The only AutoResearch tutorial you’ll ever need — Transcript

Full transcript

1. 0:00My name is David Andre and this is the
2. 0:01clearest explanation of auto research
3. 0:04you're going to find on the internet.
4. 0:05So, what is auto research? It's an open
5. 0:08source project by Andre Karpathy that
6. 0:10lets AI improve itself autonomously. You
7. 0:12have an AI agent that runs experiments
8. 0:15automatically and it keeps what works
9. 0:17and throws away everything that doesn't.
10. 0:18So, in this video, I'm going to explain
11. 0:20how auto research actually works, how
12. 0:22you can use it in your own life and
13. 0:24business because there are a lot of
14. 0:25implications, and how you can build your
15. 0:27very first auto research loop. So, if
16. 0:29you really want to be on the cutting
17. 0:31edge of AI, make sure to watch until the
18. 0:33end. First off, who is Andre Karpathy?
19. 0:35He's one of the most legendary AI
20. 0:37researchers of all time. He's one of the
21. 0:39OpenAI co-founders and he is the main
22. 0:41person behind Tesla Autopilot. Oh, yeah,
23. 0:44and he's also the guy who invented the
24. 0:45term vibe coding. Plus, he has
25. 0:47contributed a lot to the open source
26. 0:50community, especially in AI. And yes,
27. 0:52most importantly, he was born in
28. 0:54Czechoslovakia. However, this video is
29. 0:56about one of his contributions and that
30. 0:58is auto research. Karpathy had a
31. 1:00training script for the GPT-2 model that
32. 1:03he's been optimizing for many months.
33. 1:04And then he realized, "Why am I doing
34. 1:06this? Why don't I just have an AI agent
35. 1:08run different experiments in a loop to
36. 1:11figure out what's the best way to
37. 1:12optimize this model?" And this is the
38. 1:14core idea. Give AI one file, one metric,
39. 1:18and let it run hundreds, if not
40. 1:20thousands, of experiments by itself
41. 1:22while you sleep and watch it improve.
42. 1:24And yes, at the end of the video, I'll
43. 1:25show you how you can build your first
44. 1:27auto research loop. And yes, you can do
45. 1:30this even if you're a complete beginner.
46. 1:32Now, let's talk about the prepare.py
47. 1:34file because it's essential to
48. 1:35understand this. The agent cannot touch
49. 1:37prepare.py so that it can't cheat the
50. 1:39eval you set for it. Without this
51. 1:41limitation, it could rewrite the scoring
52. 1:43function to fake its results. So,
53. 1:45basically, the prepare.py file defines
54. 1:48what better means. Now, of course, if
55. 1:50you set the wrong metric, you'll get the
56. 1:51wrong results. So, here is a nice
57. 1:53graphic that visualizes the loop. First,
58. 1:56the agent comes up with a hypothesis,
59. 1:58right? A theory of what it could
60. 1:59improve, what experiment it could run.
