YouTube2Text

Gemini 4 Hydrogen è davvero così potente? (testato personalmente) — Transcript

by Simone Rizzo · 2,854 words · 463 segments · language en · Watch on YouTube

Full transcript

  1. 0:02Google is finally more active and has
  2. 0:04recently released Gemini 4 Hydrogen, a
  3. 0:06new multimodal model named after an
  4. 0:08element of the periodic table, so we
  5. 0:10have Argon, Carbon which is the smartest
  6. 0:12, and now Hydrogen.
  7. 0:18This model variant is designed for daily
  8. 0:20use, replacing the previous 3.8 Flash,
  9. 0:23and the interesting thing is that it is
  10. 0:25fast, intelligent, and relatively
  11. 0:27inexpensive, at $ 1.50 per million input
  12. 0:30tokens and $ 7.50 per million output
  13. 0:32tokens.
  14. 0:35And in today's video, I want to test
  15. 0:37this model together with you.
  16. 0:42Let’s look at all its features, and
  17. 0:44there is also news from Gemini; they
  18. 0:46released the Decisions API, and so they
  19. 0:48have also released a Decision Model,
  20. 0:50Jev-style, which is very interesting to
  21. 0:52see because, as you know, I have pushed
  22. 0:54hard on this new family of models and
  23. 0:56every day there is a new confirmation,
  24. 0:58as these frontier labs have effectively
  25. 0:59released their own Decision Model.
  26. 1:07Returning to Hydrogen, Google did
  27. 1:08something interesting: the orchestrator
  28. 1:10in Google Antigravity, where they say
  29. 1:12instead of using a single model, they
  30. 1:14are orchestrated effectively and
  31. 1:16efficiently.
  32. 1:21Specifically, this Hydrogen is the
  33. 1:23fastest for everyday work, which is then
  34. 1:25put to work in collaboration with Carbon
  35. 1:27and Argon for High Efficiency Execution
  36. 1:29Engines or Multimodal Workflows.
  37. 1:37Based on benchmarks, this model is
  38. 1:39compared with Claude Sonnet 5.5 and GPT
  39. 1:426.1 Terra; so in terms of intelligence,
  40. 1:44it compares to these two, although by
  41. 1:46benchmarks it seems to outperform them,
  42. 1:49and according to the independent
  43. 1:51Artificial Analysis benchmark, it is
  44. 1:53positioned between GPT 6.1 Terra and
  45. 1:56Claude Sonnet 5.5 Max.
  46. 2:01With 54 intelligence points, it is
  47. 2:03therefore superior to GPT 6.1 Terra and
  48. 2:06slightly inferior to Sonnet 5.5.
  49. 2:10We are now going to test this model in a
  50. 2:12practical field.
  51. 2:15It is a very capable model in coding,
  52. 2:17able to generate 3D objects, perfect for
  53. 2:19gaming, and also capable of creating
  54. 2:21landing pages and web pages with
  55. 2:23excellent UI/UX and animations.
  56. 2:28They even do a stylistic exercise here
  57. 2:30where, in Google Sheets, which is
  58. 2:31basically Excel, they have it paint
  59. 2:33images and create animations.
  60. 2:39Then we have its full capacity to create
  61. 2:41presentations and slides, or to work on
  62. 2:43financial reports, spreadsheets, and
  63. 2:45conduct research, so everything that is
  64. 2:47practically office work.
  65. 2:53Comparing Gemini 4 Hydrogen and Gemini 4
  66. 2:55Carbon, therefore the fastest versus the
  67. 2:58smartest, we really see the difference
  68. 3:01in cost and time, where above we have
  69. 3:03the example: Hydrogen cost $ 4.27 in 1
  70. 3:06hour and 23 minutes, while Carbon cost $
  71. 3:098.83 in 2 hours and 39 minutes.
  72. 3:15Looking at the output results, they are
  73. 3:17more or less the same, very comparable,
  74. 3:19so in a way, for us consumers, it's
  75. 3:21better to use Gemini 4 Hydrogen for most
  76. 3:23tasks because it saves us time and money
  77. 3:26.
  78. 3:29This pattern is also shown by these
  79. 3:31charts, where Gemini 4 Hydrogen is
  80. 3:34practically compared with Sonnet 5.5 and
  81. 3:36GPT 6.1 Terra. We can see here below the
  82. 3:39time taken per task, how many minutes it
  83. 3:41takes to complete the task, and also
  84. 3:43what score it has, so how intelligent it
  85. 3:45is.
