YouTube2Text

Google lanza su MEJOR IA (y sorprende) — Transcript

by Xavier Mitjana · 2,724 words · 412 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Google has just unveiled its best model
  2. 0:02, the new Gemini 3.6 Flash. And yet,
  3. 0:05six labs currently have a better model
  4. 0:08than Google's new one. Six. Even Meta
  5. 0:12is ahead of them. This was the reaction
  6. 0:15on X from their top AI executive. And
  7. 0:17the internet has delivered its verdict
  8. 0:19very quickly. Fiasco. And they might be
  9. 0:22right, but something doesn't add up. No
  10. 0:25one has more money, more engineers, or
  11. 0:27more chips than Google. Why would they
  12. 0:29put out such a clear fiasco on purpose?
  13. 0:32Maybe it's because they are in crisis
  14. 0:34or because in artificial intelligence
  15. 0:36there is no longer just one race, there
  16. 0:38are two, and this launch competes in
  17. 0:40the second one. Today you will discover
  18. 0:42what these two races are and why
  19. 0:44understanding them is key to making
  20. 0:45your daily use of artificial
  21. 0:47intelligence truly productive. And
  22. 0:50let's start by being fair to those
  23. 0:51pointing out the fiasco, because they
  24. 0:53have been given plenty of arguments.
  25. 0:55Yesterday Google launched three models
  26. 0:57at once, the Gemini 3.6 Flash, the 3.5
  27. 0:59Flashlite, and a version specialized in
  28. 1:02cybersecurity. But all eyes were on the
  29. 1:04first one, and for good reason. Flash
  30. 1:06is Google's workhorse, the default
  31. 1:08model that responds when you open the
  32. 1:10Gemini app and the one Google wants you
  33. 1:12to use for everything. On paper, the
  34. 1:14proposal for this new model sounds
  35. 1:16reasonable. Cheaper than its
  36. 1:18predecessor, it consumes 17%fewer
  37. 1:20tokens to solve the same task, which in
  38. 1:22other words means it does the same
  39. 1:24while spending less, but it is not a
  40. 1:26smarter model than the Gemini 3.5 Flash
  41. 1:28. In the Artificial Analysis index, the
  42. 1:32Gemini 3.6 Flash stays at 50 points,
  43. 1:35exactly the same score as the 3.5, and
  44. 1:38above it are models from six different
  45. 1:41labs: Anthropic, OpenAI, Moonshot, xAI,
  46. 1:44Zhipu, and Meta. Google is out of the
  47. 1:47top five labs for the first time, and
  48. 1:49when you go from benchmarks to a real
  49. 1:51task, things don't seem to improve. The
  50. 1:54Command Code team had three models
  51. 1:56program the same game with a single
  52. 1:58instruction. Claude 3 nailed it on the
  53. 2:00first try. 10 out of 10, an impeccable
  54. 2:03result. Gemini 3.6 Flash, well, it
  55. 2:05needed five or six iterations to
  56. 2:07deliver a worse result. If intelligence
  57. 2:10ranking is the only yardstick, there is
  58. 2:12no possible defense. Google's fiasco is
  59. 2:15real, but model intelligence is no
  60. 2:17longer the only thing that matters. And
  61. 2:20there is one piece of data from
  62. 2:21Artificial Analysis itself that is key.
  63. 2:23Despite having the same intelligence,
  64. 2:25it is now much more efficient at
  65. 2:26solving the same task. In other words,
  66. 2:29it's a cheaper model, and to understand
  67. 2:31it, let's analyze this chart carefully.
  68. 2:34This is the map of artificial
  69. 2:35intelligence today. Vertically,
  70. 2:38intelligence. Horizontally, the cost to
  71. 2:40solve a task. And the usual way to read
  72. 2:43it is vertically. The one that reaches
  73. 2:45the highest wins. But now read it
  74. 2:47horizontally, and two zones appear. On
  75. 2:51the right, the frontier models: Claude
  76. 2:533.5, GPT-4o, Claude Opus—maximum
  77. 2:55intelligence, paying up to 10 times
  78. 2:57more per task than the rest. And in the
  79. 3:01center, you find GPT-4o mini, Grok 2,
  80. 3:05Llama 3 of Meta, and Gemini 1.5 Flash.