61. 2:02Then it modifies the code, you train it
62. 2:04for around 5 minutes, then it runs the
63. 2:05evaluation to see if it's good or if
64. 2:08it's not good. If it wants to keep it,
65. 2:09it will just commit it to your Git
66. 2:11history. If it's bad and if it worsens
67. 2:13the results, it will do Git reset and
68. 2:15repeat the whole loop over and over and
69. 2:17over as many times as you want, you
70. 2:20know,
71. 2:20>> [laughter]
72. 2:20>> depending on how many dollars you want
73. 2:22to spend or tokens. And yeah, this is
74. 2:24auto research basically in a single
75. 2:25image. So, if you start it right before
76. 2:27going to bed, it can run roughly 100
77. 2:30experiments overnight. Now, by giving it
78. 2:32a fixed time budget, you make every
79. 2:34experiment directly comparable because
80. 2:35let's say you're hiring for your
81. 2:36company. If you have one applicant and
82. 2:38you give him 7 days to complete the task
83. 2:41and the other one only has 7 minutes,
84. 2:43then obviously the one who has 7 days,
85. 2:45on average, will do better. So, you want
86. 2:48to make sure this is the same in AI
87. 2:49agents. That way the agent can't cheat
88. 2:52just by training longer or by training
89. 2:54better ideas. Only the raw idea wins
90. 2:56with the same time allocated. But if you
91. 2:58think that auto research only applies to
92. 3:00the training of AI models, you are badly
93. 3:03mistaken. This has massive implications
94. 3:05across every domain of your life or
95. 3:07business. You can build a recursive
96. 3:09self-improving loop for nearly anything
97. 3:12that can be measured. Thanks to AI
98. 3:13agents, soon enough, the execution of
99. 3:15any work or task will become basically
100. 3:18free. However, what will become valuable
101. 3:20is knowing what to measure, picking the
102. 3:23right metric, and setting the right
103. 3:24constraints. This is the skill that is
104. 3:27going to make millionaires in the
105. 3:28future. And by the way, this is the
106. 3:30clearest example of what AI agents
107. 3:32actually look like in practice. Not just
108. 3:34chatbots, but real autonomous loops that
109. 3:37do meaningful work for companies or
110. 3:39individuals. And if you reason about the
111. 3:41state of AI from first principles, you
112. 3:43realize that this has been the end goal
113. 3:46all along. Andre Karpathy predicts that
114. 3:48all LLM frontier labs will do this. They
115. 3:51will run some sort of auto research.
116. 3:53This is the final boss battle. And if
117. 3:55you think about it, this is literally
118. 3:56what recursive [clears throat]
119. 3:57self-improvement will look like. And the
120. 3:59funny thing is that right now, the AI
121. 4:01labs like OpenAI and Anthropic Gemini
122. 4:03are spending tens of millions of dollars
123. 4:05on researchers, all of which are trying
124. 4:07to build this. Yet, Karpathy made it
125. 4:10completely open source. Now, to
126. 4:11understand auto research, you must
127. 4:13understand its three-file architecture.
128. 4:16First, you have program.md. This is the
129. 4:18most important file. This is the human
130. 4:20setting the goal, the constraints, and
131. 4:22the rules for the agent. Then you have
132. 4:24the train.py file. This is the one file
133. 4:26that the agent can actually change. And
134. 4:28this could be anything, by the way.
135. 4:30Code, some config, a prompt, some math
136. 4:32equation, literally anything you want to
137. 4:35optimize. And then you have prepare.py.
138. 4:37This is the metric and the evaluation
139. 4:39script. The agent cannot touch this.
140. 4:41Absolutely never. Because this is what
141. 4:43measures the result. And by the way,
142. 4:45there are several tech billionaires who
143. 4:47are going crazy over this, such as the
144. 4:48CEO of Shopify and the CEO of Stripe,
145. 4:52because these guys realized that this is
146. 4:53not just about training AI models. And
147. 4:55this perhaps the biggest misconception
148. 4:57about auto research. People think that
149. 4:59it's just about optimizing of machine
150. 5:02learning, right? Because this is what
151. 5:03Karpathy used as the example. But if you
152. 5:05actually pay attention and if you think
153. 5:07about it a little bit more, you realize
154. 5:09that as long as you have a clear metric
155. 5:11and you can run these experiments, this
156. 5:12could be used in so many different ways.
157. 5:14In marketing, in business, in testing
158. 5:16new products, in trading strategies, in
159. 5:18personal life, in weather forecast,
160. 5:21literally hundreds of possible domains.
