Deepseek a résolu le plus gros problème de l’IA — Transcript
Full transcript
- 0:00For years, the AIs that everyone uses
- 0:03have had the same problem. They cost
- 0:06billions to develop and are locked down
- 0:08by a handful of American companies. But
- 0:11recently, a Chinese AI has solved this
- 0:13problem and no one saw it coming. You
- 0:16may have already heard of it because
- 0:18over a year ago, when it was first
- 0:20released, it caused the biggest loss in
- 0:23stock market history in a single day.
- 0:26Almost $ 600 billion went up in smoke,
- 0:29and it has just repeated the feat with
- 0:31its latest artificial intelligence
- 0:34model. Tatia, her name is Dipsic, and
- 0:37she has become almost as powerful as
- 0:39the best GPT and cloud cat models,
- 0:41except that she is open source and free
- 0:43. And all this power that was once
- 0:47owned by a few American companies, can
- 0:49now be used by anyone. I will explain
- 0:52to you in 15 minutes how they did it,
- 0:54what it changes for you and how to take
- 0:56advantage of it. Now, I know what
- 0:58you're thinking. You're thinking that
- 1:01something this powerful, free, and
- 1:02Chinese, well, it smells like a trap.
- 1:04And it's true, there's one not
- 1:06necessarily where you expect it to be,
- 1:07and I'll show it to you at the end of
- 1:09this video. Before explaining how this
- 1:12AI is revolutionizing the market, let's
- 1:14see what it looks like in concrete
- 1:16terms. Here's the Deepsy interface. In
- 1:18fact, you create an account and you'll
- 1:20immediately land on a very simple
- 1:21interface that you're probably already
- 1:23familiar with because it looks a lot
- 1:24like GPT chat or Cloud. And to use it,
- 1:27it's exactly the same. You just have to
- 1:29ask it what you want. You have three
- 1:31main distinct modes: fast, expert, and
- 1:33vision. And as its name suggests, fast
- 1:36is there to help you get an answer as
- 1:38quickly as possible. Then, expert mode
- 1:41is for the most complex tasks, and
- 1:42finally vision mode is for providing it
- 1:44with images or documents so that it can
- 1:46analyze them in detail. And for the
- 1:49example, we're even going to ask him to
- 1:50create a little 3D platform game for us
- 1:52to push him to his limits and his
- 1:54thinking. We'll come back to that a
- 1:56little later when he's finished. But
- 1:57the important thing to note that you
- 1:59may have noticed on this interface is
- 2:00that you can see that he writes very,
- 2:02very quickly, much faster than Claude
- 2:04or Chat GPT and that will be very
- 2:05important in the rest of the video.
- 2:08Okay, great, the game works, nothing
- 2:10revolutionary but it was just to give
- 2:11you a little visual demo. Now, having
- 2:15this result in a single message, a
- 2:16single prompt that I sent him, is still
- 2:18impressive. Of course, if you want to
- 2:21go even further, you can simply ask him
- 2:23by refining and making changes with him
- 2:24. But the really interesting detail
- 2:26about the interface is elsewhere. On
- 2:28Chat, GPT or Cloud, you probably know
- 2:30this button that systematically pushes
- 2:33you to the paid version. Sometimes,
- 2:35it's around the € 20 version and
- 2:36sometimes even the € 200 per month
- 2:38version. And of course, with Dipsic,
- 2:40you can search for this button as long
- 2:42as you want. You won't find it. And
- 2:44it's very simple, it's because it
- 2:45doesn't exist. They don't even offer
- 2:47you to pay to use this tool. Dipsic is
- 2:50free, completely free. So naturally, I
- 2:53wanted to test their brand new AI model
- 2:55for you. I've been using it every day
- 2:58for over a week now instead of all my
- 3:00usual tools. In this video, we will
- 3:02look at four points together. First, in
- 3:04one, we're going to understand where
- 3:05this AI comes from and how a small
- 3:07Chinese lab managed to do something
- 3:09that no one had done before. In two,
- 3:11we're going to see what it's really
- 3:12capable of doing, but also its
- 3:14limitations. In three, we're going to
- 3:16see how you can use it, including for
- 3:17tasks you don't even suspect. And
- 3:19finally, in four, we'll see if all this
- 3:21is risk-free because the Chinese
- 3:22company is free, which inevitably
- 3:24raises questions. So, to understand why
- 3:27Dipsic is making so much noise, we
- 3:29first need to know where it comes from.
