15our computers starts to work. I think
- 8:19we have crossed a threshold where this
- 8:20is no longer a proof of concept, but
- 8:22rather these models are starting to be
- 8:24quite, quite intelligent to be useful
- 8:26in an agentic workflow, where we can
- 8:28start to orient them to work with
- 8:29different programs on our computers, to
- 8:31generate files, to work with our own
- 8:33local context on our computers. In
- 8:37other words, they are starting to be
- 8:38tools that we want to work with
- 8:40intensively and where in this case, the
- 8:42availability of use, the cost, and the
- 8:44agility of the models start to make a
- 8:45difference. And that is where, for me,
- 8:49Open AI has hit the nail on the head
- 8:51and Open AI knows it; and that's why
- 8:53they've also accompanied the release of
- 8:56GPT 5.6 with an update to their Chat
- 8:58GPT application to introduce the
- 9:00concept of Chat GPT Work, which is
- 9:02nothing more than inviting the general
- 9:05public to start using agentic AI as it
- 9:07should be, as it's done in 2026 and as
- 9:09many of us have been using for months
- 9:12for programming. You'd be surprised how
- 9:15many people still only understand Chat
- 9:17GPT as that interface on the web page
- 9:19that you're going to ask questions
- 9:20about and that they have no idea of all
- 9:22the agentic potential that artificial
- 9:24intelligence has right now. And this is
- 9:27the big market that these big labs want
- 9:30to attack. In fact, this is something
- 9:32that I've thought about a lot: that
- 9:33people, the general public, still have
- 9:35no idea how these tools have evolved.
- 9:37And I think that in the coming weeks
- 9:39I'm going to prepare a free tutorial
- 9:42here for the channel, to add many more
- 9:44people to the genetic era of these
- 9:46tools. And that's why the application
- 9:49that used to be Codex has now evolved
- 9:51and is the Chat GPT application, the
- 9:53super application concept, and where we
- 9:55have the tab here to choose whether we
- 9:57want to continue working as developers
- 9:59with Codex or directly with Chat GPT to
- 10:01carry out these more general tasks,
- 10:03more day-to-day tasks on our computer.
- 10:07And well, this was also changed a
- 10:09couple of days ago, which is the toggle
- 10:11that was on the website but not in the
- 10:12application and they put it there
- 10:14because I think people weren't finding
- 10:16how to chat with their usual Chat GPT.
- 10:18So, here you had the distinction
- 10:20between chatting or working, which is
- 10:22working with agentic AI. And, as you
- 10:25can see, here we have the new family of
- 10:27models; models that, as I've already
- 10:29insisted, are quite agile, quite
- 10:31capable and that have pleasantly
- 10:32surprised me, because I haven't come
- 10:34across any task in which I've felt that
- 10:36the artificial intelligence has wasted
- 10:38my time during the tests I've done this
- 10:39week. This happened with previous
- 10:43models, where you would give it a task
- 10:45and it would seem that it was moving
- 10:46towards a solution, but then when you
- 10:48realized it, it would iterate, iterate,
- 10:50iterate and please you, but in no case
- 10:52did it give you a result. Well, with
- 10:54GPT 5.6, it is the first time that I
- 10:57notice that almost 100%of the tasks
- 10:59that I have asked it to do have been
- 11:01solved quite easily. Easily with the
- 11:04Sol model, of course, Sol is easily
- 11:06done. And, for example, one of the
- 11:08first videos that I saw of the model
- 11:09working with other programs, which in
- 11:14the end is what I mean, it is the
- 11:16strong point, was this video here,
- 11:18where I saw how GPT 5.6 was, well,
- 11:20creating this model of a computer with
- 11:22a pretty spectacular final rendering in
- 11:25the 3D software Blender. It is a video
- 11:29that can be a bit misleading when you
- 11:30realize that the video actually starts
- 11:32with the model already loading assets
- 11:34from this laptop that has already been
- 11:36created; That is to say, it is not
- 11:38modeling in three dimensions, uh, what
- 11:40we are really seeing at the end, but
- 11:42simply loading the assets, generating
- 11:44the rendering configuration and