  86. 3:47We see that 4 Hydrogen is at the top
  87. 3:49left, so it is more intelligent,
  88. 3:51especially with reasoning, and it takes
  89. 3:53much less time: it takes between 6
  90. 3:55minutes here, while here it's between 10
  91. 3:57and 15.
  92. 4:02While we see how Sonnet even goes above
  93. 4:0430 and 15, whereas Terra takes 7 and 15
  94. 4:07minutes, so there is a difference of a
  95. 4:09few minutes compared to the 6.1 Terra
  96. 4:11and half an hour with Sonnet 5.5.
  97. 4:18The first thing I had it do was create a
  98. 4:203D simulation for my tests using
  99. 4:22Hydrogen with medium reasoning, a 3D
  100. 4:24visualization of how a turbine works.
  101. 4:29So let's open it now; this was all done
  102. 4:32in one shot, a single shot.
  103. 4:35We can basically see the entire 3D
  104. 4:37visualization of a turbine, so this is
  105. 4:39what it looks like from the outside when
  106. 4:41you see an airplane turbine, and this is
  107. 4:43the internal part.
  108. 4:50We can increase the throttle here, it's
  109. 4:52written there, we can change the axial
  110. 4:53exploded view, we can click on every
  111. 4:55component and read the description, so
  112. 4:57we have the fan, the compressor, the
  113. 4:59other compressor and combustor, turbine
  114. 5:01, and the other turbine.
  115. 5:08Then we have various view modes, X-ray,
  116. 5:11thermal.
  117. 5:13Oh, wow, all the heat here is in the
  118. 5:15combustor.
  119. 5:19And then we see the Brayton cycle, four
  120. 5:21phases, intake, compression, combustion
  121. 5:24, and expansion with the description.
  122. 5:29Then we can inspect the components.
  123. 5:36Reset the view and we can also remove
  124. 5:38the airflow, remove the labels, and
  125. 5:40anyway, in my opinion, this is an
  126. 5:42excellent example of 3D modeling for
  127. 5:44educational purposes, okay?
  128. 5:48So think, for example, of a teacher who
  129. 5:50has to explain to their students how a
  130. 5:52turbine works.
  131. 5:55With a single prompt, I really expected
  132. 5:57something like 5 minutes, 4 minutes; it
  133. 5:59generated this output for me, already
  134. 6:01useful for conducting a lesson on it to
  135. 6:03explain to the students exactly how the
  136. 6:05turbine works.
  137. 6:08On the gaming side, I had it create this
  138. 6:103D flight simulator over Rome, okay?
  139. 6:14Here we have the controls, flight
  140. 6:15control, environmental visual systems,
  141. 6:17and we can start from the runway or take
  142. 6:19off over Rome. Let's start from the
  143. 6:21runway.
  144. 6:22Wait, let me set it to full screen.
  145. 6:26Here we go. Anyway, the artifact is
  146. 6:28really nice that you can try directly
  147. 6:30here. I'll take part of the runway.
  148. 6:33We can set it to noon, night, sunset, we
  149. 6:35can adjust it ourselves.
  150. 6:38For destinations, we have the Colosseum
  151. 6:40, the Altar of the Fatherland, and so on
  152. 6:42.
  153. 6:43Let's take off now.
  154. 6:47There it is, we are flying over Rome.
  155. 6:50We can see the Tiber here, let's see
  156. 6:52where the Colosseum and some monuments
  157. 6:54are.
  158. 6:55There they are, look how beautiful they
  159. 6:57are.
  160. 6:58Then it tells us we are flying over Rome
  161. 7:00, we have the navigator position.
  162. 7:03There's the Colosseum, wait, let me turn
  163. 7:05, I'll show it to you.
  164. 7:07There it is.
  165. 7:10This one also took about fifteen minutes
  166. 7:12to generate.
  167. 7:16Is a piece of the Colosseum missing? Yes
  168. 7:18, a piece is missing when you look at it
  169. 7:20from inside.
  170. 7:22Then we can basically, let's see what it
  171. 7:24gave me, external, cockpit, we can
  172. 7:26change the view, cockpit.
  173. 7:30It even made the vehicles, let's say,
  174. 7:32that are moving, the traffic, the roads
  175. 7:35, the buildings.
  176. 7:38Destinations.
  177. 7:40Colosseum, there, we can set the
  178. 7:41destinations.
  179. 7:43Saint Peter's.
  180. 7:45We can fly over Saint Peter's.
  181. 7:49And this is a test to create a video
  182. 7:51game very, very quickly.