161. 5:23You can build these auto research loops
162. 5:24that don't have any human in the loop,
163. 5:26but instead swarm of AI agents running
164. 5:28experiments, discarding what doesn't
165. 5:30work and keeping what works. And as
166. 5:32Karpathy says here, we might be in the
167. 5:34early stages of the singularity. Now,
168. 5:36every AI workflow has the same
169. 5:38bottleneck, getting fresh real-world
170. 5:40data. You can build the most
171. 5:41sophisticated AI agent, but if it can't
172. 5:43see what's actually on the web right
173. 5:45now, it's flying blind. That's where
174. 5:47Oxylabs comes in. The Oxylabs web
175. 5:50scraper API lets you pull structured
176. 5:52data from basically any website. This
177. 5:54could be Amazon product listings, Google
178. 5:55search results, real estate listings,
179. 5:58all through a single API call. It
180. 6:00handles proxy rotation, captcha solving,
181. 6:02and JavaScript rendering so you don't
182. 6:04have to do that yourself. But here is
183. 6:06what makes it interesting for us AI
184. 6:08developers. Oxylabs has an official MCP
185. 6:10that you can connect directly to Cursor
186. 6:12or Cloud Code in a matter of a minute so
187. 6:14that your AI agent gets live web
188. 6:17scraping superpowers. Say you need to
189. 6:19pull competitor's pricing or scrape
190. 6:21search results or grab real estate
191. 6:23listings. Just ask it in plain English.
192. 6:25The agent calls Oxylabs, gets structured
193. 6:28data back, and reasons over it. And even
194. 6:30if you aren't a developer, Oxylabs plugs
195. 6:33straight into N8N. You can visually
196. 6:35build a workflow that scrapes Amazon
197. 6:36prices, sends them to an AI, and spits
198. 6:39out insights with zero code written by
199. 6:42you. And the best part is, Oxylabs gives
200. 6:44up to 2,000 scrape results for free so
201. 6:47you can test it yourself. So, go to
202. 6:49oxylabs.io/david
203. 6:51for the free trial. Oh, and there is no
204. 6:53credit card required. And if you run out
205. 6:54of credits, use my code david for 20%
206. 6:57off all Oxylabs plans. And thank you to
207. 7:00Oxylabs for sponsoring this video.
208. 7:02Again, I need to stress, auto research
209. 7:03is not just for machine learning. This
210. 7:05pattern works anywhere you can measure
211. 7:08an outcome, a clear outcome. And
212. 7:09Karpathy said this himself, "Any metric
213. 7:12you care about that is reasonably
214. 7:13efficient to evaluate can be auto
215. 7:16researched." So, you need one file to
216. 7:17edit, one scalar metric, and a time box
217. 7:20loop. If you can score it, you can auto
218. 7:22research it. Now, let's look at some
219. 7:24practical valuable use cases. First off,
220. 7:26trading. You can take the same auto
221. 7:27research loop, but instead of improving
222. 7:29an AI model, you can point it on a
223. 7:31trading strategy. The agent tweaks your
224. 7:33buy/sell rules, tries experiments based
225. 7:36on years of market data, and scores each
226. 7:38experiment by its sharp ratio, which is
227. 7:40basically how good are the returns to
228. 7:42the risk. And it can test hundreds of
229. 7:44different trading strategies to see
230. 7:46which one has the best returns. Here is
231. 7:47another use case, marketing. You can now
232. 7:49apply auto research to marketing with
233. 7:52emails, ad creatives, landing pages,
234. 7:54automated AB tests, headlines,
235. 7:56thumbnails, YouTube titles, any type of
236. 7:59marketing. Eric Seu put it the best,
237. 8:01"Most marketing teams run 30 experiments
238. 8:03per year. The next generation will run
239. 8:0636,000,