  81. 3:08Models that hover around 50
  82. 3:10intelligence points, but are much
  83. 3:12cheaper. This is the real league where
  84. 3:15Gemini plays. Its rival isn't Claude
  85. 3:173.5, it's GPT-4o. Mini. And why does
  86. 3:21this zone exist? Because all those
  87. 3:24models compete for the same thing:
  88. 3:26making AI agents automate tasks where
  89. 3:28they must chain many steps one after
  90. 3:29another. In other words, they want to
  91. 3:33be good in the world of agents. When
  92. 3:35you ask a chatbot something, it’s a
  93. 3:37query: a question, a response. But when
  94. 3:40an agent works for you, it executes
  95. 3:42hundreds of steps, sometimes thousands.
  96. 3:44It reads, plans, tests, makes mistakes,
  97. 3:47and corrects. And there, the cost per
  98. 3:50task is no longer an unimportant detail
  99. 3:52, since each step can accumulate tens
  100. 3:54of thousands of tokens. Efficiency is
  101. 3:57what decides whether an agent is worth
  102. 3:59it or if it bankrupts you. So, once we
  103. 4:01have this clear, it's time to put it to
  104. 4:02the test against its real competition.
  105. 4:04We're going to test it against GPT-4o
  106. 4:07mini to achieve a goal where they must
  107. 4:09chain multiple tasks. And this is the
  108. 4:13goal. Each model must create a video
  109. 4:16explaining the launch of Gemini 1.5
  110. 4:18Flash using Hyperframes. This is the
  111. 4:21type of work that artificial
  112. 4:22intelligence tools promise to solve
  113. 4:24from now on. Before we see them in
  114. 4:25action, which one do you think will do
  115. 4:27better? And here you have both tests
  116. 4:29I've run in each model's own harness.
  117. 4:31That is, to do the test with GPT-4o
  118. 4:33mini. I did it in ChatGPT, while to run
  119. 4:36the test with the Google model, I did
  120. 4:38it in AI Studio. And this is one of the
  121. 4:41biggest handicaps of using artificial
  122. 4:43intelligence. If you're like me, you
  123. 4:45probably have a subscription for an AI
  124. 4:47chat, another to generate images,
  125. 4:49another for presentations, another to
  126. 4:51build websites, and you end up jumping
  127. 4:53between 15 tabs to finish a single
  128. 4:55project. Jen Spark, which sponsors this
  129. 4:57video, has done something that really
  130. 4:59caught my attention: successfully
  131. 5:00unifying all these processes. And the
  132. 5:02fact is, Jen Spark is no small project.
  133. 5:04We are talking about Silicon Valley's
  134. 5:07first AI agent unicorn, a company
  135. 5:09valued at over 1 billion dollars. Jen
  136. 5:12Spark is a workspace that connects all
  137. 5:14the leading AI models and puts them to
  138. 5:17work together to give you finished
  139. 5:18results. You tell it what you need, and
  140. 5:22Jen delivers the finished work. Let me
  141. 5:24give you some real-world examples. With
  142. 5:26Jenark Code, I asked for a website, and
  143. 5:27it built the whole thing without me
  144. 5:29touching a single line of code. With
  145. 5:31Jen Spark Slid 5.0, I gave it a topic,
  146. 5:33and it generated a professional
  147. 5:35presentation. When it comes to creating
  148. 5:37presentations, Jen has two things that
  149. 5:39make all the difference. For one, it
  150. 5:41has "expert thinking," which uses
  151. 5:43expert workflows specific to the field
  152. 5:45to achieve a professional result. And
  153. 5:48for the other, it has a huge library of
  154. 5:49professional templates adapted to
  155. 5:51different sectors and use cases, so you
  156. 5:53never really start from scratch. And
  157. 5:55once the design and graphics are
  158. 5:57finished, everything is ready to export
  159. 5:59to PowerPoint. And finally, with Jen
  160. 6:02Spark Image, which uses Nano Banana 2,
  161. 6:04I created images for social media
  162. 6:05without opening any other tool. The
  163. 6:08best part is that with a single prompt,
  164. 6:09you get several images at once, so you
  165. 6:11can explore various creative directions
  166. 6:13at once. Compare that, for example, to
  167. 6:16Chat GPT, which usually gives you only
  168. 6:17one image per attempt. And these same
  169. 6:20images can be turned into video with a
  170. 6:21single click without leaving Jen Spark.