  183. 7:53Well, it's not perfect, as we can see
  184. 7:55here, I could have the roads improved,
  185. 7:57improve the textures, the quality, here
  186. 7:59this bus sort of flew away. But we are
  187. 8:01talking about a flash model, extremely
  188. 8:03fast.
  189. 8:07So, let's say the optimal task and use
  190. 8:09for this model would be for research,
  191. 8:11for daily work, and I would only use the
  192. 8:13more powerful models for, let's say,
  193. 8:15more complex reasoning.
  194. 8:22So if you want to develop an application
  195. 8:25, a game, I would use Claude to help it
  196. 8:27understand the project and structure the
  197. 8:29specification documents, and then, once
  198. 8:31that's done, for the execution side, I
  199. 8:33would use this model here, which is
  200. 8:35lighter, so as to save a lot of tokens.
  201. 8:42In this test, I want to see how good
  202. 8:44this model is at practically recreating
  203. 8:46the style of a web page.
  204. 8:48This is the Type Safe AI web page.
  205. 8:51See how it has this large font, these
  206. 8:53colors, this animation style?
  207. 8:57And to do that, I have an idea: I'll
  208. 8:59record a video of the web page where I
  209. 9:01scroll from top to bottom and tell it to
  210. 9:03recreate Type Safe AI for me and call it
  211. 9:05Rizzo Safe AI, and see what it comes up
  212. 9:07with and how well it manages to recreate
  213. 9:09the web page faithfully to this style
  214. 9:11here.
  215. 9:16So I recorded this video here where I
  216. 9:17scroll through the landing page,
  217. 9:19recreate the following web page in the
  218. 9:21same way, with the same design, and I'll
  219. 9:23also send you the link and call the web
  220. 9:25page
  221. 9:31Rizzo Safe AI instead of Type Safe AI.
  222. 9:37Hydrogen, standard reasoning.
  223. 9:39Alright, let’s see what it can do.
  224. 9:45And here is the result, Rizzo Safe AI.
  225. 9:48It has the same bar at the top.
  226. 9:51The first public system one model, ok.
  227. 9:55Rizzo Safe AI, yes.
  228. 10:01183 times faster, 496 times cheaper.
  229. 10:06Maybe the animations it generates are
  230. 10:08missing here.
  231. 10:11Same here. Ah no, okay, these are
  232. 10:13interactive, you can play with them
  233. 10:15yourself.
  234. 10:17Okay, let’s scroll down, let’s scroll
  235. 10:19down.
  236. 10:20Rizzo, make the eye the same and the
  237. 10:22cube below the same.
  238. 10:24There, it needs some improvement, but
  239. 10:26this is one-shot, okay?
  240. 10:28I gave it a video prompt and it
  241. 10:30generated a site that's practically
  242. 10:32identical.
  243. 10:33I could have told it to fix two or three
  244. 10:35more things, and it did it extremely
  245. 10:37quickly; it took 3 minutes.
  246. 10:41And it's so fast that it’s been added
  247. 10:43directly to Google Search itself. See
  248. 10:45here, 3.5 Flash versus 4 Hydrogen, it's
  249. 10:47significantly faster; the output quality
  250. 10:50is better.
  251. 10:56Let’s look at images, maps, infographics
  252. 10:59, tables, products, and since it’s so
  253. 11:01fast and smart, I also added it to my
  254. 11:03Academy AI assistant, where this AI
  255. 11:05assistant draws from all the information
  256. 11:07shared in our Academy Discord community
  257. 11:10, Nerd Bar, local AI hardware, all the
  258. 11:12latest news uploaded, not to mention
  259. 11:14educational materials, project ideas,
  260. 11:16job offers; in short, the Academy has
  261. 11:18built this Discord server full of
  262. 11:20information that turns into knowledge,
  263. 11:22which is difficult to search for within
  264. 11:24the Discord server.
  265. 11:32That’s why all Academy members have
  266. 11:34access to this AI agent, which was
  267. 11:35recently updated by switching the model
  268. 11:37to Gemini 4 Hydrogen, and you can ask
  269. 11:39this assistant anything.
  270. 11:45For example, I'll ask it in which live
  271. 11:47lesson did Simone talk about constraint
  272. 11:55decoding models?
  273. 11:57Look how fast it is.
  274. 12:00It’s thinking now, it has access to all
  275. 12:02the tools for semantic search, search
  276. 12:04via agentic file search in the
  277. 12:06relational database, and here it is.