- 3:31Because Dipsic is far from being a tech
- 3:34giant. It's simply a small Chinese
- 3:37laboratory that originally wasn't even
- 3:39a tech company but an investment fund.
- 3:41So it's a small team with far fewer
- 3:43people and, above all, far less money
- 3:45than the American giants. They have a
- 3:47huge handicap because due to American
- 3:49restrictions on China, they don't even
- 3:51have access to the most powerful chips,
- 3:54the ones that run all the American
- 3:55chips. So on paper, he was clearly the
- 3:58underdog. Except that it's often
- 4:00precisely this kind of handicap that
- 4:02makes a team formidable. Because when
- 4:04you have everything, well, you often
- 4:05fall into comfort. When you have
- 4:07nothing, you excel. And that's exactly
- 4:09what happened because a little over a
- 4:11year ago, they released a model that
- 4:13rivals the best American I-series for a
- 4:15fraction of the price. The numbers are
- 4:17a bit of a debate today, but we're
- 4:19talking about at least 20 times less
- 4:20than what it cost to train Chat GPT.
- 4:23And that's what set the American stock
- 4:25market on fire because overnight,
- 4:26everyone asked themselves the same
- 4:28question. If they can do it with next
- 4:31to nothing, why do others have to spend
- 4:32billions? And to give you a fairly
- 4:35simple example that explains the
- 4:36reaction of the stock markets, imagine
- 4:38that you're building your house. You
- 4:41put a lot of money into it to have the
- 4:42most beautiful house in the
- 4:44neighborhood and, the next day, you
- 4:45realize that the neighbor right next
- 4:47door built exactly the same house,
- 4:49except that he paid 10 times less. So,
- 4:53not only would I think you would be
- 4:54shocked, but you would also be very
- 4:56angry with the construction company
- 4:57that overcharged you for the same
- 4:59result. Well, that's exactly how Wall
- 5:03Street felt about Silicon Valley
- 5:05because the Americans like to do one
- 5:07thing to go faster: add more and more
- 5:09money, more machines, more power to
- 5:10crush the competition. Dipsic simply
- 5:14couldn't do that. So instead, they just
- 5:17got smarter. In fact, to help you
- 5:19understand a bit better, when you ask a
- 5:22classic AI a question, it will use all
- 5:24its power, everything it is capable of
- 5:26doing, even for a very simple question.
- 5:29Imagine for 2 seconds, you run a
- 5:30company with thousands of employees and
- 5:32you have a small, specific accounting
- 5:34question. Are you really going to call
- 5:37everyone into a meeting? Well, of
- 5:38course not. You go straight to the
- 5:40accountant and the others continue to
- 5:42work quietly on their own. Well,
- 5:44Deepsic, that's exactly what they've
- 5:45set up. They have a huge model, but for
- 5:48each question, it only wakes up the few
- 5:50specialists it really needs. All the
- 5:52others remain asleep. It has the power
- 5:54of a large model and the speed of a
- 5:56small one. So, of course, it responds
- 5:58much faster, and that's important. Now,
- 6:01this idea clearly wasn't invented by
- 6:03them. It existed long before. Except
- 6:04that before them, no one had pushed it
- 6:06that far. And above all, they've done
- 6:08something that almost no one else does
- 6:10in this shark environment. They share
- 6:12almost everything. Well, not absolutely
- 6:14everything, it's important to be
- 6:15precise. The exact recipe for training
- 6:17the model and the data they use, well,
- 6:19they keep that to themselves. But for
- 6:21everything else, they show a generosity
- 6:23rare in this environment. First, they
- 6:25make the model itself freely available,
- 6:27meaning that anyone can download it.
- 6:30And most importantly, they publish a
- 6:31huge number of papers, dozens of
- 6:33research papers in which they explain
- 6:34their findings and their tricks,
- 6:36including how to make the machine run
- 6:37much more cheaply. And that's typically
- 6:41the kind of thing that American
- 6:42laboratories keep carefully hidden,
- 6:44literally like industrial secrets. So
- 6:47now, to properly measure the impact of
- 6:49this approach on the market, let's
- 6:51imagine something. Let's take a farmer
- 6:53who finds a method to produce twice as
- 6:55much wheat with half the water.
