YouTube2Text

Deepseek a résolu le plus gros problème de l’IA — Transcript

by Yassine Sdiri · 3,319 words · 485 segments · language en · Watch on YouTube

Full transcript

  1. 0:00For years, the AIs that everyone uses
  2. 0:03have had the same problem. They cost
  3. 0:06billions to develop and are locked down
  4. 0:08by a handful of American companies. But
  5. 0:11recently, a Chinese AI has solved this
  6. 0:13problem and no one saw it coming. You
  7. 0:16may have already heard of it because
  8. 0:18over a year ago, when it was first
  9. 0:20released, it caused the biggest loss in
  10. 0:23stock market history in a single day.
  11. 0:26Almost $ 600 billion went up in smoke,
  12. 0:29and it has just repeated the feat with
  13. 0:31its latest artificial intelligence
  14. 0:34model. Tatia, her name is Dipsic, and
  15. 0:37she has become almost as powerful as
  16. 0:39the best GPT and cloud cat models,
  17. 0:41except that she is open source and free
  18. 0:43. And all this power that was once
  19. 0:47owned by a few American companies, can
  20. 0:49now be used by anyone. I will explain
  21. 0:52to you in 15 minutes how they did it,
  22. 0:54what it changes for you and how to take
  23. 0:56advantage of it. Now, I know what
  24. 0:58you're thinking. You're thinking that
  25. 1:01something this powerful, free, and
  26. 1:02Chinese, well, it smells like a trap.
  27. 1:04And it's true, there's one not
  28. 1:06necessarily where you expect it to be,
  29. 1:07and I'll show it to you at the end of
  30. 1:09this video. Before explaining how this
  31. 1:12AI is revolutionizing the market, let's
  32. 1:14see what it looks like in concrete
  33. 1:16terms. Here's the Deepsy interface. In
  34. 1:18fact, you create an account and you'll
  35. 1:20immediately land on a very simple
  36. 1:21interface that you're probably already
  37. 1:23familiar with because it looks a lot
  38. 1:24like GPT chat or Cloud. And to use it,
  39. 1:27it's exactly the same. You just have to
  40. 1:29ask it what you want. You have three
  41. 1:31main distinct modes: fast, expert, and
  42. 1:33vision. And as its name suggests, fast
  43. 1:36is there to help you get an answer as
  44. 1:38quickly as possible. Then, expert mode
  45. 1:41is for the most complex tasks, and
  46. 1:42finally vision mode is for providing it
  47. 1:44with images or documents so that it can
  48. 1:46analyze them in detail. And for the
  49. 1:49example, we're even going to ask him to
  50. 1:50create a little 3D platform game for us
  51. 1:52to push him to his limits and his
  52. 1:54thinking. We'll come back to that a
  53. 1:56little later when he's finished. But
  54. 1:57the important thing to note that you
  55. 1:59may have noticed on this interface is
  56. 2:00that you can see that he writes very,
  57. 2:02very quickly, much faster than Claude
  58. 2:04or Chat GPT and that will be very
  59. 2:05important in the rest of the video.