  278. 12:10It already found the lesson, it's lesson
  279. 12:1312 of the Jeff 360 course, you can see
  280. 12:15the lesson directly from here,
  281. 12:21where I explain this new type of model,
  282. 12:23but not only that, in the October 20th
  283. 12:25live stream I presented from minute
  284. 12:2721:23, I even show you which part of the
  285. 12:30video I talk about it in.
  286. 12:34You can see all the slides,
  287. 12:38and it really is a very, very
  288. 12:39high-performing model and fast, as you
  289. 12:41could see.
  290. 12:42Let’s ask it this other question. What
  291. 12:44computer do you recommend for running AI
  292. 12:46models locally for coding?
  293. 12:50Now it will draw from all the
  294. 12:51information that has been discussed
  295. 12:53within the academy, because there are
  296. 12:54many channels where users have shared
  297. 12:56their experience with their hardware,
  298. 12:58and that information is then cited.
  299. 13:02You see here, they are cited, my YouTube
  300. 13:04video is also cited, all the information
  301. 13:06is cited, so it is a model that
  302. 13:08practically leverages all the
  303. 13:09information that is inside the academy.
  304. 13:13Look, personal tier, pro tier, for
  305. 13:16example, Carlos Ross gave his answer
  306. 13:18regarding the 24GB M5 Pro.
  307. 13:21There is a discussion here, it gives the
  308. 13:23link, you open it on Discord and take a
  309. 13:25look at the various responses.
  310. 13:31There is also a help and questions forum
  311. 13:33where anyone can create a post and we
  312. 13:35respond below, workshops, and so as a
  313. 13:37model for the academy I am using this
  314. 13:39one now, I am finding it really, really
  315. 13:41good, and not only that, but the MCP
  316. 13:43server functionality has also been added
  317. 13:45, where you can connect your Cloud Code
  318. 13:47Agent and practically ask it anything
  319. 13:49about all the know-how and knowledge we
  320. 13:51are developing inside the academy.
  321. 13:59And the great thing is that you can ask
  322. 14:01it anything, like for example searching
  323. 14:03within the lessons for you or having it
  324. 14:05create a personalized training path, you
  325. 14:07can ask it what the latest news is on
  326. 14:08certain topics, you can ask it, for
  327. 14:10example, to search for people within the
  328. 14:12academy in a specific sector.
  329. 14:15Searching for a front-end, searching for
  330. 14:17a back-end, searching for job offers.
  331. 14:20It is truly a platform that is having a
  332. 14:22lot of success within the community.
  333. 14:26And so I wanted to show you a real use
  334. 14:27case where I put this model into
  335. 14:29production, where I am testing it,
  336. 14:31precisely, with real users who use this
  337. 14:33platform every day.
  338. 14:35And if you are also interested in being
  339. 14:37part of the academy, in the description
  340. 14:39you will find the link to the landing
  341. 14:41page rizzoaiacademy.com. It is an annual
  342. 14:43subscription that includes all courses,
  343. 14:45over 14 to date, with all weekly live
  344. 14:47sessions, typically on Tuesday evenings
  345. 14:49at 9:00 PM.
  346. 14:51Returning to the model, the nice thing
  347. 14:53is that it is already available, you can
  348. 14:55already use it for free as well, by
  349. 14:57using it directly from Google AI Studio
  350. 14:59or directly from Gemini.
  351. 15:04And once tested, it actually reflects
  352. 15:06more or less what was said in the
  353. 15:08benchmarks, it is faster and even
  354. 15:10superior to Sonnet 5.5 and GPT 6.1 Terra
  355. 15:12, even if this Gemini 4 Hydrogen
  356. 15:14represents the least intelligent in the
  357. 15:16Gemini model family, it would be like
  358. 15:18the Haiku of the Claude family, but it
  359. 15:20rivals Sonnet, so they have taken, in my
  360. 15:23opinion, an extra step, so while
  361. 15:24maintaining the lightning-fast Flash
  362. 15:27model, they have greatly increased the
  363. 15:29performance.
  364. 15:34Now let's look at the decision APIs,
  365. 15:36which surprised me a bit because when
  366. 15:38JeV came out—those System One models—
  367. 15:40everyone was skeptical, asking what
  368. 15:42these models even were, what they're for
  369. 15:44, and if they're useful.
  370. 15:46I, on the other hand, pushed them from
  371. 15:48the start, saying this would be a new
  372. 15:50family of models that everyone would
  373. 15:51start producing.
