Google lanza su MEJOR IA (y sorprende) — Transcript
Full transcript
- 0:00Google has just unveiled its best model
- 0:02, the new Gemini 3.6 Flash. And yet,
- 0:05six labs currently have a better model
- 0:08than Google's new one. Six. Even Meta
- 0:12is ahead of them. This was the reaction
- 0:15on X from their top AI executive. And
- 0:17the internet has delivered its verdict
- 0:19very quickly. Fiasco. And they might be
- 0:22right, but something doesn't add up. No
- 0:25one has more money, more engineers, or
- 0:27more chips than Google. Why would they
- 0:29put out such a clear fiasco on purpose?
- 0:32Maybe it's because they are in crisis
- 0:34or because in artificial intelligence
- 0:36there is no longer just one race, there
- 0:38are two, and this launch competes in
- 0:40the second one. Today you will discover
- 0:42what these two races are and why
- 0:44understanding them is key to making
- 0:45your daily use of artificial
- 0:47intelligence truly productive. And
- 0:50let's start by being fair to those
- 0:51pointing out the fiasco, because they
- 0:53have been given plenty of arguments.
- 0:55Yesterday Google launched three models
- 0:57at once, the Gemini 3.6 Flash, the 3.5
- 0:59Flashlite, and a version specialized in
- 1:02cybersecurity. But all eyes were on the
- 1:04first one, and for good reason. Flash
- 1:06is Google's workhorse, the default
- 1:08model that responds when you open the
- 1:10Gemini app and the one Google wants you
- 1:12to use for everything. On paper, the
- 1:14proposal for this new model sounds
- 1:16reasonable. Cheaper than its
- 1:18predecessor, it consumes 17%fewer
- 1:20tokens to solve the same task, which in
- 1:22other words means it does the same
- 1:24while spending less, but it is not a
- 1:26smarter model than the Gemini 3.5 Flash
- 1:28. In the Artificial Analysis index, the
- 1:32Gemini 3.6 Flash stays at 50 points,
- 1:35exactly the same score as the 3.5, and
- 1:38above it are models from six different
- 1:41labs: Anthropic, OpenAI, Moonshot, xAI,
- 1:44Zhipu, and Meta. Google is out of the
- 1:47top five labs for the first time, and
- 1:49when you go from benchmarks to a real
- 1:51task, things don't seem to improve. The
- 1:54Command Code team had three models
- 1:56program the same game with a single
- 1:58instruction. Claude 3 nailed it on the
- 2:00first try. 10 out of 10, an impeccable
- 2:03result. Gemini 3.6 Flash, well, it
- 2:05needed five or six iterations to
- 2:07deliver a worse result. If intelligence
- 2:10ranking is the only yardstick, there is
- 2:12no possible defense. Google's fiasco is
- 2:15real, but model intelligence is no
- 2:17longer the only thing that matters. And
- 2:20there is one piece of data from
- 2:21Artificial Analysis itself that is key.
- 2:23Despite having the same intelligence,
- 2:25it is now much more efficient at
- 2:26solving the same task. In other words,
- 2:29it's a cheaper model, and to understand
- 2:31it, let's analyze this chart carefully.
- 2:34This is the map of artificial
- 2:35intelligence today. Vertically,
- 2:38intelligence. Horizontally, the cost to
- 2:40solve a task. And the usual way to read
- 2:43it is vertically. The one that reaches
- 2:45the highest wins. But now read it
- 2:47horizontally, and two zones appear. On
- 2:51the right, the frontier models: Claude
- 2:533.5, GPT-4o, Claude Opus—maximum
- 2:55intelligence, paying up to 10 times
- 2:57more per task than the rest. And in the
- 3:01center, you find GPT-4o mini, Grok 2,
- 3:05Llama 3 of Meta, and Gemini 1.5 Flash.
- 3:08Models that hover around 50
- 3:10intelligence points, but are much
- 3:12cheaper. This is the real league where
- 3:15Gemini plays. Its rival isn't Claude
- 3:173.5, it's GPT-4o. Mini. And why does
- 3:21this zone exist? Because all those
- 3:24models compete for the same thing:
- 3:26making AI agents automate tasks where
- 3:28they must chain many steps one after
- 3:29another. In other words, they want to
- 3:33be good in the world of agents. When
- 3:35you ask a chatbot something, it’s a
- 3:37query: a question, a response. But when
- 3:40an agent works for you, it executes
- 3:42hundreds of steps, sometimes thousands.