  171. 6:24And what I like most is that here,
  172. 6:25instead of paying for four or five
  173. 6:27different subscriptions, you have
  174. 6:29everything in a single workspace for $
  175. 6:3119.99 a month. And right now, you can
  176. 6:33try it for free with welcome credits.
  177. 6:35You will find the link in the
  178. 6:37description. And let's continue with
  179. 6:39this test I conducted on both models.
  180. 6:41As I was saying, in the case of GPT 5.6
  181. 6:43Luna, we did it within its own harness,
  182. 6:45which is Codex. I simply asked it to
  183. 6:47create a video with Hyperframes; it
  184. 6:49installed the Hyperframes skills, and
  185. 6:51then I told it what video I wanted. A
  186. 6:5330-second promo explaining the new
  187. 6:55Gemini 3.6 Flash features, and I gave
  188. 6:57it the link to the news on the Google
  189. 6:59blog. From there, it simply defined the
  190. 7:02brief; I told it to continue, and with
  191. 7:05that, it delivered the video, which
  192. 7:07looks like this. What if your agent
  193. 7:09could think more while spending less?
  194. 7:11Introducing Gemini 3.6 Flash. Code
  195. 7:14knowledge and multimodal towers. Writes
  196. 7:17code. Understands long contexts. Works
  197. 7:21with context, images, audio, video, and
  198. 7:24PDFs. Result: up to 17%fewer output
  199. 7:28tokens and lower cost per task.
  200. 7:31Available on Gemini API, Google AI
  201. 7:34Studio, and the Gemini app. Start
  202. 7:36building today.
  203. 7:38On the other hand, the experience with
  204. 7:39Antigravity hasn't been as satisfactory
  205. 7:41. Why? Because I asked it for exactly
  206. 7:43the same thing. The first instruction:
  207. 7:45I want to create a video with
  208. 7:46hyperframes. From there, the second
  209. 7:48instruction was the same. A 30-second
  210. 7:50promo based on the blog post. And the
  211. 7:54problem we had here is that once I
  212. 7:56validated the brief, it delivered a
  213. 7:59version without sound, so I had to give
  214. 8:02it a follow-up instruction telling it I
  215. 8:04wanted a voiceover in Spanish and audio
  216. 8:07effects. And here I think it got a bit
  217. 8:11confused when using the skills and
  218. 8:13returned a result like this.
  219. 8:16Introducing Cheminite 3.6 Flash.
  220. 8:20Advanced reasoning quality with fewer
  221. 8:24tokens, 12%faster, and up to 20%better
  222. 8:27performance in a truly multimodal
  223. 8:31architecture. Agentic code, vision,
  224. 8:35audio, and deep reasoning. Pro-level
  225. 8:39intelligence at flash speed. Start
  226. 8:42building today in Google Studio.
  227. 8:45The result, as you could see, is that
  228. 8:47Luna got there more directly and did it
  229. 8:49better. It understood the task and
  230. 8:51executed it almost in one go. Gemini
  231. 8:53needed an extra correction along the
  232. 8:55way, and one of its decisions took a
  233. 8:57toll, a slight audio desync that taints
  234. 8:59the final result. Mind you, the path it
  235. 9:02chose complicated the process, but even
  236. 9:05so, if you put the two videos side by
  237. 9:07side, the difference is clear. GPT 5.6,
  238. 9:10Luna is better. If you bet on Luna, you
  239. 9:13were right. If you bet on Gemini, note
  240. 9:16the nuance. With a specific prompt
  241. 9:18about the process, it could have
  242. 9:19completed the task, but the fact is
  243. 9:21that it did worse. Although in this
  244. 9:23example, I want you to focus on another
  245. 9:25detail. Both did a job that a year and
  246. 9:27a half ago would have seemed like
  247. 9:29science fiction, even for the most
  248. 9:31powerful model of the time. And that is
  249. 9:33the key that the big artificial
  250. 9:34intelligence companies have already
  251. 9:36understood. For many common tasks, you
  252. 9:38no longer need the best model; you need
  253. 9:41the right model. And I'm going to prove
  254. 9:43it to you. A few days ago, I did this
  255. 9:44same test with Chemite 3. Chemite 3
  256. 9:47scores 57 in the Artificial Analysis
  257. 9:50ranking, seven points above Gemini and
  258. 9:52six points above GPT 5.6 Luna. One of
  259. 9:56the most powerful models on the market.
