Gemini 4 Hydrogen è davvero così potente? (testato personalmente) — Transcript
Full transcript
- 0:02Google is finally more active and has
- 0:04recently released Gemini 4 Hydrogen, a
- 0:06new multimodal model named after an
- 0:08element of the periodic table, so we
- 0:10have Argon, Carbon which is the smartest
- 0:12, and now Hydrogen.
- 0:18This model variant is designed for daily
- 0:20use, replacing the previous 3.8 Flash,
- 0:23and the interesting thing is that it is
- 0:25fast, intelligent, and relatively
- 0:27inexpensive, at $ 1.50 per million input
- 0:30tokens and $ 7.50 per million output
- 0:32tokens.
- 0:35And in today's video, I want to test
- 0:37this model together with you.
- 0:42Let’s look at all its features, and
- 0:44there is also news from Gemini; they
- 0:46released the Decisions API, and so they
- 0:48have also released a Decision Model,
- 0:50Jev-style, which is very interesting to
- 0:52see because, as you know, I have pushed
- 0:54hard on this new family of models and
- 0:56every day there is a new confirmation,
- 0:58as these frontier labs have effectively
- 0:59released their own Decision Model.
- 1:07Returning to Hydrogen, Google did
- 1:08something interesting: the orchestrator
- 1:10in Google Antigravity, where they say
- 1:12instead of using a single model, they
- 1:14are orchestrated effectively and
- 1:16efficiently.
- 1:21Specifically, this Hydrogen is the
- 1:23fastest for everyday work, which is then
- 1:25put to work in collaboration with Carbon
- 1:27and Argon for High Efficiency Execution
- 1:29Engines or Multimodal Workflows.
- 1:37Based on benchmarks, this model is
- 1:39compared with Claude Sonnet 5.5 and GPT
- 1:426.1 Terra; so in terms of intelligence,
- 1:44it compares to these two, although by
- 1:46benchmarks it seems to outperform them,
- 1:49and according to the independent
- 1:51Artificial Analysis benchmark, it is
- 1:53positioned between GPT 6.1 Terra and
- 1:56Claude Sonnet 5.5 Max.
- 2:01With 54 intelligence points, it is
- 2:03therefore superior to GPT 6.1 Terra and
- 2:06slightly inferior to Sonnet 5.5.
- 2:10We are now going to test this model in a
- 2:12practical field.
- 2:15It is a very capable model in coding,
- 2:17able to generate 3D objects, perfect for
- 2:19gaming, and also capable of creating
- 2:21landing pages and web pages with
- 2:23excellent UI/UX and animations.
- 2:28They even do a stylistic exercise here
- 2:30where, in Google Sheets, which is
- 2:31basically Excel, they have it paint
- 2:33images and create animations.
- 2:39Then we have its full capacity to create
- 2:41presentations and slides, or to work on
- 2:43financial reports, spreadsheets, and
- 2:45conduct research, so everything that is
- 2:47practically office work.
- 2:53Comparing Gemini 4 Hydrogen and Gemini 4
- 2:55Carbon, therefore the fastest versus the
- 2:58smartest, we really see the difference
- 3:01in cost and time, where above we have
- 3:03the example: Hydrogen cost $ 4.27 in 1
- 3:06hour and 23 minutes, while Carbon cost $
- 3:098.83 in 2 hours and 39 minutes.
- 3:15Looking at the output results, they are
- 3:17more or less the same, very comparable,
- 3:19so in a way, for us consumers, it's
- 3:21better to use Gemini 4 Hydrogen for most
- 3:23tasks because it saves us time and money
- 3:26.
- 3:29This pattern is also shown by these
- 3:31charts, where Gemini 4 Hydrogen is
- 3:34practically compared with Sonnet 5.5 and
- 3:36GPT 6.1 Terra. We can see here below the
- 3:39time taken per task, how many minutes it
- 3:41takes to complete the task, and also
- 3:43what score it has, so how intelligent it
- 3:45is.
