The AI Takeover Has Completely Backfired — Transcript

Full transcript

1. 0:00AI was supposed to replace all software
2. 0:02engineers.
3. 0:03>> Nobody has to program.
4. 0:04>> There are no programmers in 5 years.
5. 0:06>> are going to have an AI that can
6. 0:08effectively be a sort of mid-level
7. 0:11engineer that you have at your company
8. 0:13that can write code.
9. 0:14>> We're not adding any more software
10. 0:15engineers next year.
11. 0:16>> That is why companies had multiple
12. 0:17layoffs, right? So why then are software
13. 0:20engineering job postings climbing again?
14. 0:23Why is Anthropic offering its software
15. 0:25engineers $750,000
16. 0:28a year annually? Their CEO had predicted
17. 0:30in 2025 that AI will write all code
18. 0:33within a year. When it comes to AI, the
19. 0:35numbers have never added up. Companies
20. 0:37worldwide are projected to spend $2.7
21. 0:40trillion
22. 0:41on AI in 2026 alone. And how much of
23. 0:44this money are companies actually making
24. 0:46back? 56% of CEOs reported zero benefits
25. 0:50from implementing AI in productivity and
26. 0:52revenue. And the problems are piling up.
27. 0:54Companies [snorts] are rehiring people
28. 0:56they laid off, burning through a lot of
29. 0:58revenue in terms of AI tokens, and
30. 1:01facing lawsuits. So the claim that AI
31. 1:03will replace all software engineers has
32. 1:05officially failed. In this video, we're
33. 1:07going to investigate four things: why
34. 1:09companies went through the cycle of
35. 1:10layoffs and rehiring, why CEOs keep
36. 1:13making such extreme predictions, the big
37. 1:16problem with AI-generated code, and the
38. 1:18skills you need to stand out as a
39. 1:19software engineer in this AI era.
40. 1:22Let's jump right into it. To understand
41. 1:24why companies started laying people off,
42. 1:26we need to separate two timelines that
43. 1:28get mashed up together. The rise of AI
44. 1:30coding tools and post-pandemic tech
45. 1:33correction. If we go back in time, 2018
46. 1:35was the first time AI had kind of
47. 1:38started coding. Microsoft introduced
48. 1:40IntelliCode. It used machine learning to
49. 1:42rank autocomplete suggestions based on
50. 1:45the context of your code. [music] In
51. 1:462019, DeepTabNine used GPT-2. This
52. 1:49predicted code across multiple
53. 1:51programming languages. These tools
54. 1:53helped developers type and code faster.
55. 1:56The first major shift came in June 2021.
56. 1:59This [snorts] was when GitHub launched
57. 2:01Copilot as an AI programming partner. It
58. 2:05could suggest entire lines of code and
59. 2:07functions, but a developer still had to
60. 2:10choose the problem, view and test the
61. 2:12code, and then deploy it. Then ChatGPT
62. 2:14launched in November 2022. Suddenly,
63. 2:17anybody could write code with plain
64. 2:20[music] English. But just 5 days later,
65. 2:22Stack Overflow temporarily banned
66. 2:24ChatGPT-generated coding answers.
67. 2:27Why? Because many answers looked very
68. 2:30convincing but were completely wrong. AI
69. 2:32was able to suddenly imitate the most
70. 2:34important part of programming, which was
71. 2:37just generating code,
72. 2:39but it wasn't essentially producing the
73. 2:41right code. However, this was not what
74. 2:44caused most of the layoffs in that time.
75. 2:46Tech companies had hired aggressively
76. 2:48during the pandemic. This was because
77. 2:50interest rates were low and the demand
78. 2:52of online products had exploded. So when
79. 2:55the interest rates rose again and growth
80. 2:57slowed down, companies had to cut these
81. 2:59people off that they had mass hired. In
82. 3:01America, there were 233,000
83. 3:04job positions cut off in 2023.
84. 3:07However, AI was explicitly cited for
85. 3:10only about 4,200 of them. AI was part of
86. 3:14the story but definitely not the main
87. 3:15cause behind this first wave of layoffs.
88. 3:18Then in March 2024, Cognition introduced
89. 3:21Devin, the first AI software engineer.
90. 3:24Devin successfully resolved only 13% of
91. 3:27coding issues, which was impressive but
92. 3:29nowhere near replacing a real software
93. 3:32engineer. But executives saw this and
94. 3:35immediately made this assumption that
95. 3:38because software engineers could write
96. 3:40code fast because of AI, we would need
97. 3:43lesser software engineers and lesser
98. 3:45investment. But all that AI had
99. 3:47automated was code generation, which is
100. 3:49not the only role of a software
101. 3:51engineer. AI had not automated system
102. 3:53architecture, security, debugging,
103. 3:56production, ownership, and
104. 3:57responsibility when something breaks. So
105. 4:00So was never AI replace software
106. 4:01engineers. The cycle was companies over
107. 4:04hired, then they cut costs, so they had
108. 4:06to cut people. They assumed AI would
109. 4:08close the gap, but they started rehiring
110. 4:11when unautomated work started piling up.
111. 4:13And that brings us to the next question.