  60. 2:08Okay, great, the game works, nothing
  61. 2:10revolutionary but it was just to give
  62. 2:11you a little visual demo. Now, having
  63. 2:15this result in a single message, a
  64. 2:16single prompt that I sent him, is still
  65. 2:18impressive. Of course, if you want to
  66. 2:21go even further, you can simply ask him
  67. 2:23by refining and making changes with him
  68. 2:24. But the really interesting detail
  69. 2:26about the interface is elsewhere. On
  70. 2:28Chat, GPT or Cloud, you probably know
  71. 2:30this button that systematically pushes
  72. 2:33you to the paid version. Sometimes,
  73. 2:35it's around the € 20 version and
  74. 2:36sometimes even the € 200 per month
  75. 2:38version. And of course, with Dipsic,
  76. 2:40you can search for this button as long
  77. 2:42as you want. You won't find it. And
  78. 2:44it's very simple, it's because it
  79. 2:45doesn't exist. They don't even offer
  80. 2:47you to pay to use this tool. Dipsic is
  81. 2:50free, completely free. So naturally, I
  82. 2:53wanted to test their brand new AI model
  83. 2:55for you. I've been using it every day
  84. 2:58for over a week now instead of all my
  85. 3:00usual tools. In this video, we will
  86. 3:02look at four points together. First, in
  87. 3:04one, we're going to understand where
  88. 3:05this AI comes from and how a small
  89. 3:07Chinese lab managed to do something
  90. 3:09that no one had done before. In two,
  91. 3:11we're going to see what it's really
  92. 3:12capable of doing, but also its
  93. 3:14limitations. In three, we're going to
  94. 3:16see how you can use it, including for
  95. 3:17tasks you don't even suspect. And
  96. 3:19finally, in four, we'll see if all this
  97. 3:21is risk-free because the Chinese
  98. 3:22company is free, which inevitably
  99. 3:24raises questions. So, to understand why
  100. 3:27Dipsic is making so much noise, we
  101. 3:29first need to know where it comes from.
  102. 3:31Because Dipsic is far from being a tech
  103. 3:34giant. It's simply a small Chinese
  104. 3:37laboratory that originally wasn't even
  105. 3:39a tech company but an investment fund.
  106. 3:41So it's a small team with far fewer
  107. 3:43people and, above all, far less money
  108. 3:45than the American giants. They have a
  109. 3:47huge handicap because due to American
  110. 3:49restrictions on China, they don't even
  111. 3:51have access to the most powerful chips,
  112. 3:54the ones that run all the American
  113. 3:55chips. So on paper, he was clearly the
  114. 3:58underdog. Except that it's often
  115. 4:00precisely this kind of handicap that
  116. 4:02makes a team formidable. Because when
  117. 4:04you have everything, well, you often
  118. 4:05fall into comfort. When you have
  119. 4:07nothing, you excel. And that's exactly
  120. 4:09what happened because a little over a
  121. 4:11year ago, they released a model that
  122. 4:13rivals the best American I-series for a
  123. 4:15fraction of the price. The numbers are
  124. 4:17a bit of a debate today, but we're
  125. 4:19talking about at least 20 times less
  126. 4:20than what it cost to train Chat GPT.
  127. 4:23And that's what set the American stock
  128. 4:25market on fire because overnight,
  129. 4:26everyone asked themselves the same
  130. 4:28question. If they can do it with next
  131. 4:31to nothing, why do others have to spend
  132. 4:32billions? And to give you a fairly
  133. 4:35simple example that explains the
  134. 4:36reaction of the stock markets, imagine
  135. 4:38that you're building your house. You
  136. 4:41put a lot of money into it to have the
  137. 4:42most beautiful house in the
  138. 4:44neighborhood and, the next day, you
  139. 4:45realize that the neighbor right next
  140. 4:47door built exactly the same house,
  141. 4:49except that he paid 10 times less. So,
  142. 4:53not only would I think you would be
  143. 4:54shocked, but you would also be very
  144. 4:56angry with the construction company
  145. 4:57that overcharged you for the same
  146. 4:59result. Well, that's exactly how Wall
  147. 5:03Street felt about Silicon Valley
  148. 5:05because the Americans like to do one
  149. 5:07thing to go faster: add more and more
  150. 5:09money, more machines, more power to
  151. 5:10crush the competition. Dipsic simply
  152. 5:14couldn't do that. So instead, they just
  153. 5:17got smarter. In fact, to help you
  154. 5:19understand a bit better, when you ask a
  155. 5:22classic AI a question, it will use all
  156. 5:24its power, everything it is capable of
  157. 5:26doing, even for a very simple question.