- 3:47We see that 4 Hydrogen is at the top
- 3:49left, so it is more intelligent,
- 3:51especially with reasoning, and it takes
- 3:53much less time: it takes between 6
- 3:55minutes here, while here it's between 10
- 3:57and 15.
- 4:02While we see how Sonnet even goes above
- 4:0430 and 15, whereas Terra takes 7 and 15
- 4:07minutes, so there is a difference of a
- 4:09few minutes compared to the 6.1 Terra
- 4:11and half an hour with Sonnet 5.5.
- 4:18The first thing I had it do was create a
- 4:203D simulation for my tests using
- 4:22Hydrogen with medium reasoning, a 3D
- 4:24visualization of how a turbine works.
- 4:29So let's open it now; this was all done
- 4:32in one shot, a single shot.
- 4:35We can basically see the entire 3D
- 4:37visualization of a turbine, so this is
- 4:39what it looks like from the outside when
- 4:41you see an airplane turbine, and this is
- 4:43the internal part.
- 4:50We can increase the throttle here, it's
- 4:52written there, we can change the axial
- 4:53exploded view, we can click on every
- 4:55component and read the description, so
- 4:57we have the fan, the compressor, the
- 4:59other compressor and combustor, turbine
- 5:01, and the other turbine.
- 5:08Then we have various view modes, X-ray,
- 5:11thermal.
- 5:13Oh, wow, all the heat here is in the
- 5:15combustor.
- 5:19And then we see the Brayton cycle, four
- 5:21phases, intake, compression, combustion
- 5:24, and expansion with the description.
- 5:29Then we can inspect the components.
- 5:36Reset the view and we can also remove
- 5:38the airflow, remove the labels, and
- 5:40anyway, in my opinion, this is an
- 5:42excellent example of 3D modeling for
- 5:44educational purposes, okay?
- 5:48So think, for example, of a teacher who
- 5:50has to explain to their students how a
- 5:52turbine works.
- 5:55With a single prompt, I really expected
- 5:57something like 5 minutes, 4 minutes; it
- 5:59generated this output for me, already
- 6:01useful for conducting a lesson on it to
- 6:03explain to the students exactly how the
- 6:05turbine works.
- 6:08On the gaming side, I had it create this
- 6:103D flight simulator over Rome, okay?
- 6:14Here we have the controls, flight
- 6:15control, environmental visual systems,
- 6:17and we can start from the runway or take
- 6:19off over Rome. Let's start from the
- 6:21runway.
- 6:22Wait, let me set it to full screen.
- 6:26Here we go. Anyway, the artifact is
- 6:28really nice that you can try directly
- 6:30here. I'll take part of the runway.
- 6:33We can set it to noon, night, sunset, we
- 6:35can adjust it ourselves.
- 6:38For destinations, we have the Colosseum
- 6:40, the Altar of the Fatherland, and so on
- 6:42.
- 6:43Let's take off now.
- 6:47There it is, we are flying over Rome.
- 6:50We can see the Tiber here, let's see
- 6:52where the Colosseum and some monuments
- 6:54are.
- 6:55There they are, look how beautiful they
- 6:57are.
- 6:58Then it tells us we are flying over Rome
- 7:00, we have the navigator position.
- 7:03There's the Colosseum, wait, let me turn
- 7:05, I'll show it to you.
- 7:07There it is.
- 7:10This one also took about fifteen minutes
- 7:12to generate.
- 7:16Is a piece of the Colosseum missing? Yes
- 7:18, a piece is missing when you look at it
- 7:20from inside.
- 7:22Then we can basically, let's see what it
- 7:24gave me, external, cockpit, we can
- 7:26change the view, cockpit.
- 7:30It even made the vehicles, let's say,
- 7:32that are moving, the traffic, the roads
- 7:35, the buildings.
- 7:38Destinations.
- 7:40Colosseum, there, we can set the
- 7:41destinations.
- 7:43Saint Peter's.
- 7:45We can fly over Saint Peter's.