240. 8:07aka roughly 100 per day. The agent will
241. 8:10modify the copy and measure the
242. 8:11conversions and decide whether to keep
243. 8:13this experiment or to discard it." The
244. 8:15exact same loop as before. Now, before I
245. 8:17show you more use cases how auto
246. 8:19research can benefit your life, please
247. 8:21consider subscribing. 25% of you are
248. 8:23subscribed, which means the vast
249. 8:25majority of you watching right now are
250. 8:27not subscribed. So, if you want to see
251. 8:29more high-quality
252. 8:31in your YouTube recommended, please take
253. 8:332 seconds, go below the video, and click
254. 8:35the subscribe button. It helps out more
255. 8:37than you think. So, to all of you who
256. 8:39just subscribed, thank you. The next use
257. 8:41case is for developers. You can point
258. 8:43auto research at basically any code base
259. 8:46and say, "Make it faster." And people
260. 8:48are also using auto research to
261. 8:49fine-tune open source AI models so they
262. 8:51run faster locally on your laptop or
263. 8:54phone. So, expect in the next 6 months
264. 8:56insane breakthroughs in terms of what is
265. 8:59possible to run on phone. I would even
266. 9:01say that we will have Sonnet 4.6 quality
267. 9:03models runnable on iPhones in three or
268. 9:07four months. This is a prediction. Let's
269. 9:09see if it's correct. Another use case is
270. 9:10prompt engineering. Auto research can
271. 9:12fine-tune the system instructions behind
272. 9:14all of your AI agents. So, Harrison
273. 9:16Chase, the founder of LangChain, which
274. 9:18is a billion-dollar company, said,
275. 9:19"Agents mess up because they don't have
276. 9:21the right context." And system prompts
277. 9:23are part of that context. So, auto
278. 9:25research can find new ways to phrase
279. 9:28things. Better language, maybe even
280. 9:29different language, you know? Maybe
281. 9:31instead of English, it can use Polish or
282. 9:33Czech or German, whatever. It can try
283. 9:35different levels, like beginner, you
284. 9:38know, college level, PhD level, to see
285. 9:41which prompt works the best for all of
286. 9:43your AI agents. So, these are the three
287. 9:45conditions that decide whether your auto
288. 9:47research loop will be successful or will
289. 9:50be a failure. Number one, a clear
290. 9:51metric. One number, a clear direction
291. 9:53you want to go in. And number two, an
292. 9:55automated evaluation where there isn't a
293. 9:57human in the loop, right? If you, the
294. 10:00human, need to be in the loop, it will
295. 10:01be so slow, and it will be not be auto
296. 10:03research. Sure, it still can be
297. 10:05research, but it will not be auto
298. 10:07research. It will not be running while
299. 10:08you sleep. And number three, there is
300. 10:10one file that the agent can change. Not
301. 10:12two, not zero, one file. And you need
302. 10:15all of these three conditions for auto
303. 10:17research to work. Now, here's where auto
304. 10:19research will fail. Brand design, UX,
305. 10:21pricing, anything where better is
306. 10:24subjective. Now, for example, in
307. 10:25pricing, you could have it succeed if
308. 10:27you have a large volume of traffic to
309. 10:28your pricing page, and you can quickly
310. 10:30AB test different pricing to see highest
311. 10:33cash collected, but for most businesses,
312. 10:35it will not be effective, right? Because
313. 10:37the quality of that makes better is
314. 10:39subjective, or the loop is too slow. The
315. 10:41loop needs objective metric. If the
316. 10:43success is a judgment call or a feeling,
317. 10:46the agent cannot tell what's working.