- 11:45reaching a result like what we see here
- 11:47, well, for me it is quite useful as a
- 11:49person who does not know how to use
- 11:51this software. So I wanted to try it
- 11:53out. And look what I mean when I say
- 11:56that the model works. The test simply
- 11:58consisted of me installing Blender on
- 12:00my own computer and telling it, "I just
- 12:02installed Blender on my computer, but I
- 12:04don't know how to use it. Even so, I've
- 12:07seen that there are MCPs so that you
- 12:09can use it directly for me. Configure
- 12:11it and do a usage test. I'm not going
- 12:13to give you any more information, I'm
- 12:14not going to tell you how to configure
- 12:16it, figure it out. And in a matter of 4
- 12:18minutes, look, the model manages to
- 12:21configure its own Blender MCP. It
- 12:24starts communicating, performs
- 12:25different tests in an automated way,
- 12:27closing this loop to check that its
- 12:29result is correct; it makes its own
- 12:31render where it says MCP OK, that is,
- 12:33it is now connected to Blender and is
- 12:35working. And voila, in 4 minutes, the
- 12:38truth is that we have everything
- 12:40connected. And the next test that I
- 12:42wanted to try was to upload a photo,
- 12:44this photo here, and ask it to recreate
- 12:46this scene and room in Blender. This is
- 12:49a task that, obviously, the model has
- 12:51to be very fluent in Blender to do. to
- 12:54do it professionally, but I simply
- 12:56wanted to see how advanced this vision
- 12:58capability could be to see the image
- 13:00and translate this into operations
- 13:02within Blender, which resulted, after,
- 13:04look, 4 or 5 minutes of work, in this
- 13:07scene here that I found very funny.
- 13:10There are people who have been
- 13:11commenting on the shape of the
- 13:12microphone, guys, please, uh, behave.
- 13:14But it really is quite surprising. Look
- 13:16, don't you think? I don't even know
- 13:18how to move around Blender. How can I,
- 13:19uh, rotate the camera here? I'm trying
- 13:21to, okay, do a rotation like this. Look
- 13:23, I don't, I don't know Blender and yet
- 13:26the artificial intelligence has already
- 13:28given me this here as a base, okay?
- 13:31Which as I say, is not a result that is
- 13:33even remotely useful for a professional
- 13:35working with 3D, but it begins to give
- 13:37us clues as to where these models are
- 13:39going to evolve in the coming years.
- 13:43Quite a fan of my face, honestly, but
- 13:45look at the little details, uh, of the
- 13:47lights, the audio waveform, Not bad,
- 13:49right? It's not too bad or too good,
- 13:53but since I want to show you that this
- 13:55can actually start to have a real
- 13:57professional utility, I'm going to take
- 14:00you to a domain that I'm much more
- 14:02familiar with, which has to do with the
- 14:04use of After Effects. After Effects is
- 14:08the software that I've always used to
- 14:10create many of the animations that you
- 14:12see on this and my other channels, and
- 14:14that, well, requires a lot of work to
- 14:16make a short 5 or 10 second animation,
- 14:18especially when it comes to placing a
- 14:20lot of elements, making designs, trying
- 14:22to create that first sketch of what you
- 14:24want to animate and, then, well, yes,
- 14:26getting into the animation process.
- 14:28It's a lot of time. And so I wanted to
- 14:30try it out and I said," Hey, I've seen
- 14:32that there are ways to connect
- 14:34artificial intelligence to programs. "
- 14:36For example, a Blender MCP for After
- 14:38Effects, well I didn't know if it
- 14:40existed, it didn't exist, so I'm going
- 14:43to directly ask GPT 5.6 if it can make
- 14:45one with After Effects so that I can
- 14:47then control it autonomously. I'm not
- 14:51even asking it to configure an MCP, but
- 14:53rather I'm asking Chat GPT to build it,
- 14:55to implement it. And believe me, I
- 14:58didn't think this was going to work so
- 15:01well. 6 minutes 38 seconds, the model
- 15:04working and suddenly it tells me," Yes,
- 15:06the After Effects MCP is already
- 15:07implemented, compiled and connected to
- 15:09the project. It includes 17 tools that
- 15:12allow us to do this type of thing now
- 15:15using artificial intelligence itself.