  158. 5:29Imagine for 2 seconds, you run a
  159. 5:30company with thousands of employees and
  160. 5:32you have a small, specific accounting
  161. 5:34question. Are you really going to call
  162. 5:37everyone into a meeting? Well, of
  163. 5:38course not. You go straight to the
  164. 5:40accountant and the others continue to
  165. 5:42work quietly on their own. Well,
  166. 5:44Deepsic, that's exactly what they've
  167. 5:45set up. They have a huge model, but for
  168. 5:48each question, it only wakes up the few
  169. 5:50specialists it really needs. All the
  170. 5:52others remain asleep. It has the power
  171. 5:54of a large model and the speed of a
  172. 5:56small one. So, of course, it responds
  173. 5:58much faster, and that's important. Now,
  174. 6:01this idea clearly wasn't invented by
  175. 6:03them. It existed long before. Except
  176. 6:04that before them, no one had pushed it
  177. 6:06that far. And above all, they've done
  178. 6:08something that almost no one else does
  179. 6:10in this shark environment. They share
  180. 6:12almost everything. Well, not absolutely
  181. 6:14everything, it's important to be
  182. 6:15precise. The exact recipe for training
  183. 6:17the model and the data they use, well,
  184. 6:19they keep that to themselves. But for
  185. 6:21everything else, they show a generosity
  186. 6:23rare in this environment. First, they
  187. 6:25make the model itself freely available,
  188. 6:27meaning that anyone can download it.
  189. 6:30And most importantly, they publish a
  190. 6:31huge number of papers, dozens of
  191. 6:33research papers in which they explain
  192. 6:34their findings and their tricks,
  193. 6:36including how to make the machine run
  194. 6:37much more cheaply. And that's typically
  195. 6:41the kind of thing that American
  196. 6:42laboratories keep carefully hidden,
  197. 6:44literally like industrial secrets. So
  198. 6:47now, to properly measure the impact of
  199. 6:49this approach on the market, let's
  200. 6:51imagine something. Let's take a farmer
  201. 6:53who finds a method to produce twice as
  202. 6:55much wheat with half the water.
  203. 6:57Normally, what does he do? Well, he
  204. 6:58keeps his method well hidden to himself
  205. 7:00. He crushes his competition and sells
  206. 7:02his wheat at the price he wants. Well,
  207. 7:04in Psic, they do the exact opposite.
  208. 7:06They publicly explain how they did it
  209. 7:08for free so that even their opponents
  210. 7:10can draw inspiration from it. And that
  211. 7:13triggers two phenomena. The first is
  212. 7:15competition. Dipsic sells a nia at a
  213. 7:17fraction of the American price. As a
  214. 7:20result, all the others are forced to
  215. 7:21fall into line or explain to everyone
  216. 7:23why they are 10 times more expensive.
  217. 7:25And the second one is even more
  218. 7:27interesting. Since almost everything is
  219. 7:30published, other AI labs can read,
  220. 7:31understand, and apply their findings to
  221. 7:33lower their own costs. In other words,
  222. 7:36Dipsic doesn't just undercut the prices
  223. 7:38of the biggest AI suppliers, it also
  224. 7:41gives other labs the means to undercut
  225. 7:43them. And you'll see why it snowballs.
  226. 7:45Because since they did that, other
  227. 7:47giants, particularly in China, have
  228. 7:49started to open up their models in turn
  229. 7:50. And today, Chinese nias cost 4 to 10
  230. 7:53times less than their American
  231. 7:55equivalents thanks to these famous
  232. 7:57Dipsic findings. So now, we might begin
  233. 7:59to wonder exactly why they are doing
  234. 8:01all of this? And frankly, it's pretty
  235. 8:03smart of them. When you're the
  236. 8:05challenger in this market, making
  237. 8:07technology free for everyone is the
  238. 8:08best way to destroy the lead of those
  239. 8:10who are ahead. And by the way, even you
  240. 8:13behind this screen in your daily life
  241. 8:15can see the results of this approach.
