Cómo el ingeniero de IA #1 del mundo usa la IA — Transcript

Full transcript

1. 0:01Andrej Karpathy is arguably one of the
2. 0:03best AI engineers in the world; he
3. 0:04co-founded OpenAI, led AI at Tesla,
4. 0:06coined the term "bit-flip coding," and
5. 0:08has been working at Anthropic, the
6. 0:10company behind Claude, since May.
7. 0:13In other words, he is one of the people
8. 0:14who best understands how these models
9. 0:16work on the inside. In the last two
10. 0:17weeks, he shared on his blog, X, and
11. 0:19Hacker News how he uses AI in his daily
12. 0:22life, and what he does is almost the
13. 0:24opposite of what most people do. So I’ve
14. 0:26organized it into seven principles,
15. 0:28explaining in each what he does, why it
16. 0:30works internally, and what changes
17. 0:32compared to how AI works by default. I’m
18. 0:35Fran, I’ve been a DevOps engineer for
19. 0:37over 10 years, and today I use Claude,
20. 0:38Code, and all these AI models to run my
21. 0:40daily life and my business. Well, let's
22. 0:42start with what's normal, and I didn't
23. 0:44make this up. OpenAI conducted a study
24. 0:46analyzing over a million ChatGPT
25. 0:48conversations to see how people use it
26. 0:50and discovered that nearly half of the
27. 0:52messages are questions. The person wants
28. 0:54advice or a piece of data to make a
29. 0:56decision. In programming, usage is
30. 0:58barely 4%, and when people ask it to
31. 1:00write, two out of three times it’s just
32. 1:02pasting their own text for it to be
33. 1:04corrected.
34. 1:06In other words, most people use AI as a
35. 1:07consultant; they ask it, read the answer
36. 1:09, and then do the work themselves.
37. 1:11The AI advises, and you execute.
38. 1:13Karpathy wrote on his blog that he no
39. 1:15longer does anything resembling
40. 1:17traditional engineering; he doesn't read
41. 1:19code, doesn't write code, and doesn't
42. 1:21run terminal commands—he just talks to
43. 1:23Claude. It’s the reverse of the norm: he
44. 1:25makes the AI do all the work, and he
45. 1:27only decides what gets done and checks
46. 1:29how it turned out.
47. 1:31This is one of the most important shifts
48. 1:33because it’s when you have to empower
49. 1:34yourself with judgment and direction,
50. 1:36while the AI is the execution.
51. 1:39It does everything you direct and create
52. 1:41in your mind. Why does this work? It’s a
53. 1:43simple idea. Reviewing a result is much
54. 1:46cheaper than producing it. Knowing if a
55. 1:48page loads, if a report makes sense, or
56. 1:50if a test passes takes minutes, whereas
57. 1:52doing it yourself takes hours. So, when
58. 1:54the model produces and you review, your
59. 1:55time is spent on what matters most:
60. 1:57deciding what gets done and judging if
61. 1:59it turned out well.
62. 2:00And an important nuance is that not
63. 2:02reading the code isn't the same as not
64. 2:04reviewing it, because you review the
65. 2:05result, but not line by line, and you
66. 2:07set limits on what people can touch.
67. 2:09I made a separate video about that which
68. 2:11I recommend right here. Regarding effort
69. 2:13, it’s normal to think more is better;
70. 2:15that is, if the model can think more, it
71. 2:17should always think as much as possible
72. 2:19.
73. 2:20That’s what we tend to think by default
74. 2:22, but Karpathy uses Claude Code with
75. 2:24Opus at medium effort as his default
76. 2:26setting, not maximum.
77. 2:27The concept behind it is that when the
78. 2:29model thinks, it’s generating internal
79. 2:31text before responding, and more effort
80. 2:33means more of that text. And that, uh,
81. 2:36has diminishing returns. For a difficult
82. 2:38task, thinking helps more, but for a
83. 2:40normal task, the model overthinks, takes
84. 2:42longer, spends more of your limit, and
85. 2:44sometimes complicates things for itself
86. 2:46and ends up worse. What does this change
87. 2:47for you? Effort is a dial, not a "more
88. 2:50is better" setting; use medium for
89. 2:51day-to-day by default, and high or
90. 2:53maximum only when something failed twice
91. 2:55or when it’s a big decision. If you want
92. 2:58the details on each level, I have a
93. 3:00specific video talking about the
94. 3:01different usage levels. I’ll leave it
95. 3:03here for you to check out. Another
96. 3:05important pattern that is highly
97. 3:06recommended and has been widely
98. 3:08suggested on the internet is very long
99. 3:10prompts. For example, mega-prompts are
100. 3:12the norm today.
101. 3:14For instance, "you are an expert with 20
102. 3:15years of experience, do this, don't do
103. 3:17that, and take care of these 15 points.