- 15:17In other words, we have created the
- 15:19bridge that will now serve as the
- 15:21infrastructure so that the AI itself
- 15:23can work with the software. Here I was
- 15:25still skeptical, I was still thinking,
- 15:27well, let's see what it has done for me
- 15:29, let's see what it can do now. And I
- 15:32told him, now to create an animation of
- 15:34a GPT chat-type menu, where you write
- 15:35the text of a prompt, enter the cursor,
- 15:37click and then the response is
- 15:39generated. And this, now yes, just
- 15:43working for 2 minutes, 2 minutes, the
- 15:45result has been an After Effects
- 15:46composition with all these elements
- 15:48that just naming all the elements, all
- 15:51the layers that he has put here, would
- 15:53have taken me more work than these 2
- 15:55minutes and he has created an animation
- 15:57like this, okay? A GPT 2023 chat-type
- 16:01interface. The cursor enters, a bit of
- 16:03a weird animation, but okay. And we see
- 16:06how all the elements start to be drawn
- 16:08and we have a pseudo-animation of what
- 16:09I asked him for. Unlike the Blender
- 16:12project, where I don't know if this
- 16:14works or not, here I can tell you that,
- 16:16at least this as a base, gets me a lot
- 16:19of work done. Because here the only
- 16:21thing I would have to do now is polish:
- 16:23polish these animation errors, this
- 16:25thing that is misplaced, change fonts,
- 16:27change styles, but this is already a
- 16:28base that gets me ahead of 80%of the
- 16:30work it would take me to make this
- 16:32animation. I really can't emphasize how
- 16:36brutal this is. So much so that of
- 16:38course, I started doing more tests and
- 16:40asked him, "Okay, now I want you to
- 16:41create another composition where a
- 16:43three-dimensional orbit of a gradient
- 16:44descent-type animation is visualized in
- 16:463D." And here I have to tell you that
- 16:48this was a hydration process. The model
- 16:50did not reach an initial solution that
- 16:53was even close to professional, but
- 16:54here I wanted to see, not the one-shot
- 16:57effect of the model, but to iterate,
- 16:59iterate, iterate and see if I could get
- 17:01a really good result. And well, I kept
- 17:03testing. It didn't take me that long
- 17:05either, but look how after five or six
- 17:07prompts I already have an animation
- 17:09that I'm starting to like quite a bit.
- 17:12This is already a very interesting
- 17:13composition that I was able to end up
- 17:15with something like this, okay? A
- 17:18composition that, for me, is already 98
- 17:20%of what a good animation would be, and
- 17:23the final touch was to give it this
- 17:25more digital style and to mark the
- 17:27color, but the complicated part of the
- 17:29animation was all done by the damn AI.
- 17:32An animation whose result looks like
- 17:35this. This blows my mind because this
- 17:39is an animation that if an editor had
- 17:41given it to me, it would be 10 out of
- 17:4310. It is the perfect animation from my
- 17:45point of view, my criteria of what
- 17:46gradient descent would be. And it is an
- 17:49animation that, as I say, the AI has
- 17:51done 80 to 90%of the work for me. And
- 17:53not only that, but it is the agility
- 17:55with which you can start moving between
- 17:57different ideas. Imagine that now I
- 17:59don't want the stroke in the gradient
- 18:01descent to be this descent here, but
- 18:03rather different ones, because the
- 18:04normal thing would be to redo almost
- 18:06the entire animation in After Effects.
- 18:09This would take me a lot of manual work
- 18:11, but here I would simply ask Chat GPT.