  242. 8:16This pressure that they have managed to
  243. 8:18create, well, it even affects the AI
  244. 8:19that you use every day. And if we go
  245. 8:22back to our example of the farmer, the
  246. 8:23price of the infinite bit has dropped
  247. 8:25for everyone, even for those who have
  248. 8:27never heard of our farmer and his new
  249. 8:28method. So Lia cost a fortune to create
  250. 8:32and run. It was locked to a few people
  251. 8:35and Dipsic has just solved this problem
  252. 8:37not just for them but for an entire
  253. 8:40industry. And of course, when something
  254. 8:42costs next to nothing, we can afford to
  255. 8:44give it away for free and that's where
  256. 8:46you really get into the story. Okay, a
  257. 8:49free AI, but does it really hold its
  258. 8:52own against GPT chat and Claud? Because
  259. 8:54free AI that talks absolute nonsense is
  260. 8:56useless and nobody even cares about it.
  261. 8:59So, as we saw in the demo, it is
  262. 9:00capable of doing quite complex things.
  263. 9:03Then, if we look at the comparisons on
  264. 9:05some of the hardest thinking tests to
  265. 9:07update the different model levels, well
  266. 9:09, we realize that it doesn't just
  267. 9:10follow the American I's. From time to
  268. 9:13time, on certain tests, it actually
  269. 9:14comes out ahead. In concrete terms, if
  270. 9:16you ask it a complicated problem,
  271. 9:18something involving logic or math, well
  272. 9:21instead of giving you a random answer,
  273. 9:23it will take the time to reason step by
  274. 9:25step in front of you, exactly like the
  275. 9:27best AIs of the moment do. Then, like
  276. 9:29the best models, it has an enormous
  277. 9:31memory. Its latest model, version 4 Pro
  278. 9:34, is capable of remembering up to 1
  279. 9:36million tokens at a time, which is
  280. 9:38about 750,000 words. Basically, it's as
  281. 9:41if you gave it the entire Lord of the
  282. 9:43Rings trilogy at once and it still had
  283. 9:46room left. And in there, you'll be able
  284. 9:48to ask it for any specific information,
  285. 9:49hidden somewhere, and it'll be able to
  286. 9:51find it. For example, a very specific
  287. 9:53detail, page 800, line 14. After that,
  288. 9:55since the best can translate, write,
  289. 9:57code and analyze, for the vast majority
  290. 10:00of what you would do with AI on a daily
  291. 10:02basis, honestly, you wouldn't even
  292. 10:04notice the difference with paid AIs.
  293. 10:06Now, is it better than Chat GPT or
  294. 10:08Claude? Well, for me, no. And I'll be
  295. 10:11honest with you, on a lot of things,
  296. 10:13it's neck and neck with these AIs. But
  297. 10:15when it comes to the toughest tasks,
  298. 10:17the most advanced reasoning questions,
  299. 10:19or the big files to analyze, the most
  300. 10:20expensive American models still
  301. 10:22maintain a slight lead, even more so
  302. 10:23since the release of the latest Fable 5
  303. 10:25model from Claude or the Chat GPT model
  304. 10:27. But what I really appreciate about
  305. 10:30them is that dips say it themselves and
  306. 10:32fully assume it. In their own report,
  307. 10:34they write in black and white that they
  308. 10:36are about 6 months behind the best
  309. 10:38models on the market. And a company
  310. 10:40that itself recognizes its delay in
  311. 10:42this field. Honestly, this is rare
  312. 10:44enough to be highlighted. So now, let's
  313. 10:46be clear, Deepsic clearly does not
  314. 10:48crush everyone. It's just almost as
  315. 10:50good, free, and open. And for most
  316. 10:53people, that's more than enough. Now,
  317. 10:57how can you actually use it? Well, the
  318. 11:00easiest way, as you've already seen at
  319. 11:02the beginning of the video, is the
  320. 11:03website or the app on your phone. You
  321. 11:05create an account and it's free. But
  322. 11:07there's one thing that JPT or Cloud
  323. 11:09Chat doesn't let you do, and that
  324. 11:10changes a lot of things. Like Deepsic's
  325. 11:12AI, it's open source, so it doesn't
  326. 11:15just live in a single application. It
  327. 11:17can be found almost anywhere in other
  328. 11:19applications if you want. It's in other
  329. 11:21tools connected to other services. So
  330. 11:24you're never locked into a single
  331. 11:25provider, and that allows you to do
  332. 11:27incredible things. For the more
  333. 11:29advanced among you, you may have
  334. 11:31already heard of cloud code or codex.