112. 4:15If the evidence was never that clear,
113. 4:17why did so many CEOs make such extreme
114. 4:20predictions? Well, you got to understand
115. 4:22that for a CEO, you are not the target
116. 4:24audience. Building AI is extremely
117. 4:27expensive. As mentioned earlier, there's
118. 4:292.7 trillion dollars invested into AI
119. 4:32this year. Most of that money came from
120. 4:34investors, and that amount is so much
121. 4:37greater than the profit made by any of
122. 4:40these companies. So the investors expect
123. 4:42an explanation on how they'll get a
124. 4:45return on that money they spent. So when
125. 4:48a CEO says AI will write the code,
126. 4:50investors hear lower labor costs, high
127. 4:54profit margin, and the ability to
128. 4:56produce more code without hiring more
129. 4:58people. This prediction became a
130. 5:00financial story. This story being yes,
131. 5:03AI is expensive today, but tomorrow
132. 5:05it'll dramatically reduce costs. But if
133. 5:08you listen carefully to the language
134. 5:09used by these CEOs, they keep [music]
135. 5:11saying AI will replace writing code.
136. 5:14They never explicitly really mention
137. 5:16that AI will replace all software
138. 5:17engineers. Because that's not the same
139. 5:19thing. Writing code, again, is not the
140. 5:21only thing that a software engineer
141. 5:22does. Even if AI comes in and writes 90%
142. 5:25of the code, somebody has to come in and
143. 5:27decide what to build. They need to
144. 5:29design the system, understand business
145. 5:31requirements, review the generated code,
146. 5:34test it, make sure it's secure, and then
147. 5:35ship it. And this is not the first time
148. 5:38programming has gone through a major
149. 5:39shift like this.
150. 5:40>> [music]
151. 5:40>> In 1940s, programmers worked directly
152. 5:43with machine instructions. In 1947,
153. 5:46Kathleen Booth began developing one of
154. 5:48the first assembly languages. In 1957,
155. 5:51IBM released Fortran. This was the first
156. 5:53high-level programming language. So
157. 5:55throughout the years, software engineers
158. 5:56have moved to a higher level of
159. 5:58abstraction. AI assisted coding is just
160. 6:00the next step in that progression. But
161. 6:02here's the actual problem with AI
162. 6:04generated code. First is bloating. AI
163. 6:06tends to add a lot of unnecessary
164. 6:08functions and methods, even when a few
165. 6:11lines of code could solve the problem.
166. 6:12For example, here's a simple function
167. 6:15that I wrote. It converts a name to a
168. 6:17username. Here is the same function
169. 6:19generated by AI. Both produce the same
170. 6:21result, but the AI code is so bloated
171. 6:24and unnecessarily complicated. This also
172. 6:26creates technical debt. Teams start
173. 6:29become responsible for all of this code
174. 6:31that nobody really understands. Second
175. 6:33is duplication. Instead of finding and
176. 6:35reusing a function that already exists
177. 6:37in the code base, AI may simply create
178. 6:39another version of it. Third is costs.
179. 6:41Without token limits or guardrails, if
180. 6:43you're not careful, an AI agent can
181. 6:46easily just re-comb the entire
182. 6:48repository for a simple change. This can
183. 6:50dramatically increase costs. And the
184. 6:52last point is subtle bugs. AI can
185. 6:55introduce mistakes that might seem
186. 6:56common sense to human beings, but
187. 6:59>> [music]
188. 6:59>> they're not easy to catch in tests. For
189. 7:01example, [snorts] they might just handle
190. 7:03time in local time instead of UTC. So,
191. 7:06the bugs would look harmless in
192. 7:08isolation, or you might not even catch
193. 7:10them, but they can lead to big problems
194. 7:12in production. So, now that you know
195. 7:13that AI is not replacing developers,
196. 7:16what skills should you focus on to stand
197. 7:18out in this AI job market? First, learn
198. 7:21system design. It is no longer something
199. 7:23only senior engineers are expected to
200. 7:24understand. Juniors and even interns
201. 7:27should now know the fundamentals. If
202. 7:29you're unsure where to begin, I have a
203. 7:30video here for you to get started. It
204. 7:33covers all the basics of system design.
205. 7:35Second, become extremely good at reading
206. 7:38and debugging code. You should use AI to
207. 7:40generate code, but if somebody asks you
208. 7:42how the code works, you should be able
209. 7:44to explain it line by line. You should
210. 7:46be able to justify why all of it exists
211. 7:50and why you made all the decisions that
212. 7:52you did. Third, learn how to manage AI
213. 7:54tools properly. Give agents only the
214. 7:56context they need. Set token and
215. 7:58iteration limits. Restrict their
216. 7:59permissions and require manual approval
217. 8:01before it performs any sensitive tasks.
218. 8:04Finally, build projects that are
219. 8:05actually useful in the real world.
220. 8:08Nobody needs another to-do list app. If
221. 8:10you need ideas, I have another video for
222. 8:12you on five AI projects that you should
223. 8:14pursue to get hired. Whatever you build
224. 8:17should solve a real problem. It should
225. 8:19save somebody time, reduce a business
226. 8:21expense, or have the potential to earn
227. 8:23revenue. Because anybody can generate
228. 8:25lines of code now, it's all about
229. 8:27execution. I hope you found this video
230. 8:29helpful. Please give it a thumbs up and
231. 8:31subscribe for more.