- 18:14Now I want the composition itself to
- 18:16have different paths that descend on
- 18:18different sides simultaneously with
- 18:20different colors. I send it, then he
- 18:23works. I'm going to expand your current
- 18:24composition with three simultaneous
- 18:26optimizations. I'll keep the yellow one
- 18:28on the right, I'll add different paths.
- 18:30He understands the task perfectly. Look
- 18:32, he's been here for 6 minutes and he's
- 18:34already started showing us the
- 18:36different paths he's just designed and
- 18:38the final result is what you're seeing
- 18:40on the screen now. How crazy, Carlos.
- 18:46Does this mean that GPT 5.6 has done
- 18:48something for you that Fable 5 couldn't
- 18:50? No, Fable 5 could probably have made
- 18:53its own MCP in After Effects. It could
- 18:56probably have worked with the
- 18:57compositions I just showed you, but the
- 18:59truth is that I probably would have
- 19:01spent the fee before getting a result
- 19:03like the ones we've seen. And that's
- 19:05the key to everything. I think right
- 19:08now that these agentic models have
- 19:10become a Swiss army knife that we can
- 19:12connect to Excel, Word, PDF, After
- 19:14Effects, Blender, finding ourselves
- 19:16limited in how much we can use these
- 19:18models is quite frustrating. And Open
- 19:22AI knows this and that's why they've
- 19:24been playing with all that efficiency,
- 19:26all that extra computing that they seem
- 19:27to have doing massive resets all this
- 19:29week. I think they've reset our GPT
- 19:33usage quotas about four or five times.
- 19:35Furthermore, in some cases they are
- 19:37cumulative resets that you can save to
- 19:40apply later. They are giving us a huge
- 19:42amount of inference, somewhat justified
- 19:45by this evolution in the number of
- 19:47users they are having; which has had,
- 19:49well, with the release of GPT 5.6, a
- 19:52meteoric rise. And they know that they
- 19:54have to fight this battle, they know
- 19:56that at this point users have to start
- 19:58testing Igenetics, they have to start
- 20:00seeing that it already works. And they
- 20:03know that while Antropic is drowning
- 20:04due to its lack of computing, they can
- 20:06play at being the Daddy White who gives
- 20:08you more inference. But be careful,
- 20:12because although your competitor now
- 20:14seems to be showing some weakness, we
- 20:16know on the one hand that Antropic
- 20:18always comes back very strong and on
- 20:20the other hand your margins with open
- 20:22source models continue to narrow.
- 20:26Another big piece of news we've seen
- 20:28this week has been the release of the
- 20:31new Química 3 model from the Moonshot
- 20:33laboratory, a Chinese laboratory that
- 20:36has released the largest, open source,
- 20:38open weight model, as they say, the
- 20:41largest available to date, almost
- 20:43doubling what was the largest model we
- 20:45had until now: a 1.6 model which was
- 20:48the Dipsic model. In this case, the
- 20:51Kimica 3 model, notice how it rises to
- 20:53a size scale that is already starting
- 20:56to be equivalent to what is possibly
- 20:58GPT 5.6 sol or the model of Antropic
- 21:00Fable 5. And of course, as happens in
- 21:02the world of deep learning, the scale
- 21:04often determines your capabilities,
- 21:07leaving us now with a model that, look
- 21:09at many of the benchmarks presented
- 21:11here by the people at Moonshot, notice
- 21:13how Chemistry 3 is above, in many cases
- 21:15, the Opus 4.8 model of GPT 5.5. The
- 21:20models that were frontier models just a
- 21:22few weeks ago and are very close, not
- 21:24surpassing and I wouldn't say in the
- 21:26same league, but very close to what
- 21:28would be GPT 5.6 Sol or Fable 5, and a
- 21:30model that, well, at this scale,