104. 3:18"
105. 3:19But that is now in the past. Karpathy
106. 3:21says the opposite. His prompts are
107. 3:23getting shorter and shorter, and in his
108. 3:25words, "dumber." To deploy a project, he
109. 3:27gave Claude the server address and five
110. 3:30words: update, deploy, secure, harden,
111. 3:32and optimize. Other times, he just
112. 3:34writes "check the latest on
113. 3:35cybersecurity," and Claude launches
114. 3:37several agents and solves it. Even so,
115. 3:39he clarifies that this works for him
116. 3:41with the newest models. And why does it
117. 3:42work? He explains it like this. If you
118. 3:44give it too much, you restrict the model
119. 3:46with too many requirements.
120. 3:48That is, these models today are so
121. 3:50powerful that if you put too many
122. 3:51restrictions on them, you are limiting
123. 3:53their power.
124. 3:54Instead, if you give very brief
125. 3:56directions, like this example, the model
126. 3:58will draw more potential from itself and
127. 4:01go further than it would if you
128. 4:02restricted it. Think of it this way.
129. 4:04Every requirement you add closes a door
130. 4:07. With few, the model uses everything it
131. 4:09knows to find the best solution.
132. 4:11With many, requirements get shelved and
133. 4:13it's easy for them to clash, like being
134. 4:16fast but thorough; the model tries to do
135. 4:18both and does neither well.
136. 4:20And verbs carry knowledge. I'll tell you
137. 4:23from a DevOps perspective. The term "
138. 4:24endurecer," or "harden" in English. For
139. 4:26a server, it means very specific things
140. 4:28. Access only with a key, not a password
141. 4:30.
142. 4:31Close all ports that aren't in use.
143. 4:33Install security patches. Only block
144. 4:35those trying to guess the password. A
145. 4:37single word equals a list of 20 tasks.
146. 4:40The model already knows that list; you
147. 4:41don't have to write it out. Right. So,
148. 4:43what changes for you? Write the goal,
149. 4:45not the steps. But keep in mind this
150. 4:47works because Karpathy knows how to
151. 4:49evaluate the result. So if you don't
152. 4:50know, for example, what "hardening" is,
153. 4:52you always end up with "explain what you
154. 4:54did and how I verify it." It is
155. 4:56important to always check the result and
156. 4:59ensure it is correct. Have it explain it
157. 5:01like an airplane manual. I found this
158. 5:02super interesting and I'm going to start
159. 5:04applying it daily. Normally, you ask the
160. 5:06AI something and it replies with four
161. 5:08super long, well-written, polite
162. 5:09paragraphs, but it's hard to extract the
163. 5:11main idea and we waste a lot of time
164. 5:13reading it. There is a technical
165. 5:14explanation for that.
166. 5:16These models are partly trained by
167. 5:17people choosing which response they like
168. 5:19best, and people tend to reward long
169. 5:21answers because they seem more complete
170. 5:23, but the result is that the model
171. 5:25learned to write too much, and Karpathy
172. 5:27has a trick for that: he asks it to
173. 5:28explain in ASD STE1.
174. 5:31What is that? It's the Simplified
175. 5:33Technical English used by the aerospace
176. 5:35industry for aircraft maintenance
177. 5:37manuals. It was born in the 80s because
178. 5:39mechanics all over the world needed to
179. 5:41understand exactly the same thing, with
180. 5:42no room for misunderstandings. It has
181. 5:44very strict rules. Sentences of at most
182. 5:4620 words for instructions, active voice
183. 5:49. Always state who does what, one idea
184. 5:51per sentence, and a dictionary of about
185. 5:53900 approved words. It is so strict that
186. 5:55sometimes Karpathy only asks for 80%of
187. 5:57that standard and goes even further. It
188. 5:59has a scale. A diagram is better than
189. 6:01text, an interactive webpage is better
190. 6:03than a diagram, and what excites him
191. 6:05most is asking for an explanatory video
192. 6:06. The logic is that it takes less effort
193. 6:08for the model to make the image than for
194. 6:10you to read 1,000 words. And what does
195. 6:11this change for you? The response format
196. 6:13is requested as well. You don't have to
197. 6:15accept a wall of text; the rule works
198. 6:17the same in Spanish. Ask for sentences
199. 6:19of at most 20 words, active voice, and
200. 6:21one idea per sentence. I’ve left all the
201. 6:23prompts for these recommendations in a
202. 6:25free guide linked in the description.
203. 6:27Your own tests are worth more than
204. 6:28benchmarks.
205. 6:30Usually, a new model comes out, you see
206. 6:31a chart where it wins all the tests, and
207. 6:33you think it’s the best.