- 3:44It reads, plans, tests, makes mistakes,
- 3:47and corrects. And there, the cost per
- 3:50task is no longer an unimportant detail
- 3:52, since each step can accumulate tens
- 3:54of thousands of tokens. Efficiency is
- 3:57what decides whether an agent is worth
- 3:59it or if it bankrupts you. So, once we
- 4:01have this clear, it's time to put it to
- 4:02the test against its real competition.
- 4:04We're going to test it against GPT-4o
- 4:07mini to achieve a goal where they must
- 4:09chain multiple tasks. And this is the
- 4:13goal. Each model must create a video
- 4:16explaining the launch of Gemini 1.5
- 4:18Flash using Hyperframes. This is the
- 4:21type of work that artificial
- 4:22intelligence tools promise to solve
- 4:24from now on. Before we see them in
- 4:25action, which one do you think will do
- 4:27better? And here you have both tests
- 4:29I've run in each model's own harness.
- 4:31That is, to do the test with GPT-4o
- 4:33mini. I did it in ChatGPT, while to run
- 4:36the test with the Google model, I did
- 4:38it in AI Studio. And this is one of the
- 4:41biggest handicaps of using artificial
- 4:43intelligence. If you're like me, you
- 4:45probably have a subscription for an AI
- 4:47chat, another to generate images,
- 4:49another for presentations, another to
- 4:51build websites, and you end up jumping
- 4:53between 15 tabs to finish a single
- 4:55project. Jen Spark, which sponsors this
- 4:57video, has done something that really
- 4:59caught my attention: successfully
- 5:00unifying all these processes. And the
- 5:02fact is, Jen Spark is no small project.
- 5:04We are talking about Silicon Valley's
- 5:07first AI agent unicorn, a company
- 5:09valued at over 1 billion dollars. Jen
- 5:12Spark is a workspace that connects all
- 5:14the leading AI models and puts them to
- 5:17work together to give you finished
- 5:18results. You tell it what you need, and
- 5:22Jen delivers the finished work. Let me
- 5:24give you some real-world examples. With
- 5:26Jenark Code, I asked for a website, and
- 5:27it built the whole thing without me
- 5:29touching a single line of code. With
- 5:31Jen Spark Slid 5.0, I gave it a topic,
- 5:33and it generated a professional
- 5:35presentation. When it comes to creating
- 5:37presentations, Jen has two things that
- 5:39make all the difference. For one, it
- 5:41has "expert thinking," which uses
- 5:43expert workflows specific to the field
- 5:45to achieve a professional result. And
- 5:48for the other, it has a huge library of
- 5:49professional templates adapted to
- 5:51different sectors and use cases, so you
- 5:53never really start from scratch. And
- 5:55once the design and graphics are
- 5:57finished, everything is ready to export
- 5:59to PowerPoint. And finally, with Jen
- 6:02Spark Image, which uses Nano Banana 2,
- 6:04I created images for social media
- 6:05without opening any other tool. The
- 6:08best part is that with a single prompt,
- 6:09you get several images at once, so you
- 6:11can explore various creative directions
- 6:13at once. Compare that, for example, to
- 6:16Chat GPT, which usually gives you only
- 6:17one image per attempt. And these same
- 6:20images can be turned into video with a
- 6:21single click without leaving Jen Spark.
- 6:24And what I like most is that here,
- 6:25instead of paying for four or five
- 6:27different subscriptions, you have
- 6:29everything in a single workspace for $
- 6:3119.99 a month. And right now, you can
- 6:33try it for free with welcome credits.
- 6:35You will find the link in the
- 6:37description. And let's continue with
- 6:39this test I conducted on both models.
- 6:41As I was saying, in the case of GPT 5.6
- 6:43Luna, we did it within its own harness,
- 6:45which is Codex. I simply asked it to
- 6:47create a video with Hyperframes; it
- 6:49installed the Hyperframes skills, and
- 6:51then I told it what video I wanted. A
- 6:5330-second promo explaining the new
- 6:55Gemini 3.6 Flash features, and I gave
- 6:57it the link to the news on the Google
- 6:59blog. From there, it simply defined the
- 7:02brief; I told it to continue, and with
- 7:05that, it delivered the video, which
- 7:07looks like this. What if your agent
- 7:09could think more while spending less?
- 7:11Introducing Gemini 3.6 Flash. Code
- 7:14knowledge and multimodal towers. Writes
- 7:17code. Understands long contexts. Works
- 7:21with context, images, audio, video, and
- 7:24PDFs. Result: up to 17%fewer output
- 7:28tokens and lower cost per task.
- 7:31Available on Gemini API, Google AI
- 7:34Studio, and the Gemini app. Start
- 7:36building today.