- 21:31knowing how the capabilities frontiers
- 21:34are in the world of deep learning, we
- 21:36will surely see that in some tasks it
- 21:38gets closer to Fable 5, even surpassing
- 21:40it, or in other tasks to GPT 5.6 or
- 21:42surpassing it. That is to say, we have
- 21:45a new great model with all the
- 21:46capabilities that they can offer. With
- 21:48one difference that the previous ones
- 21:50do not have, which is that this model
- 21:52is openweight. This model, large
- 21:54organizations that do not want to be
- 21:55blocked because the United States
- 21:57suddenly cuts off access to Open AI or
- 21:59anthropogenic that wants to have
- 22:01control over data submissions, privacy,
- 22:03can use these models in their own
- 22:05organization. It will be expensive, it
- 22:08will require a lot of infrastructure,
- 22:10but having a model with these
- 22:11characteristics open to the public is a
- 22:13real earthquake. Look how, going back
- 22:16to Artificial Analysis Intelligence, uh
- 22:19, Chemistry 3 is in third position,
- 22:22that is, it has made a huge leap,
- 22:24overtaking Opus, GPT 5.6 Terra, GPT 5.5
- 22:27, Grock 4.5, Sonet, overtaking all
- 22:29these models. And notice that if we
- 22:31look at it in terms of cost, we would
- 22:33be looking at a model that is much more
- 22:35similar to what GPT 5.6 would be in its
- 22:37maximum configuration, the Sol model,
- 22:39than what the Fable 5 model would be,
- 22:41okay? As we've seen before, it's a much
- 22:43more expensive and cumbersome model.
- 22:46It's crazy. In a 2026 where Antropic,
- 22:48with Mos and Fable, had broken the
- 22:50growth trend and, suddenly, had given
- 22:52us a model more capable than what we
- 22:54all expected AI to be able to achieve
- 22:56this year, I was already surprised that
- 22:58Open AI was able to respond with
- 23:00something that was more or less up to
- 23:02par, which is the model we've analyzed
- 23:04today. Imagine my surprise and that of
- 23:08the entire community when we saw that
- 23:10an open source competitor had managed
- 23:12to offer a model that is also up to par
- 23:14. The difference in capabilities
- 23:17between private models, which is the
- 23:18black line we see here, which ends with
- 23:20Fable 5, versus the open source models,
- 23:22which is the blue line and which is not
- 23:24updated. What we're seeing here is GLM
- 23:265.2. Keep this in mind. GLM 5.2 from ZI
- 23:29, which was a model that we celebrated
- 23:31a few weeks ago when it came out
- 23:33because it was a model that, wow, had a
- 23:36lot of capabilities despite being an
- 23:38open source model. Do you see this
- 23:40margin here? This difference. Now look
- 23:42at how Chemistry 3 compares to, well,
- 23:45the Fable 5 models, GPT 5.6 sol, all
- 23:48the Frontier models, and GLM 5.2. Look
- 23:52at the jump in agentic programming
- 23:54capabilities. Deep SWE 67.5 vs. 46.2.
- 23:59Program Bench 77 vs. 63.7, Frontier SWE
- 24:0481 vs. 67.3. Marathon SWE 42 vs. 13.
- 24:10It's outrageous. Still, as you know,
- 24:13these are benchmarks and benchmarks are
- 24:15the devil's work. We all know this, so
- 24:18we must let the community try it out
- 24:20and see for themselves what kind of
- 24:21sensations this new Chemistry 3 model
- 24:23transmits to all of us. Still, to close
- 24:27, I wanted to do a comparative test
- 24:29between Fable 5, GPT 5.6, 6 and
- 24:30Chemistry 3 in one of the tasks that,
- 24:32well, many of us usually care about and
- 24:34we know that in general GPT models have
- 24:36failed enormously, and that is: "
- 24:38Generate a frontend for me, generate
- 24:40something pretty, please, and show it
- 24:42to me on the screen without it being
- 24:44little boxes inside little boxes and
- 24:45purple colors everywhere." And the
- 24:48prompt that I tried was the following.