- 6:57Normally, what does he do? Well, he
- 6:58keeps his method well hidden to himself
- 7:00. He crushes his competition and sells
- 7:02his wheat at the price he wants. Well,
- 7:04in Psic, they do the exact opposite.
- 7:06They publicly explain how they did it
- 7:08for free so that even their opponents
- 7:10can draw inspiration from it. And that
- 7:13triggers two phenomena. The first is
- 7:15competition. Dipsic sells a nia at a
- 7:17fraction of the American price. As a
- 7:20result, all the others are forced to
- 7:21fall into line or explain to everyone
- 7:23why they are 10 times more expensive.
- 7:25And the second one is even more
- 7:27interesting. Since almost everything is
- 7:30published, other AI labs can read,
- 7:31understand, and apply their findings to
- 7:33lower their own costs. In other words,
- 7:36Dipsic doesn't just undercut the prices
- 7:38of the biggest AI suppliers, it also
- 7:41gives other labs the means to undercut
- 7:43them. And you'll see why it snowballs.
- 7:45Because since they did that, other
- 7:47giants, particularly in China, have
- 7:49started to open up their models in turn
- 7:50. And today, Chinese nias cost 4 to 10
- 7:53times less than their American
- 7:55equivalents thanks to these famous
- 7:57Dipsic findings. So now, we might begin
- 7:59to wonder exactly why they are doing
- 8:01all of this? And frankly, it's pretty
- 8:03smart of them. When you're the
- 8:05challenger in this market, making
- 8:07technology free for everyone is the
- 8:08best way to destroy the lead of those
- 8:10who are ahead. And by the way, even you
- 8:13behind this screen in your daily life
- 8:15can see the results of this approach.
- 8:16This pressure that they have managed to
- 8:18create, well, it even affects the AI
- 8:19that you use every day. And if we go
- 8:22back to our example of the farmer, the
- 8:23price of the infinite bit has dropped
- 8:25for everyone, even for those who have
- 8:27never heard of our farmer and his new
- 8:28method. So Lia cost a fortune to create
- 8:32and run. It was locked to a few people
- 8:35and Dipsic has just solved this problem
- 8:37not just for them but for an entire
- 8:40industry. And of course, when something
- 8:42costs next to nothing, we can afford to
- 8:44give it away for free and that's where
- 8:46you really get into the story. Okay, a
- 8:49free AI, but does it really hold its
- 8:52own against GPT chat and Claud? Because
- 8:54free AI that talks absolute nonsense is
- 8:56useless and nobody even cares about it.
- 8:59So, as we saw in the demo, it is
- 9:00capable of doing quite complex things.
- 9:03Then, if we look at the comparisons on
- 9:05some of the hardest thinking tests to
- 9:07update the different model levels, well
- 9:09, we realize that it doesn't just
- 9:10follow the American I's. From time to
- 9:13time, on certain tests, it actually
- 9:14comes out ahead. In concrete terms, if
- 9:16you ask it a complicated problem,
- 9:18something involving logic or math, well
- 9:21instead of giving you a random answer,
- 9:23it will take the time to reason step by
- 9:25step in front of you, exactly like the
- 9:27best AIs of the moment do. Then, like
- 9:29the best models, it has an enormous
- 9:31memory. Its latest model, version 4 Pro
- 9:34, is capable of remembering up to 1
- 9:36million tokens at a time, which is
- 9:38about 750,000 words. Basically, it's as
- 9:41if you gave it the entire Lord of the
- 9:43Rings trilogy at once and it still had
- 9:46room left. And in there, you'll be able
- 9:48to ask it for any specific information,
- 9:49hidden somewhere, and it'll be able to
- 9:51find it. For example, a very specific
- 9:53detail, page 800, line 14. After that,
- 9:55since the best can translate, write,
- 9:57code and analyze, for the vast majority
- 10:00of what you would do with AI on a daily
- 10:02basis, honestly, you wouldn't even
- 10:04notice the difference with paid AIs.