  335. 11:33These are AI agents that don't just
  336. 11:35respond to you. They're capable of
  337. 11:36taking action, they can write code,
  338. 11:38create files, launch commands, all
  339. 11:40autonomously on your computer. In
  340. 11:42reality, cloud code or Codex are just
  341. 11:45tools. The real engine behind it is the
  342. 11:48GPT 5.5 model, for example, or the Opus
  343. 11:504.8 model from Entrepiquic. Well, for
  344. 11:53most of these tools, you are capable of
  345. 11:55manually changing the engine, which
  346. 11:56allows you to run them using Dipsic as
  347. 11:58a replacement instead. Cloud Code, for
  348. 12:00example, can connect directly to Dipsic
  349. 12:03if you want, and there are free tools
  350. 12:05like Open Code that are designed
  351. 12:07specifically to accept any model.
  352. 12:09However, in this case, you have to
  353. 12:10switch to pay-per-use. This is called
  354. 12:13using the API. And since Dips costs a
  355. 12:15fraction of the price of Clode or Chat
  356. 12:17GPT for their model, you will often pay
  357. 12:19much, much less for the same use. Now,
  358. 12:23be careful. If we start from this
  359. 12:25principle, the Clude Code agent, for
  360. 12:27example, will run on your machine, on
  361. 12:28your computer. However, the engine, the
  362. 12:31AI model itself, runs on the cloud, and
  363. 12:33most of the time, it is on Deepsy's
  364. 12:35servers. Remember this, we'll talk
  365. 12:38about it again in 2 minutes because
  366. 12:39this is exactly where many people fall
  367. 12:41into the trap. Since this video began,
  368. 12:44we have repeated one word: open source.
  369. 12:46And I think it's time to explain what
  370. 12:49that really means because a lot of
  371. 12:51people still get it wrong. In general,
  372. 12:53people think that open source is free.
  373. 12:55It means that you can download LIA and
  374. 12:57run it comfortably at home on your
  375. 12:58small computer. Except that's where
  376. 13:00it's quite misleading. Because to get
  377. 13:02an idea of the power of a DIIA model,
  378. 13:04we're going to systematically look at
  379. 13:06the same thing: its number of
  380. 13:07parameters. Basically, the more
  381. 13:09parameters, the more powerful the model
  382. 13:11. And the latest version of Deepsic,
  383. 13:13the most powerful one that rivals the
  384. 13:16best models, has exactly 1,600 billion
  385. 13:19parameters, which is a lot of zeros. So
  386. 13:21yes, remember, as we said, with each
  387. 13:23question, it will only trigger a very
  388. 13:25small part of it. Except that to run it
  389. 13:27, you still need to have these 1,600
  390. 13:29billion parameters on hand. And it's so
  391. 13:32huge that you need real professional
  392. 13:34equipment, not just your little
  393. 13:35portable MacBook. And here, we're
  394. 13:37talking about a machine at several tens
  395. 13:39of thousands of euros. So open source
  396. 13:41doesn't necessarily mean it's
  397. 13:43systematically on your computer. It
  398. 13:45means that the model is just public and
  399. 13:48that anyone can get hold of it and
  400. 13:49offer it. And believe me, that has a
  401. 13:52huge consequence. A Chinese one is free
  402. 13:56. The legitimate question that everyone
  403. 13:59might ask is, does my data go to China?