318. 10:48So, it will optimize in a random
319. 10:49direction. Also, I want to stress this
320. 10:51again. If you give it a bad metric, it
321. 10:53will very confidently optimize the wrong
322. 10:55thing. So, here is Andrej Karpathy's end
323. 10:57vision. In the early 2000s, there was a
324. 10:59project SETI@home that let anyone donate
325. 11:03their spare computer power to research
326. 11:05for alien life. And Andrej Karpathy
327. 11:07wants to do the same model, the same
328. 11:09idea, but for AI research, where you
329. 11:11have millions of AI agents distributed
330. 11:14across thousands of computers, and you
331. 11:16can actually allocate where that
332. 11:18research goes towards. All right, so
333. 11:20now, let me show you how to actually
334. 11:22build your own auto research loop from
335. 11:24scratch, even if you're a complete
336. 11:26beginner. Just stick with me for the
337. 11:27next 5 minutes, and you'll be ahead of
338. 11:2999.9% of people in AI who just pay
339. 11:32ChatGPT subscription and they think
340. 11:34they're advanced. They cannot even dream
341. 11:36about having their own auto research,
342. 11:38but you are like a couple of steps away
343. 11:39from having that. So, lock in. So, first
344. 11:41of all, this is the GitHub repo. I'm
345. 11:43going to link it below the video. This
346. 11:45is the auto research repository from
347. 11:47Karpathy. I'm going to star it as well,
348. 11:48cuz it's a great project. Now, if you
349. 11:50are not familiar with GitHub, you don't
350. 11:52need to panic. You don't need to
351. 11:53understand much. It's just a way to
352. 11:56store code, okay? Efficient way to store
353. 11:58coding projects. That's all you need to
354. 12:00understand. You don't need to be a Git
355. 12:02expert, nothing like that. All we need
356. 12:04to do is click here on the code, and
357. 12:06click copy here. That's it. Next, you
358. 12:08need an IDE. So, either VS Code or
359. 12:10Cursor. I'm going to be using Cursor
360. 12:11here. Just install an IDE of some sort,
361. 12:14and we're going to use the coding agent.
362. 12:15I'm going to use Claude Code. So, I'm
363. 12:16going to open an empty project here.
364. 12:19Boom. Nothing in here. You can see on
365. 12:21the right, no files. And I'm going to
366. 12:23launch Claude. Dangerously skip
367. 12:26permissions. Boom. Enter. So, here it
368. 12:28is, Claude Code. Now, I'm going to say,
369. 12:31create a new folder {slash} original in
370. 12:35our repo root level, and in there, clone
371. 12:38this GitHub repository. Boom. Paste it
372. 12:41in. And again, Claude Code will know
373. 12:43what to do exactly. So, it will figure
374. 12:45out, okay, everything is empty, and then
375. 12:46it created original folder and cloned it
376. 12:48in there. Now, why did I want to do it?
377. 12:50Well, because I want to create a
378. 12:51separate folder where we're going to
379. 12:52build something with auto research for
380. 12:55us, but I also want to keep the original
381. 12:57repository here, so that we can use it
382. 12:59as a reference. Next, I'm going to say,
383. 13:01now create {slash} website folder root
384. 13:05level, and in there, build the following
385. 13:08project. XML text. Boom. Project. So, I
386. 13:12have described a clear vision for a
387. 13:13simple web app. And what I'm going to
388. 13:15show you is how you can use auto
389. 13:17research to optimize any website you
390. 13:19have, right? Whether it's a personal
391. 13:20website, whether it's your AI startup,
392. 13:22any type of website to optimize the
393. 13:24loading times or anything else that you
394. 13:26can measure, but loading times loading
395. 13:28times are very easy to measure, so
396. 13:29they're a great candidate for an auto
397. 13:31research loop. So, I'm just having
398. 13:33Claude Code with bypass all permissions
399. 13:36build this, so that we can
400. 13:38build our own auto research loop. And in
401. 13:39fact, while this is running, let me
402. 13:40launch a Codex in Yolo mode. Boom. We
403. 13:44put it here. So, I'm going to rename it,
404. 13:45so it's clear for you guys what it's
405. 13:47what. Codex CLI. GPT-4.6 is actually
406. 13:50really good at
407. 13:51fixing and debugging, but Claude Code
408. 13:54with Opus 4.6 is really incredible to
409. 13:57work with, especially with fast mode. I
410. 13:58know it costs a lot of money if you do
411. 13:59{slash} fast, but I highly recommend it,
412. 14:02and I also highly recommend it inside of
413. 14:04Codex. And keep it on high. Extra high
414. 14:06is usually overkill. So, for Codex, I'm
415. 14:08going to say, create a benchmark.mjs
416. 14:11file inside of {slash} original folder
417. 14:14where it makes sense. We're going to
418. 14:17give it the eval instructions. Hit hand
419. 14:19hit enter. Okay, actually, first I'm
420. 14:20going to say, read the structure of the
421. 14:23Actually, that's my bad. It should be in
422. 14:25the website here. Website folder, not
423. 14:26original, because it should read the
424. 14:29original It shows the original repo by
425. 14:32Karpathy to understand where the eval
426. 14:35script should be. Boom. So, this is
427. 14:38going to use Puppeteer to test the speed
428. 14:41locally of this website.