- 24:51Develop the frontend of what would be a
- 24:53theme park like Disneyland Paris or
- 24:54Disney World, but dedicated to the
- 24:56world of deep learning and artificial
- 24:58intelligence. We are going to mix two
- 25:00worlds here to see how creative the AI
- 25:02is. Generate all the assets needed to
- 25:04create a result that is rich in details
- 25:06and professional design. I will first
- 25:08show you the result that after 16
- 25:11minutes of work GPT 5.5 5 in high
- 25:13configuration has generated for me. The
- 25:16result is not bad, but we see some of
- 25:18the usual flaws of a GPT model, because
- 25:20as I was saying, it used to be a lot of
- 25:22boxes, a format that our eyes are
- 25:24already used to seeing and that stands
- 25:26out a lot, since it is made with
- 25:28artificial intelligence. I think there
- 25:31are some very interesting things. Here
- 25:32it has created assets for us such as
- 25:33the park map itself. Look how we have
- 25:36Neural Kingdom, Vision Valley, Language
- 25:37Lagoon, Robotics Groove, Ethics Garden.
- 25:41It's pretty cool and we can see the
- 25:42different areas of the map. The result
- 25:44is pretty good, we won't deny it, but
- 25:46we could do better. And this is
- 25:48something we have to celebrate because
- 25:50Open AI has been working on this very
- 25:51thing. Look, what we are seeing here is
- 25:54a visualization that encodes different
- 25:56frontends generated by, in this case,
- 25:58GPT 5.6. And notice how it is clearly
- 26:02seen that within this map of different
- 26:03styles, there are a few white gaps that
- 26:05people have been studying to see what
- 26:07they correspond to. And what they have
- 26:09found is that these white spaces, look
- 26:11at this area that is headed by blue
- 26:13dots, that is, frontends that GPT 5.5
- 26:15did generate and that GPT 5.6 or 6 Sol
- 26:17does not generate, otherwise you would
- 26:19see orange dots. Here they correspond
- 26:22to frontends with purple designs or
- 26:24designs that, for example, have boxes,
- 26:26fonts with a specific style. In other
- 26:30words, Open AI has done post-training
- 26:32work to remove, to eliminate these
- 26:34flaws that we know GPT frontends have.
- 26:39Thus, the result generated by the Luna
- 26:41model, the most basic model in the Open
- 26:43AI family, is a slightly more
- 26:44aesthetically pleasing web page. This
- 26:47is already starting to look like what a
- 26:49museum or amusement park page could be.
- 26:51There are certainly errors. We see that
- 26:54there are things that do not load,
- 26:55there is information that is not really
- 26:57generated well, but here we are working
- 26:58with the Luna model which, let's
- 27:00remember, is the least capable. But at
- 27:02least we have a frontend that we could
- 27:04say is a little better and, at least,
- 27:06different from what we are used to. If
- 27:09we now look at the result of GPT SOL,
- 27:11the most advanced model, this is the
- 27:13web page that it has created for us
- 27:15with assets that take advantage of the
- 27:17image generation capacity. Well, we see
- 27:20how these images have been generated to
- 27:22enrich our web page and we already see
- 27:24a very professional web page, okay? I
- 27:27see some things that do continue, well,
- 27:30getting a little closer to the idea of
- 27:32a web page made by artificial
- 27:34intelligence, but we are already
- 27:35starting to see results that are more
- 27:38attractive in terms of design. I really
- 27:41like the idea of Lent Space as a roller
- 27:43coaster of a park about artificial
- 27:45intelligence, quite, quite cool. And to
- 27:49compare Fable 5 with the same prompt,
- 27:51Fable 5 High, returns a web page that
- 27:53is this one here, which of course,
- 27:55since Fable 5 does not have the image
- 27:57generation capacity, it has had to bet
- 28:00a lot on SVG. It's not bad, but it's
- 28:03not the design of a web page as I was
- 28:05thinking, which in this case would be a
- 28:07professional web page for an amusement
- 28:09park and in this case it comes across
- 28:11as something a bit strange. Note that
- 28:13Fable 5 is a model that Antropic models
- 28:15have always been quite competent in
- 28:17terms of design, but I'm not seeing a
- 28:19result here that I can compare, for
- 28:21example, with what GPT 5.6 Sol has done
- 28:24with the same Prompt. What I am seeing
- 28:26is that at least in terms of content,
- 28:28some of the things I'm seeing are much
- 28:30more original, right? Inference input,
- 28:32training pass, family batch. Hm. Surely
- 28:35if you start reading things around here
- 28:37, you'll find some very, very funny
- 28:38details. For example, minimum height.