  404. 14:01Let's be clear, on the official website
  405. 14:03, the answer is yes. And Dipsy doesn't
  406. 14:05even hide it. As you've just seen, it's
  407. 14:08written in black and white in their
  408. 14:09privacy policy. It collects, processes,
  409. 14:12and stores your personal data in the
  410. 14:14Republic of China, your messages, the
  411. 14:16files you send, your history, your
  412. 14:18login address. And in fact, in the
  413. 14:21United States and Australia, Zipsic has
  414. 14:23been banned from all government devices
  415. 14:25. In Italy, the app was even pulled
  416. 14:28from both the Apple and Play stores.
  417. 14:30And now, you might be thinking, “
  418. 14:32Superias,” all this to tell me that
  419. 14:33it’s too risky to use it, and
  420. 14:35that’s exactly where you fall into
  421. 14:36the trap. Because the trap is to
  422. 14:38believe that you only have one entry
  423. 14:40point, the official Deepsic website.
  424. 14:43Because remember, as we’ve seen
  425. 14:45several times, Deepsic is open source,
  426. 14:47so anyone can host it, including
  427. 14:49trusted players. And today, you can use
  428. 14:52exactly the same model at Amazon,
  429. 14:54Microsoft, or European hosts. And then,
  430. 14:57well, your data no longer goes to China
  431. 14:58at all. They stay with the host you
  432. 15:00have chosen under American or European
  433. 15:03law, not Chinese law. And that's the
  434. 15:05real strength of open source because
  435. 15:07you can never run a closed AI like Chat
  436. 15:09JPT or Claude yourself, or even with
  437. 15:11the host of your choice. An open AI
  438. 15:14like dipsic, yes, you will always have
  439. 15:16the choice. And even if tomorrow de
  440. 15:18Dieps decides to close everything, you
  441. 15:19will still be able to use it. The model
  442. 15:22is Chinese, it's true, but where you
  443. 15:24decide to send your data, well, that's
  444. 15:25up to you and ultimately, you have
  445. 15:27total control. I'm also going to give
  446. 15:30you my honest opinion on dipsic. Is it
  447. 15:32worth it? If you want a very powerful,
  448. 15:35free AI for writing, translating,
  449. 15:37coding, and learning on a daily basis,
  450. 15:39then honestly yes, it's well worth it.
  451. 15:41Try it, you have nothing to lose, it's
  452. 15:43free. If you are a developer and you
  453. 15:46run AI on a very large scale, meaning
  454. 15:48you use it a lot, then the price
  455. 15:49clearly changes everything and can
  456. 15:51literally divide your token bills by
  457. 15:5310. If you handle sensitive personal or
  458. 15:55professional information, then be
  459. 15:56careful, think twice before pasting
  460. 15:58everything on the official website and
  461. 16:00always keep in mind where your data is
  462. 16:02going. Because basically, what dipsy
  463. 16:04has shown us is that AI, the most
  464. 16:06powerful, is no longer reserved for a
  465. 16:07handful of giants. It is becoming a
  466. 16:09common good and becoming commoditized.
  467. 16:12And the hardest thing today is not even
  468. 16:13having access to AI, it is mainly
  469. 16:15knowing what to do with it. That is
  470. 16:16exactly why we created a specialized
  471. 16:19school to train the next generation of
  472. 16:21experts in generative AI. You will find
  473. 16:23all the key training courses to master
  474. 16:25the subject and improve your skills,
  475. 16:27weekly workshops with my team and
  476. 16:28myself, and all the technical support
  477. 16:30you need to never get stuck. Today,
  478. 16:32there are more than 700 experts active
  479. 16:34in the market, and some of our students
  480. 16:36, like Nadia, have already landed their
  481. 16:38first contract worth several thousand
  482. 16:40euros. If you are interested, the link
  483. 16:41is in the description. I hope you liked
  484. 16:43this video and now better grasp the
  485. 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.