- 7:49And this is a test to create a video
- 7:51game very, very quickly.
- 7:53Well, it's not perfect, as we can see
- 7:55here, I could have the roads improved,
- 7:57improve the textures, the quality, here
- 7:59this bus sort of flew away. But we are
- 8:01talking about a flash model, extremely
- 8:03fast.
- 8:07So, let's say the optimal task and use
- 8:09for this model would be for research,
- 8:11for daily work, and I would only use the
- 8:13more powerful models for, let's say,
- 8:15more complex reasoning.
- 8:22So if you want to develop an application
- 8:25, a game, I would use Claude to help it
- 8:27understand the project and structure the
- 8:29specification documents, and then, once
- 8:31that's done, for the execution side, I
- 8:33would use this model here, which is
- 8:35lighter, so as to save a lot of tokens.
- 8:42In this test, I want to see how good
- 8:44this model is at practically recreating
- 8:46the style of a web page.
- 8:48This is the Type Safe AI web page.
- 8:51See how it has this large font, these
- 8:53colors, this animation style?
- 8:57And to do that, I have an idea: I'll
- 8:59record a video of the web page where I
- 9:01scroll from top to bottom and tell it to
- 9:03recreate Type Safe AI for me and call it
- 9:05Rizzo Safe AI, and see what it comes up
- 9:07with and how well it manages to recreate
- 9:09the web page faithfully to this style
- 9:11here.
- 9:16So I recorded this video here where I
- 9:17scroll through the landing page,
- 9:19recreate the following web page in the
- 9:21same way, with the same design, and I'll
- 9:23also send you the link and call the web
- 9:25page
- 9:31Rizzo Safe AI instead of Type Safe AI.
- 9:37Hydrogen, standard reasoning.
- 9:39Alright, let’s see what it can do.
- 9:45And here is the result, Rizzo Safe AI.
- 9:48It has the same bar at the top.
- 9:51The first public system one model, ok.
- 9:55Rizzo Safe AI, yes.
- 10:01183 times faster, 496 times cheaper.
- 10:06Maybe the animations it generates are
- 10:08missing here.
- 10:11Same here. Ah no, okay, these are
- 10:13interactive, you can play with them
- 10:15yourself.
- 10:17Okay, let’s scroll down, let’s scroll
- 10:19down.
- 10:20Rizzo, make the eye the same and the
- 10:22cube below the same.
- 10:24There, it needs some improvement, but
- 10:26this is one-shot, okay?
- 10:28I gave it a video prompt and it
- 10:30generated a site that's practically
- 10:32identical.
- 10:33I could have told it to fix two or three
- 10:35more things, and it did it extremely
- 10:37quickly; it took 3 minutes.
- 10:41And it's so fast that it’s been added
- 10:43directly to Google Search itself. See
- 10:45here, 3.5 Flash versus 4 Hydrogen, it's
- 10:47significantly faster; the output quality
- 10:50is better.
- 10:56Let’s look at images, maps, infographics
- 10:59, tables, products, and since it’s so
- 11:01fast and smart, I also added it to my
- 11:03Academy AI assistant, where this AI
- 11:05assistant draws from all the information
- 11:07shared in our Academy Discord community
- 11:10, Nerd Bar, local AI hardware, all the
- 11:12latest news uploaded, not to mention
- 11:14educational materials, project ideas,
- 11:16job offers; in short, the Academy has
- 11:18built this Discord server full of
- 11:20information that turns into knowledge,
- 11:22which is difficult to search for within
- 11:24the Discord server.
- 11:32That’s why all Academy members have
- 11:34access to this AI agent, which was
- 11:35recently updated by switching the model
- 11:37to Gemini 4 Hydrogen, and you can ask
- 11:39this assistant anything.
- 11:45For example, I'll ask it in which live
- 11:47lesson did Simone talk about constraint
- 11:55decoding models?
- 11:57Look how fast it is.
- 12:00It’s thinking now, it has access to all
- 12:02the tools for semantic search, search
- 12:04via agentic file search in the
- 12:06relational database, and here it is.