429. 14:43Um
430. 14:44it we should have the project running on
431. 14:45localhost 3000. Let me see. So,
432. 14:47basically, I asked it to create a simple
433. 14:49portfolio website with Express and
434. 14:51static files. So, let's see what it
435. 14:53built. Here's what it looks like. Alex
436. 14:56Morgan. Uh just a simple portfolio
437. 14:58website. Like, you could see this is
438. 14:59literally like 10, 15 years ago, every
439. 15:01website looked like that, right?
440. 15:03Everybody who like is new to HTML and
441. 15:05CSS would have stuff like this in high
442. 15:07school and think they're the best
443. 15:09website designer. I'm guilty of that as
444. 15:10well. But anyways, let's
445. 15:12see whether Codex finished this. Okay,
446. 15:15it did. So, we have the setup. So, now
447. 15:17let's understand the repo, right?
448. 15:19Remember, there are the three main
449. 15:20files. program.md, train.py, and
450. 15:23prepare.py. So, the original program.md
451. 15:26looks like this from Karpathy, and he
452. 15:28wrote this himself. And this is a very
453. 15:30useful instruction, by the way. Feel
454. 15:32free to steal this prompt into many of
455. 15:33your agentic projects where you want the
456. 15:36agent to keep going forever, or as close
457. 15:39to forever as possible. But we need to
458. 15:41kind of write our own. Okay, so let's go
459. 15:43back into Claude Code, and I'm going to
460. 15:44tell it what to do. To CD into the
461. 15:46website folder, and to install
462. 15:48Puppeteer, and run this baseline script.
463. 15:51We're going to [clears throat] benchmark
464. 15:52our website of our expert Alex Morgan
465. 15:55portfolio designer to see how fast or
466. 15:58slow this is, and then we're going to
467. 16:01run auto research to optimize the living
468. 16:03out of it, okay? Okay, let's see. Okay,
469. 16:05we have the results. So, Puppeteer ran,
470. 16:07and it closed the Chrome faster than I
471. 16:09could see it, or it ran in the
472. 16:11background, but basically, we have the
473. 16:12medium load time, 50 ms,
474. 16:15which is
475. 16:16not that bad. So, we'll see whether auto
476. 16:18research can can optimize this. Okay, so
477. 16:20now, the most important part of any auto
478. 16:22research loop is the program.md. So, I'm
479. 16:24going to go into our website folder,
480. 16:26create a new file, program.md, and this
481. 16:30is the main file.
482. 16:32And actually, we can use inspiration
483. 16:34from Karpathy to say like, so read the
484. 16:36program.md inside of {slash} original,
485. 16:40and then build our new {slash} website
486. 16:43{slash} program.md file on top of that,
487. 16:46but relevant to our website speed
488. 16:50benchmarking auto research objective.
489. 16:53Okay? So, we're going to borrow Andrej
490. 16:55Karpathy's from engineering that he
491. 16:57optimized quite heavily for his own
492. 16:59project, and we're going to have Claude
493. 17:01Code rewrite that into our own
494. 17:03program.md, which is going to be
495. 17:05relevant to our own project. So, again,
496. 17:06feel free to steal this for whatever you
497. 17:07want. Any business use case, any
498. 17:09marketing use case, any make money use
499. 17:12case, anything where there's a clear
500. 17:13metric, you can create your own
501. 17:15program.md, and boom, just like that, I
502. 17:17used Claude Code to write 128 lines of
503. 17:20instructions and to adapt it to this
504. 17:22project. So, now, the program.md is
505. 17:23highly relevant to this specific
506. 17:25project. Okay, so I'm going to say, do
507. 17:27this next. Give it clear instruction of
508. 17:30what I want Claude Code to do. To
509. 17:32stage this, commit this as a baseline.