  260. 9:57It competes directly with Fabel and GPT
  261. 9:595.6. Well, look at its result when I
  262. 10:03asked it to create a video to explain
  263. 10:05itself.
  264. 10:06Everyone claims frontier intelligence.
  265. 10:09We bring the receipts. Kimi k-3. 2.8
  266. 10:12trillion parameters. 1 million context
  267. 10:16tokens. Open. Case 001. A chip designed
  268. 10:20by the model itself in 48 hours. 8700
  269. 10:24tokens per second. Case 002. A GPU
  270. 10:28compiler written from scratch, on par
  271. 10:30with and sometimes above Triden. Case
  272. 10:33003. Two weeks of research solved in 2
  273. 10:37hours. 20 papers. 3000 lines of code.
  274. 10:41Frontier performance. The best open
  275. 10:43model in the world. Kimi k-3. Available
  276. 10:46today at kimi.com. The open frontier.
  277. 10:49As you can see, the results are
  278. 10:51perfectly comparable. Although Gemini
  279. 10:53lags a bit behind, we see how a model
  280. 10:55in its category, Luna, competes with
  281. 10:57another at the frontier, in this case
  282. 10:59Chemistry 3. In fact, I would dare to
  283. 11:02say that Luna did a little better than
  284. 11:04Kimi, and this is what will throw off
  285. 11:06many analyses from now on. If a model
  286. 11:08that can be considered a flop and an
  287. 11:10outstanding model can deliver similar
  288. 11:12results in the most common tasks, what
  289. 11:14exactly are you buying when you pay for
  290. 11:16the smartest model? And at this point,
  291. 11:19we must understand that the main
  292. 11:21conclusion of this video is more about
  293. 11:23you as a user of tools than about the
  294. 11:24latest Google novelty. We have spent
  295. 11:27years measuring all models from a
  296. 11:28single point of view. Which one is
  297. 11:31smarter? Which one is more capable?
  298. 11:33Every launch, the same ritual. Open the
  299. 11:35ranking, see who is at the top. And
  300. 11:37it's normal that it was this way,
  301. 11:38because until a couple of years ago
  302. 11:40this was very relevant as models were
  303. 11:42still very limited. But that is over.
  304. 11:46When it comes to completing more and
  305. 11:48more relevant tasks in the real world,
  306. 11:50these extra points of intelligence
  307. 11:51affect the result less and less, so
  308. 11:53measuring intelligence may have ceased
  309. 11:55to mean measuring what concerns the
  310. 11:57user most today: their bill. So, why
  311. 12:00are there labs prioritizing moving up
  312. 12:02one more point in that ranking? Because
  313. 12:05not everyone prioritizes the same race.
  314. 12:10The first race is that of capability,
  315. 12:12being the smartest at any cost. This
  316. 12:15competition is prioritized above all by
  317. 12:18pure labs, Open Console or Antropicon
  318. 12:20Fabel. Why? Because they live off that.
  319. 12:22A lab without another business needs
  320. 12:24challenges. They are their argument for
  321. 12:27securing the next billions in
  322. 12:28investment. Even if those models
  323. 12:30currently lose money when they serve
  324. 12:32them to their users. The benchmark for
  325. 12:35them is not just a product metric; it
  326. 12:37is also a promise to investors. They
  327. 12:39need to be at the frontier. In contrast
  328. 12:42, the second race consists of
  329. 12:43prioritizing efficiency—models that
  330. 12:45are good enough at the lowest possible
  331. 12:47cost—and that is where companies
  332. 12:48whose business is different compete.