- 12:10It already found the lesson, it's lesson
- 12:1312 of the Jeff 360 course, you can see
- 12:15the lesson directly from here,
- 12:21where I explain this new type of model,
- 12:23but not only that, in the October 20th
- 12:25live stream I presented from minute
- 12:2721:23, I even show you which part of the
- 12:30video I talk about it in.
- 12:34You can see all the slides,
- 12:38and it really is a very, very
- 12:39high-performing model and fast, as you
- 12:41could see.
- 12:42Let’s ask it this other question. What
- 12:44computer do you recommend for running AI
- 12:46models locally for coding?
- 12:50Now it will draw from all the
- 12:51information that has been discussed
- 12:53within the academy, because there are
- 12:54many channels where users have shared
- 12:56their experience with their hardware,
- 12:58and that information is then cited.
- 13:02You see here, they are cited, my YouTube
- 13:04video is also cited, all the information
- 13:06is cited, so it is a model that
- 13:08practically leverages all the
- 13:09information that is inside the academy.
- 13:13Look, personal tier, pro tier, for
- 13:16example, Carlos Ross gave his answer
- 13:18regarding the 24GB M5 Pro.
- 13:21There is a discussion here, it gives the
- 13:23link, you open it on Discord and take a
- 13:25look at the various responses.
- 13:31There is also a help and questions forum
- 13:33where anyone can create a post and we
- 13:35respond below, workshops, and so as a
- 13:37model for the academy I am using this
- 13:39one now, I am finding it really, really
- 13:41good, and not only that, but the MCP
- 13:43server functionality has also been added
- 13:45, where you can connect your Cloud Code
- 13:47Agent and practically ask it anything
- 13:49about all the know-how and knowledge we
- 13:51are developing inside the academy.
- 13:59And the great thing is that you can ask
- 14:01it anything, like for example searching
- 14:03within the lessons for you or having it
- 14:05create a personalized training path, you
- 14:07can ask it what the latest news is on
- 14:08certain topics, you can ask it, for
- 14:10example, to search for people within the
- 14:12academy in a specific sector.
- 14:15Searching for a front-end, searching for
- 14:17a back-end, searching for job offers.
- 14:20It is truly a platform that is having a
- 14:22lot of success within the community.
- 14:26And so I wanted to show you a real use
- 14:27case where I put this model into
- 14:29production, where I am testing it,
- 14:31precisely, with real users who use this
- 14:33platform every day.
- 14:35And if you are also interested in being
- 14:37part of the academy, in the description
- 14:39you will find the link to the landing
- 14:41page rizzoaiacademy.com. It is an annual
- 14:43subscription that includes all courses,
- 14:45over 14 to date, with all weekly live
- 14:47sessions, typically on Tuesday evenings
- 14:49at 9:00 PM.
- 14:51Returning to the model, the nice thing
- 14:53is that it is already available, you can
- 14:55already use it for free as well, by
- 14:57using it directly from Google AI Studio
- 14:59or directly from Gemini.
- 15:04And once tested, it actually reflects
- 15:06more or less what was said in the
- 15:08benchmarks, it is faster and even
- 15:10superior to Sonnet 5.5 and GPT 6.1 Terra
- 15:12, even if this Gemini 4 Hydrogen
- 15:14represents the least intelligent in the
- 15:16Gemini model family, it would be like
- 15:18the Haiku of the Claude family, but it
- 15:20rivals Sonnet, so they have taken, in my
- 15:23opinion, an extra step, so while
- 15:24maintaining the lightning-fast Flash
- 15:27model, they have greatly increased the
- 15:29performance.
- 15:34Now let's look at the decision APIs,
- 15:36which surprised me a bit because when
- 15:38JeV came out—those System One models—
- 15:40everyone was skeptical, asking what
- 15:42these models even were, what they're for
- 15:44, and if they're useful.