- 10:06Now, is it better than Chat GPT or
- 10:08Claude? Well, for me, no. And I'll be
- 10:11honest with you, on a lot of things,
- 10:13it's neck and neck with these AIs. But
- 10:15when it comes to the toughest tasks,
- 10:17the most advanced reasoning questions,
- 10:19or the big files to analyze, the most
- 10:20expensive American models still
- 10:22maintain a slight lead, even more so
- 10:23since the release of the latest Fable 5
- 10:25model from Claude or the Chat GPT model
- 10:27. But what I really appreciate about
- 10:30them is that dips say it themselves and
- 10:32fully assume it. In their own report,
- 10:34they write in black and white that they
- 10:36are about 6 months behind the best
- 10:38models on the market. And a company
- 10:40that itself recognizes its delay in
- 10:42this field. Honestly, this is rare
- 10:44enough to be highlighted. So now, let's
- 10:46be clear, Deepsic clearly does not
- 10:48crush everyone. It's just almost as
- 10:50good, free, and open. And for most
- 10:53people, that's more than enough. Now,
- 10:57how can you actually use it? Well, the
- 11:00easiest way, as you've already seen at
- 11:02the beginning of the video, is the
- 11:03website or the app on your phone. You
- 11:05create an account and it's free. But
- 11:07there's one thing that JPT or Cloud
- 11:09Chat doesn't let you do, and that
- 11:10changes a lot of things. Like Deepsic's
- 11:12AI, it's open source, so it doesn't
- 11:15just live in a single application. It
- 11:17can be found almost anywhere in other
- 11:19applications if you want. It's in other
- 11:21tools connected to other services. So
- 11:24you're never locked into a single
- 11:25provider, and that allows you to do
- 11:27incredible things. For the more
- 11:29advanced among you, you may have
- 11:31already heard of cloud code or codex.
- 11:33These are AI agents that don't just
- 11:35respond to you. They're capable of
- 11:36taking action, they can write code,
- 11:38create files, launch commands, all
- 11:40autonomously on your computer. In
- 11:42reality, cloud code or Codex are just
- 11:45tools. The real engine behind it is the
- 11:48GPT 5.5 model, for example, or the Opus
- 11:504.8 model from Entrepiquic. Well, for
- 11:53most of these tools, you are capable of
- 11:55manually changing the engine, which
- 11:56allows you to run them using Dipsic as
- 11:58a replacement instead. Cloud Code, for
- 12:00example, can connect directly to Dipsic
- 12:03if you want, and there are free tools
- 12:05like Open Code that are designed
- 12:07specifically to accept any model.
- 12:09However, in this case, you have to
- 12:10switch to pay-per-use. This is called
- 12:13using the API. And since Dips costs a
- 12:15fraction of the price of Clode or Chat
- 12:17GPT for their model, you will often pay
- 12:19much, much less for the same use. Now,
- 12:23be careful. If we start from this
- 12:25principle, the Clude Code agent, for
- 12:27example, will run on your machine, on
- 12:28your computer. However, the engine, the
- 12:31AI model itself, runs on the cloud, and
- 12:33most of the time, it is on Deepsy's
- 12:35servers. Remember this, we'll talk
- 12:38about it again in 2 minutes because
- 12:39this is exactly where many people fall
- 12:41into the trap. Since this video began,
- 12:44we have repeated one word: open source.
- 12:46And I think it's time to explain what
- 12:49that really means because a lot of
- 12:51people still get it wrong. In general,
- 12:53people think that open source is free.
- 12:55It means that you can download LIA and
- 12:57run it comfortably at home on your
- 12:58small computer. Except that's where
- 13:00it's quite misleading. Because to get
- 13:02an idea of the power of a DIIA model,
- 13:04we're going to systematically look at
- 13:06the same thing: its number of
- 13:07parameters. Basically, the more
- 13:09parameters, the more powerful the model
- 13:11. And the latest version of Deepsic,
- 13:13the most powerful one that rivals the
- 13:16best models, has exactly 1,600 billion
- 13:19parameters, which is a lot of zeros. So
- 13:21yes, remember, as we said, with each
- 13:23question, it will only trigger a very
- 13:25small part of it. Except that to run it
- 13:27, you still need to have these 1,600
- 13:29billion parameters on hand. And it's so
- 13:32huge that you need real professional
- 13:34equipment, not just your little
- 13:35portable MacBook. And here, we're
- 13:37talking about a machine at several tens
- 13:39of thousands of euros. So open source
- 13:41doesn't necessarily mean it's
- 13:43systematically on your computer. It
- 13:45means that the model is just public and
- 13:48that anyone can get hold of it and
- 13:49offer it. And believe me, that has a
- 13:52huge consequence. A Chinese one is free
- 13:56. The legitimate question that everyone
- 13:59might ask is, does my data go to China?