510. 17:35So, remember, if an experiment is
511. 17:37successful, it commits the result into
512. 17:40history.that tweak. If it's not, it does
513. 17:42get reset and try something else. Now,
514. 17:44I'm going to say, read program.md, run
515. 17:47baseline benchmark first, record
516. 17:49results.tsv,
517. 17:50then begin the experiment loop. Do not
518. 17:52stop or ask me anything. Just keep
519. 17:53running experiments automatically. And
520. 17:55there we go. We have our first auto
521. 17:57research loop running. Maybe I could
522. 17:59have created a separate agent in a
523. 18:01separate Claude Code or Codex. And
524. 18:03obviously, you can try different agents
525. 18:04and see which, you know, which one
526. 18:06performs the best. But
527. 18:08as you can see, it's not So, look,
528. 18:10slightly worse looks like noise. Let me
529. 18:11rerun to confirm. Still worse. Following
530. 18:13protocol, revert. Okay, amazing. So, it
531. 18:15ran an experiment, and the speed of the
532. 18:16website was worse. So, it's going to try
533. 18:18something else. And this is the whole
534. 18:20point of auto research. I'm not doing
535. 18:21anything. My hands are up. And even if I
536. 18:23was doing this, first of all, I would
537. 18:24need to be a solid front-end developer.
538. 18:27And second of all, I couldn't do it so
539. 18:28quickly, right? So, even if you are a
540. 18:30great front-end developer and you know
541. 18:31how to optimize websites, you're still
542. 18:33Okay, now it found an improvement.
543. 18:35You're still not going to beat an AI
544. 18:37agent that can generate hundreds of
545. 18:38tokens per second. So, it found an
546. 18:39improvement. Look at this. 33
547. 18:41milliseconds instead of 50. So, it's
548. 18:43already down 34% in a matter of less
549. 18:45than a minute. This is the power of auto
550. 18:47research. Okay, and look at this.
551. 18:49Another one. 28 milliseconds. Another
552. 18:5115% improvement in a matter of two or
553. 18:54three minutes, guys. Yeah, auto research
554. 18:55is insane, and I have a feeling that
555. 18:56this is not the last video I'm going to
556. 18:58make on auto research. So again, if you
557. 19:00want me to make more content on auto
558. 19:02research, make sure to subscribe. And if
559. 19:04you are someone who's building his own
560. 19:05AI startup, you want to build a real AI
561. 19:07business, then listen up, because in my
562. 19:09accelerator, we work closely with a
563. 19:11handful of founders for 6 months to help
564. 19:13them scale aggressively. And in fact,
565. 19:15for the rest of March, we are offering
566. 19:16free idea validation calls. So, if you
567. 19:19have an AI app you want to build, and
568. 19:21you want to turn it into a real
569. 19:22profitable business, then make sure to
570. 19:24click the second link in the
571. 19:25description, which will take you to the
572. 19:26landing page to see if you qualify for a
573. 19:29free idea validation call. But again,
574. 19:31this is only for serious founders who
575. 19:33actually want to build a real AI
576. 19:34business. That being said, thank you
577. 19:36guys for watching. I'm going to let this
578. 19:37auto research loop run for a bit. We're
579. 19:39already on a 25 milliseconds. That's
580. 19:42crazy. Already half in a matter of 4
581. 19:44minutes. Let's see where this gets, and
582. 19:46yeah, I'm going to update you on
583. 19:48Twitter. So, make sure to follow me on
584. 19:49Twitter, and have a great productive
585. 19:51week.