- 15:46I, on the other hand, pushed them from
- 15:48the start, saying this would be a new
- 15:50family of models that everyone would
- 15:51start producing.
- 15:56Then OpenAI started, Perplexity followed
- 15:58, Cloudflare made theirs, many versions
- 16:00were released, and we in the academy
- 16:03even made our own version, Rizzo Flow,
- 16:05which has almost 1,000 stars on GitHub
- 16:07and is a System One model based on two
- 16:09decoder models of 1.7 billion and 4
- 16:11billion parameters, which I trained and
- 16:14fine-tuned to have calibrated
- 16:15probabilities, just like JeV did with
- 16:18his model.
- 16:22But if you want to dig deeper into this
- 16:24, there is a dedicated video where I
- 16:25explain how I built the model, its
- 16:27performance, how I trained it—I explain
- 16:29everything in detail.
- 16:32Back to Gemini, the model they created
- 16:35is completely multimodal, so it receives
- 16:38text, JSON, images, audio, video, and
- 16:40PDFs as input.
- 16:44We can ask our usual questions, so it
- 16:46takes the state and the question, and
- 16:48the model responds with the three
- 16:49primitives: boolean, choice, and score.
- 16:54I think other properties and primitives
- 16:56will be added, like "box," which was
- 16:58added to the latest JeV version because
- 17:00it handles images and gives you the
- 17:02bounding box of objects, which is very
- 17:04useful.
- 17:07Looking at the performance, according to
- 17:10them, it beats TypeSafe JeV 1.14, the
- 17:12latest one, and it outperforms OpenAI's
- 17:15GPT-6 Luna Decisions and Gemini 3.5
- 17:17Flashlight.
- 17:22And thanks to its multimodal nature,
- 17:24they show right away in this demo that
- 17:26it is practically perfect for robotics.
- 17:29Why?
- 17:37Because by giving it images, audio, text
- 17:39, and JSON as input, and given that it's
- 17:41super fast at making decisions, you need
- 17:43that real-time capability in robotics—
- 17:45you need to make decisions in real time
- 17:48—so it's perfect for many use cases,
- 17:50such as understanding what's happening
- 17:52on the fly, working with huge amounts of
- 17:54data, or for automations and web
- 17:56scraping, or for document systems if you
- 17:58need to filter many resumes, documents,
- 18:00and ads.
- 18:09So I'm pleased to see another player
- 18:11investing in decision models by
- 18:13releasing their own, which costs 10
- 18:15cents per million input tokens and, of
- 18:17course, zero for output tokens since it
- 18:19doesn't generate output tokens, it's all
- 18:21input.
- 18:24And Artificial Analysis recently added a
- 18:26new leaderboard called Decision Index
- 18:28version 2.1, where we can see the latest
- 18:31models.
- 18:36We see Gemini Decision 1 tied with Jeff
- 18:391.14, Elephant 9B, Liquid AI's D1, D1
- 18:42Ultra, and Glide V 1.2; so many are
- 18:44coming out, like Cloudflare's multimodal
- 18:47Clef with different primitives—some do
- 18:50forecasting, and some even extract text
- 18:52from documents.
- 18:57It’s absolutely insane what is happening
- 19:00with Decision Models.
- 19:02In conclusion, Google, in my opinion, is
- 19:04flourishing again, releasing models with
- 19:06great frequency. Carbon is like a
- 19:08top-three best model, the smartest one,
- 19:11and Hydrogen is the fastest—truly the
- 19:13fastest one out there. In fact, I’ve put
- 19:15it into production inside my academy.
- 19:17I’m finding it works very well; we even
- 19:19experimented by having it build small
- 19:21applications, web pages, simulations,
- 19:23and video games. They also created the
- 19:25Decision API, so they are investing in
- 19:27Decision Models, and I’m very happy
- 19:29about that because the more competition
- 19:31there is, the more we all benefit. I
- 19:33hope you enjoyed this video; write a
- 19:35comment below with your thoughts, and as
- 19:36always, follow me to stay updated on the
- 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.