- 14:01Let's be clear, on the official website
- 14:03, the answer is yes. And Dipsy doesn't
- 14:05even hide it. As you've just seen, it's
- 14:08written in black and white in their
- 14:09privacy policy. It collects, processes,
- 14:12and stores your personal data in the
- 14:14Republic of China, your messages, the
- 14:16files you send, your history, your
- 14:18login address. And in fact, in the
- 14:21United States and Australia, Zipsic has
- 14:23been banned from all government devices
- 14:25. In Italy, the app was even pulled
- 14:28from both the Apple and Play stores.
- 14:30And now, you might be thinking, “
- 14:32Superias,” all this to tell me that
- 14:33it’s too risky to use it, and
- 14:35that’s exactly where you fall into
- 14:36the trap. Because the trap is to
- 14:38believe that you only have one entry
- 14:40point, the official Deepsic website.
- 14:43Because remember, as we’ve seen
- 14:45several times, Deepsic is open source,
- 14:47so anyone can host it, including
- 14:49trusted players. And today, you can use
- 14:52exactly the same model at Amazon,
- 14:54Microsoft, or European hosts. And then,
- 14:57well, your data no longer goes to China
- 14:58at all. They stay with the host you
- 15:00have chosen under American or European
- 15:03law, not Chinese law. And that's the
- 15:05real strength of open source because
- 15:07you can never run a closed AI like Chat
- 15:09JPT or Claude yourself, or even with
- 15:11the host of your choice. An open AI
- 15:14like dipsic, yes, you will always have
- 15:16the choice. And even if tomorrow de
- 15:18Dieps decides to close everything, you
- 15:19will still be able to use it. The model
- 15:22is Chinese, it's true, but where you
- 15:24decide to send your data, well, that's
- 15:25up to you and ultimately, you have
- 15:27total control. I'm also going to give
- 15:30you my honest opinion on dipsic. Is it
- 15:32worth it? If you want a very powerful,
- 15:35free AI for writing, translating,
- 15:37coding, and learning on a daily basis,
- 15:39then honestly yes, it's well worth it.
- 15:41Try it, you have nothing to lose, it's
- 15:43free. If you are a developer and you
- 15:46run AI on a very large scale, meaning
- 15:48you use it a lot, then the price
- 15:49clearly changes everything and can
- 15:51literally divide your token bills by
- 15:5310. If you handle sensitive personal or
- 15:55professional information, then be
- 15:56careful, think twice before pasting
- 15:58everything on the official website and
- 16:00always keep in mind where your data is
- 16:02going. Because basically, what dipsy
- 16:04has shown us is that AI, the most
- 16:06powerful, is no longer reserved for a
- 16:07handful of giants. It is becoming a
- 16:09common good and becoming commoditized.
- 16:12And the hardest thing today is not even
- 16:13having access to AI, it is mainly
- 16:15knowing what to do with it. That is
- 16:16exactly why we created a specialized
- 16:19school to train the next generation of
- 16:21experts in generative AI. You will find
- 16:23all the key training courses to master
- 16:25the subject and improve your skills,
- 16:27weekly workshops with my team and
- 16:28myself, and all the technical support
- 16:30you need to never get stuck. Today,
- 16:32there are more than 700 experts active
- 16:34in the market, and some of our students
- 16:36, like Nadia, have already landed their
- 16:38first contract worth several thousand
- 16:40euros. If you are interested, the link
- 16:41is in the description. I hope you liked
- 16:43this video and now better grasp the
- 16:45impact of Deepsx on the market.
About this transcript
This page contains the full transcript of Deepseek a résolu le plus gros problème de l’IA by Yassine Sdiri, generated from the public captions YouTube serves with the video. The transcript has 3,319 words across 485 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.