  374. 15:56Then OpenAI started, Perplexity followed
  375. 15:58, Cloudflare made theirs, many versions
  376. 16:00were released, and we in the academy
  377. 16:03even made our own version, Rizzo Flow,
  378. 16:05which has almost 1,000 stars on GitHub
  379. 16:07and is a System One model based on two
  380. 16:09decoder models of 1.7 billion and 4
  381. 16:11billion parameters, which I trained and
  382. 16:14fine-tuned to have calibrated
  383. 16:15probabilities, just like JeV did with
  384. 16:18his model.
  385. 16:22But if you want to dig deeper into this
  386. 16:24, there is a dedicated video where I
  387. 16:25explain how I built the model, its
  388. 16:27performance, how I trained it—I explain
  389. 16:29everything in detail.
  390. 16:32Back to Gemini, the model they created
  391. 16:35is completely multimodal, so it receives
  392. 16:38text, JSON, images, audio, video, and
  393. 16:40PDFs as input.
  394. 16:44We can ask our usual questions, so it
  395. 16:46takes the state and the question, and
  396. 16:48the model responds with the three
  397. 16:49primitives: boolean, choice, and score.
  398. 16:54I think other properties and primitives
  399. 16:56will be added, like "box," which was
  400. 16:58added to the latest JeV version because
  401. 17:00it handles images and gives you the
  402. 17:02bounding box of objects, which is very
  403. 17:04useful.
  404. 17:07Looking at the performance, according to
  405. 17:10them, it beats TypeSafe JeV 1.14, the
  406. 17:12latest one, and it outperforms OpenAI's
  407. 17:15GPT-6 Luna Decisions and Gemini 3.5
  408. 17:17Flashlight.
  409. 17:22And thanks to its multimodal nature,
  410. 17:24they show right away in this demo that
  411. 17:26it is practically perfect for robotics.
  412. 17:29Why?
  413. 17:37Because by giving it images, audio, text
  414. 17:39, and JSON as input, and given that it's
  415. 17:41super fast at making decisions, you need
  416. 17:43that real-time capability in robotics—
  417. 17:45you need to make decisions in real time
  418. 17:48—so it's perfect for many use cases,
  419. 17:50such as understanding what's happening
  420. 17:52on the fly, working with huge amounts of
  421. 17:54data, or for automations and web
  422. 17:56scraping, or for document systems if you
  423. 17:58need to filter many resumes, documents,
  424. 18:00and ads.
  425. 18:09So I'm pleased to see another player
  426. 18:11investing in decision models by
  427. 18:13releasing their own, which costs 10
  428. 18:15cents per million input tokens and, of
  429. 18:17course, zero for output tokens since it
  430. 18:19doesn't generate output tokens, it's all
  431. 18:21input.
  432. 18:24And Artificial Analysis recently added a
  433. 18:26new leaderboard called Decision Index
  434. 18:28version 2.1, where we can see the latest
  435. 18:31models.
  436. 18:36We see Gemini Decision 1 tied with Jeff
  437. 18:391.14, Elephant 9B, Liquid AI's D1, D1
  438. 18:42Ultra, and Glide V 1.2; so many are
  439. 18:44coming out, like Cloudflare's multimodal
  440. 18:47Clef with different primitives—some do
  441. 18:50forecasting, and some even extract text
  442. 18:52from documents.
  443. 18:57It’s absolutely insane what is happening
  444. 19:00with Decision Models.
  445. 19:02In conclusion, Google, in my opinion, is
  446. 19:04flourishing again, releasing models with
  447. 19:06great frequency. Carbon is like a
  448. 19:08top-three best model, the smartest one,
  449. 19:11and Hydrogen is the fastest—truly the
  450. 19:13fastest one out there. In fact, I’ve put
  451. 19:15it into production inside my academy.
  452. 19:17I’m finding it works very well; we even
  453. 19:19experimented by having it build small
  454. 19:21applications, web pages, simulations,
  455. 19:23and video games. They also created the
  456. 19:25Decision API, so they are investing in
  457. 19:27Decision Models, and I’m very happy
  458. 19:29about that because the more competition
  459. 19:31there is, the more we all benefit. I
  460. 19:33hope you enjoyed this video; write a
  461. 19:35comment below with your thoughts, and as
  462. 19:36always, follow me to stay updated on the
  463. 19:38world of AI.

About this transcript

This page contains the full transcript of Gemini 4 Hydrogen è davvero così potente? (testato personalmente) by Simone Rizzo, generated from the public captions YouTube serves with the video. The transcript has 2,854 words across 463 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.