  333. 12:51Meta's business is not selling models,
  334. 12:52it is selling advertising, and
  335. 12:54artificial intelligence helps its
  336. 12:55social networks function better. Think
  337. 12:57of X, which answers millions of
  338. 13:00questions every day. There, every cent
  339. 13:02per answer counts. And Google, Google
  340. 13:05has an entire empire of services—the
  341. 13:07search engine, Gmail, Docs, Android,
  342. 13:09YouTube—that can improve thanks to
  343. 13:11artificial intelligence, but that
  344. 13:13improvement is only justified if it
  345. 13:15pays off. For them, the model is not
  346. 13:17the product, it is the infrastructure.
  347. 13:19And although deep down they all have
  348. 13:21arguments to say they compete on both
  349. 13:23fronts, they prioritize different
  350. 13:24things. And the fact is that, at this
  351. 13:26point, it seems Google is still
  352. 13:27competing and promising top-tier models
  353. 13:29. In the announcement of Gemini 3.6
  354. 13:31Flash, Google confirmed that Gemini 3.6
  355. 13:33Pro is already in the development phase
  356. 13:35, that they have given access to the
  357. 13:37first companies, and that they have
  358. 13:39also started training Gemini 4, in
  359. 13:41their own words, the most ambitious
  360. 13:42model in their history. But the reality
  361. 13:45is that they have presented a model
  362. 13:47that seems to not be at the level of
  363. 13:48its competition, and we still haven't
  364. 13:50resolved the crux of the matter. Is
  365. 13:52Gemini 3.6 Flash good or is it bad?
  366. 13:56Well, look, perhaps that is not the
  367. 13:58right question. If you only look at
  368. 14:02Gemini 3.6 Flash in the race for raw
  369. 14:05power, it loses. You've already seen it
  370. 14:07; it has up to six labs with better
  371. 14:09models ahead of it. But if you look at
  372. 14:12specific capabilities, such as
  373. 14:13interpreting charts and documents,
  374. 14:15which would be the ChartQA benchmark,
  375. 14:17it has the best result of the entire
  376. 14:18comparison there. When working with
  377. 14:21huge documents, it far exceeds its
  378. 14:23predecessor, going from 27%to 54%. And
  379. 14:27in computer use, that is to say,
  380. 14:29handling a computer as you would, it
  381. 14:32reaches 83%in the OSWorld benchmark.
  382. 14:34Ahead not only of its predecessor, but
  383. 14:37also ahead of GPT 5.6 Luna and Grok 4.5
  384. 14:39. And all that while solving tasks in a
  385. 14:42more efficient way than its previous
  386. 14:44version. We often complain that a model
  387. 14:47isn't Einstein for tasks that don't
  388. 14:49need an Einstein. You don't need an
  389. 14:51Einstein to grade middle school
  390. 14:52homework, for example. The key lies in
  391. 14:55getting the model right. After all, no
  392. 14:58one builds a company by hiring only
  393. 14:59Einsteins. You hire an architect to
  394. 15:01design the building, workers to build
  395. 15:03it, and someone to check that it
  396. 15:04doesn't fall down. Each one is the
  397. 15:06right fit for their specific part. That
  398. 15:08is where artificial intelligence is
  399. 15:10heading. The near future isn't about
  400. 15:12choosing one model, but orchestrating
  401. 15:13them. A large model to plan, an
  402. 15:15efficient one to iterate and execute
  403. 15:17without going broke, and another to
  404. 15:19audit the result. A key skill in the
  405. 15:21future will be knowing how to build and
  406. 15:23maintain systems that are very good and
  407. 15:25cheap at solving specific scenarios.
  408. 15:27And at this point, if you want to see a
  409. 15:30full test of the model threatening the
  410. 15:32leadership of Fabel 5 and GPT 5.6 Sol,
  411. 15:34the Chemist 3, you can see it in this
  412. 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.