- 28:41This bothered me a lot when I went to
- 28:42amusement parks as a child, the minimum
- 28:44height thing. 1.20 m or 120 million
- 28:47parameters. Okay, these are little
- 28:49details that I think really demonstrate
- 28:51the originality of these models and
- 28:52where I think Fall 5 has done a better
- 28:54job, right? Look, 1X1 convolution
- 28:56avenue, latent valley 10 minutes from
- 28:58the real world, nothing crazy. All
- 29:02weights reserved. There's a lot,
- 29:05there's a lot of humor here, eh, I like
- 29:07it, I like it, but it's not a
- 29:09professional result. And the latest
- 29:12contender, Kimica 3, tested from the
- 29:14Moonshot platform itself, has generated
- 29:16an agentic loop that we can see here,
- 29:19quite cool as it shows you the
- 29:20evolution of what it's building, okay?
- 29:24It works and works and then it
- 29:26generates this web page that if we look
- 29:28at it in full screen, it shows us
- 29:29Neurolandia. I don't know exactly what
- 29:33image generation model it has used, but
- 29:34anyway, this already shows more
- 29:36integration than Fable 5 by default.
- 29:39And yes, I'm seeing an aesthetic that
- 29:41reminds me more of what GPT 5.5 used to
- 29:44be for us; Well, these chips, these
- 29:46types of boxes, these more purple
- 29:48colors, so many boxes just for the sake
- 29:51of putting boxes. And yes, it reminds
- 29:54me, as I say, more of the design of
- 29:56previous AIs than what they offer us
- 29:58now, well, models like GPT 5.6. Even so
- 30:01, a pretty rich result with lots of
- 30:03little things. Here we have the mascot.
- 30:05Say hello to Neuro. We have information
- 30:08about different parades. We have an
- 30:10attempt at a map with SVG that's also
- 30:12pretty cool. Token Town, Gradient Peaks
- 30:15, Vector Jungle, Generative Gardens.
- 30:17Why isn't there a park for it? It's a
- 30:19pretty cool theme. The Gun mansion as a
- 30:23fast backend attraction, data carousel.
- 30:26Wow, I should have paid for it. And as
- 30:30I said: well, a result that, at the
- 30:31frontend level, doesn't appeal to me as
- 30:33much as the previous one, but that
- 30:35demonstrates the capabilities of this
- 30:37Kimica 3 model, which comes in like a
- 30:38bull in a china shop to turn the world
- 30:40of artificial intelligence upside down,
- 30:42where none of the big labs can rest on
- 30:44their laurels. Google, I speak for you.
- 30:47Where are you? What's wrong? Are you
- 30:49okay? You promised you would release a
- 30:52Gemini 3.5 Pro model in April. You told
- 30:55me, "In a month, Carlos, I'm going to
- 30:57release Gemini 3.5 Pro. What happened?
- 30:59Haven't you been able to keep up with
- 31:01your competitors?" Well, if any move
- 31:03were to happen in the coming weeks,
- 31:05guys and girls, stay tuned because
- 31:06we'll be reporting on it here, we'll be
- 31:08analyzing it sooner or later. Well, it
- 31:10doesn't matter. A rigorous analysis
- 31:12like the ones I always bring you to
- 31:13keep you up to date with what's
- 31:15happening in the world of artificial
- 31:16intelligence. If you liked the video,
- 31:18don't forget to share it. I'll leave
- 31:20you a couple more videos here that you
- 31:22can check out. Thanks again to the
- 31:23sponsor of the video Nord VPN. I'll
- 31:26leave the link to the discount I
- 31:27mentioned at the beginning below and
- 31:29I'll see you with more AI in a future
- 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.