208. 6:34Karpathy says he basically doesn't trust
209. 6:36benchmarks, and there are technical
210. 6:38reasons why. Many of those tests are
211. 6:40public, so their questions end up in the
212. 6:42training data. The model doesn't reason
213. 6:44; it remembers, and others are already
214. 6:46saturated. All models score nearly 100,
215. 6:48so they don't distinguish anything
216. 6:49anymore. There is a law that sums it up
217. 6:51. When a measure becomes a target, it
218. 6:53ceases to be a good measure. He
219. 6:54recommends looking at an index that
220. 6:56tracks many benchmarks and analyzes
221. 6:58which ones are worth it. But the most
222. 7:00interesting thing is that he likes weird
223. 7:02tests.
224. 7:03He shared one someone else did, for
225. 7:05example, asking the model "land or water
226. 7:06" for over 16,000 coordinates on the
227. 7:08planet.
228. 7:09You plot each answer as a dot, and the
229. 7:11world map appears without internet, just
230. 7:13using what the model has stored inside.
231. 7:15What does this change for you? Build
232. 7:17your own test. For example, gather 10
233. 7:19real tasks from your work. Ideally, the
234. 7:21times the AI failed you, and test them
235. 7:23with every new model. That score is
236. 7:25worth more than any chart. And if you
237. 7:27can turn something into a score, agents
238. 7:28can improve it. For example, this week
239. 7:31another creator did exactly that with
240. 7:33his app's tools. First, he got the test
241. 7:35to 100%, and then he cut the text in
242. 7:37half without losing the score. Software
243. 7:39is just the wrapper. What is valuable is
244. 7:41the data. Normally, you have a problem
245. 7:43and you look for an application to solve
246. 7:45it. You find one that does 70%of it, and
247. 7:47you pay the subscription. Karpathy was
248. 7:49fed up with YouTube; in his words, it
249. 7:51seemed like a wall of junk content, so
250. 7:53he made his own. I talked about it in
251. 7:55yesterday's video; it's called
252. 7:56HackerTube. It’s a YouTube with only the
253. 7:58videos that people on Hacker News have
254. 7:59recommended over and over for years.
255. 8:01Over 6,000 videos in about two hours of
256. 8:03work. The most interesting thing is what
257. 8:05he did next. He didn't publish the app;
258. 8:07he published the data in a file and
259. 8:09wrote, "Point your Claude at this and
260. 8:10ask for your own version." The concept
261. 8:12is that when building software costs
262. 8:14almost zero, the app is no longer what's
263. 8:16valuable. It is just a wrapper. What
264. 8:18matters is the data and the judgment you
265. 8:20use with it. I see this every day with
266. 8:21clients. AI can now copy almost any app
267. 8:24. What the client pays for is the result
268. 8:25based on their data. What does this
269. 8:27change for you? Before paying for
270. 8:28another subscription, ask yourself if
271. 8:30you should build it instead. And if you
272. 8:32have a business, your advantage is no
273. 8:34longer software, but the data only you
274. 8:35possess. One expensive model to direct
275. 8:37and a cheap one to work. The usual way:
276. 8:40one chat, one model for everything. And
277. 8:42if you have 500 rows of data, you paste
278. 8:44them into the chat and wait. For
279. 8:45HackerTube, Karpathy did something else
280. 8:47.
281. 8:48Opus, the large model, downloaded all
282. 8:50the Hacker News comments, filtered them
283. 8:52, and passed what remained to Sonnet, a
284. 8:53cheaper model, to process them one by
285. 8:55one.
286. 8:56Over 90%of the work was done by the
287. 8:58cheap model. The expensive one only
288. 9:00directed and built the app. There are
289. 9:01two concepts here. First, code is free
290. 9:03and accurate. Everything that can be
291. 9:05filtered with rules, filter it
292. 9:06beforehand. Without AI. AI is for what
293. 9:09requires judgment. And second, not every
294. 9:11task needs a more expensive model. The
295. 9:13big one plans and decides. The small one
296. 9:15handles the volume, so your subscription
297. 9:17covers things that were previously
298. 9:18unthinkable. What does this change for
299. 9:20you? Stop thinking about a chat, think
300. 9:22about a team: one that directs and
301. 9:23several that execute. This very video
302. 9:25started that way. Four parallel agents
303. 9:27reviewed thousands of videos to find
304. 9:28what topic to cover today. If you look
305. 9:30closely, the seven principles have
306. 9:31something in common. Karpathy doesn't
307. 9:33use any strange tricks. He says himself
308. 9:35that what he does looks so simple it
309. 9:36almost embarrasses him. Treat AI like a
310. 9:38highly capable colleague. Give it the
311. 9:40goal, not the steps. Ask it to explain
312. 9:42briefly, check the result, and don't
313. 9:44blindly trust the charts. Test it.
314. 9:46If I had to choose one to start with
315. 9:48today, it's number three: short prompts
316. 9:50with the goal. That changes everything
317. 9:52else. Tell me in the comments which of
318. 9:53the seven you already do and which one
319. 9:54is the hardest for you. And if you want
320. 9:56to understand principle two in depth,
321. 9:58the video on effort is right here. Thank
322. 10:00you very much for following along with
323. 10:01this video.