- 7:38On the other hand, the experience with
- 7:39Antigravity hasn't been as satisfactory
- 7:41. Why? Because I asked it for exactly
- 7:43the same thing. The first instruction:
- 7:45I want to create a video with
- 7:46hyperframes. From there, the second
- 7:48instruction was the same. A 30-second
- 7:50promo based on the blog post. And the
- 7:54problem we had here is that once I
- 7:56validated the brief, it delivered a
- 7:59version without sound, so I had to give
- 8:02it a follow-up instruction telling it I
- 8:04wanted a voiceover in Spanish and audio
- 8:07effects. And here I think it got a bit
- 8:11confused when using the skills and
- 8:13returned a result like this.
- 8:16Introducing Cheminite 3.6 Flash.
- 8:20Advanced reasoning quality with fewer
- 8:24tokens, 12%faster, and up to 20%better
- 8:27performance in a truly multimodal
- 8:31architecture. Agentic code, vision,
- 8:35audio, and deep reasoning. Pro-level
- 8:39intelligence at flash speed. Start
- 8:42building today in Google Studio.
- 8:45The result, as you could see, is that
- 8:47Luna got there more directly and did it
- 8:49better. It understood the task and
- 8:51executed it almost in one go. Gemini
- 8:53needed an extra correction along the
- 8:55way, and one of its decisions took a
- 8:57toll, a slight audio desync that taints
- 8:59the final result. Mind you, the path it
- 9:02chose complicated the process, but even
- 9:05so, if you put the two videos side by
- 9:07side, the difference is clear. GPT 5.6,
- 9:10Luna is better. If you bet on Luna, you
- 9:13were right. If you bet on Gemini, note
- 9:16the nuance. With a specific prompt
- 9:18about the process, it could have
- 9:19completed the task, but the fact is
- 9:21that it did worse. Although in this
- 9:23example, I want you to focus on another
- 9:25detail. Both did a job that a year and
- 9:27a half ago would have seemed like
- 9:29science fiction, even for the most
- 9:31powerful model of the time. And that is
- 9:33the key that the big artificial
- 9:34intelligence companies have already
- 9:36understood. For many common tasks, you
- 9:38no longer need the best model; you need
- 9:41the right model. And I'm going to prove
- 9:43it to you. A few days ago, I did this
- 9:44same test with Chemite 3. Chemite 3
- 9:47scores 57 in the Artificial Analysis
- 9:50ranking, seven points above Gemini and
- 9:52six points above GPT 5.6 Luna. One of
- 9:56the most powerful models on the market.
- 9:57It competes directly with Fabel and GPT
- 9:595.6. Well, look at its result when I
- 10:03asked it to create a video to explain
- 10:05itself.
- 10:06Everyone claims frontier intelligence.
- 10:09We bring the receipts. Kimi k-3. 2.8
- 10:12trillion parameters. 1 million context
- 10:16tokens. Open. Case 001. A chip designed
- 10:20by the model itself in 48 hours. 8700
- 10:24tokens per second. Case 002. A GPU
- 10:28compiler written from scratch, on par
- 10:30with and sometimes above Triden. Case
- 10:33003. Two weeks of research solved in 2
- 10:37hours. 20 papers. 3000 lines of code.
- 10:41Frontier performance. The best open
- 10:43model in the world. Kimi k-3. Available
- 10:46today at kimi.com. The open frontier.
- 10:49As you can see, the results are
- 10:51perfectly comparable. Although Gemini
- 10:53lags a bit behind, we see how a model
- 10:55in its category, Luna, competes with
- 10:57another at the frontier, in this case
- 10:59Chemistry 3. In fact, I would dare to
- 11:02say that Luna did a little better than
- 11:04Kimi, and this is what will throw off
- 11:06many analyses from now on. If a model
- 11:08that can be considered a flop and an
- 11:10outstanding model can deliver similar
- 11:12results in the most common tasks, what
- 11:14exactly are you buying when you pay for
- 11:16the smartest model? And at this point,
- 11:19we must understand that the main
- 11:21conclusion of this video is more about
- 11:23you as a user of tools than about the
- 11:24latest Google novelty. We have spent
- 11:27years measuring all models from a
- 11:28single point of view. Which one is
- 11:31smarter? Which one is more capable?
- 11:33Every launch, the same ritual. Open the
- 11:35ranking, see who is at the top. And
- 11:37it's normal that it was this way,
- 11:38because until a couple of years ago
- 11:40this was very relevant as models were
- 11:42still very limited. But that is over.
- 11:46When it comes to completing more and
- 11:48more relevant tasks in the real world,
- 11:50these extra points of intelligence
- 11:51affect the result less and less, so
- 11:53measuring intelligence may have ceased
- 11:55to mean measuring what concerns the
- 11:57user most today: their bill. So, why
- 12:00are there labs prioritizing moving up
- 12:02one more point in that ranking? Because
- 12:05not everyone prioritizes the same race.
- 12:10The first race is that of capability,
- 12:12being the smartest at any cost. This
- 12:15competition is prioritized above all by
- 12:18pure labs, Open Console or Antropicon
- 12:20Fabel. Why? Because they live off that.
- 12:22A lab without another business needs
- 12:24challenges. They are their argument for
- 12:27securing the next billions in
- 12:28investment. Even if those models
- 12:30currently lose money when they serve
- 12:32them to their users. The benchmark for
- 12:35them is not just a product metric; it
- 12:37is also a promise to investors. They
- 12:39need to be at the frontier. In contrast
- 12:42, the second race consists of
- 12:43prioritizing efficiency—models that
- 12:45are good enough at the lowest possible
- 12:47cost—and that is where companies
- 12:48whose business is different compete.
- 12:51Meta's business is not selling models,
- 12:52it is selling advertising, and
- 12:54artificial intelligence helps its
- 12:55social networks function better. Think
- 12:57of X, which answers millions of
- 13:00questions every day. There, every cent
- 13:02per answer counts. And Google, Google
- 13:05has an entire empire of services—the
- 13:07search engine, Gmail, Docs, Android,
- 13:09YouTube—that can improve thanks to
- 13:11artificial intelligence, but that
- 13:13improvement is only justified if it
- 13:15pays off. For them, the model is not
- 13:17the product, it is the infrastructure.
- 13:19And although deep down they all have
- 13:21arguments to say they compete on both
- 13:23fronts, they prioritize different
- 13:24things. And the fact is that, at this
- 13:26point, it seems Google is still
- 13:27competing and promising top-tier models
- 13:29. In the announcement of Gemini 3.6
- 13:31Flash, Google confirmed that Gemini 3.6
- 13:33Pro is already in the development phase
- 13:35, that they have given access to the
- 13:37first companies, and that they have
- 13:39also started training Gemini 4, in
- 13:41their own words, the most ambitious
- 13:42model in their history. But the reality
- 13:45is that they have presented a model
- 13:47that seems to not be at the level of
- 13:48its competition, and we still haven't
- 13:50resolved the crux of the matter. Is
- 13:52Gemini 3.6 Flash good or is it bad?
- 13:56Well, look, perhaps that is not the
- 13:58right question. If you only look at
- 14:02Gemini 3.6 Flash in the race for raw
- 14:05power, it loses. You've already seen it
- 14:07; it has up to six labs with better
- 14:09models ahead of it. But if you look at
- 14:12specific capabilities, such as
- 14:13interpreting charts and documents,
- 14:15which would be the ChartQA benchmark,
- 14:17it has the best result of the entire
- 14:18comparison there. When working with
- 14:21huge documents, it far exceeds its
- 14:23predecessor, going from 27%to 54%. And
- 14:27in computer use, that is to say,
- 14:29handling a computer as you would, it
- 14:32reaches 83%in the OSWorld benchmark.
- 14:34Ahead not only of its predecessor, but
- 14:37also ahead of GPT 5.6 Luna and Grok 4.5
- 14:39. And all that while solving tasks in a
- 14:42more efficient way than its previous
- 14:44version. We often complain that a model
- 14:47isn't Einstein for tasks that don't
- 14:49need an Einstein. You don't need an
- 14:51Einstein to grade middle school
- 14:52homework, for example. The key lies in
- 14:55getting the model right. After all, no
- 14:58one builds a company by hiring only
- 14:59Einsteins. You hire an architect to
- 15:01design the building, workers to build
- 15:03it, and someone to check that it
- 15:04doesn't fall down. Each one is the
- 15:06right fit for their specific part. That
- 15:08is where artificial intelligence is
- 15:10heading. The near future isn't about
- 15:12choosing one model, but orchestrating
- 15:13them. A large model to plan, an
- 15:15efficient one to iterate and execute
- 15:17without going broke, and another to
- 15:19audit the result. A key skill in the
- 15:21future will be knowing how to build and
- 15:23maintain systems that are very good and
- 15:25cheap at solving specific scenarios.
- 15:27And at this point, if you want to see a
- 15:30full test of the model threatening the
- 15:32leadership of Fabel 5 and GPT 5.6 Sol,
- 15:34the Chemist 3, you can see it in this
- 15:37video. See you in the next video.
About this transcript
This page contains the full transcript of Google lanza su MEJOR IA (y sorprende) by Xavier Mitjana, generated from the public captions YouTube serves with the video. The transcript has 2,724 words across 412 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.