AI Replacing Software Engineers Has Official Failed — Transcript

Full transcript

1. 0:00For about 3 years, we were all told the
2. 0:02same story. AI is going to replace
3. 0:04knowledge workers starting with software
4. 0:06engineers. Companies will run leaner,
5. 0:08save billions, and the people who don't
6. 0:10adapt get left behind. And the loudest
7. 0:13version of that came from the CEO of an
8. 0:16AI lab who warned that AI could wipe out
9. 0:18half of all entry-level white collar
10. 0:21jobs and push unemployment to somewhere
11. 0:24between 10% and 20%. Here's what
12. 0:26happened next. That same company's head
13. 0:29of economics published an analysis of 18
14. 0:32months of their own labor market
15. 0:34research and his conclusion was that
16. 0:37they don't see a significant impact of
17. 0:39AI on the US labor market. Not yet
18. 0:41anyway. And he's not the only one
19. 0:43backing off. Reporting earlier this year
20. 0:45covered both him and the CEO of OpenAI
21. 0:48softening the jobs apocalypse prediction
22. 0:52they would spend a year making.
23. 0:54Meanwhile, on the ground, companies that
24. 0:56fired people because of AI have started
25. 1:00quietly hiring them back. So, in this
26. 1:02video, I want to walk you through the
27. 1:04whole thing properly. how we all got
28. 1:06convinced AI was going to delete these
29. 1:09jobs, what the actual lie was underneath
30. 1:12the layoffs, the three specific reasons
31. 1:14the replacements did not work, why the
32. 1:16rehiring is happening so quietly, and
33. 1:19then the part nobody wants to say out
34. 1:21loud, which is who genuinely did get
35. 1:24hurt by all of this because it wasn't
36. 1:26the people you would think. To
37. 1:28understand why this fell apart, you have
38. 1:30to go back to before chat GPT even
39. 1:33existed. About 5 years ago, GitHub
40. 1:35Copilot launched and it was not an AI
41. 1:38engineer. All it did was finish your
42. 1:40function while you type like
43. 1:42autocomplete for code. But that was
44. 1:44enough. Articles started running about
45. 1:46whether AI would replace developers,
46. 1:48whether this was an existential threat
47. 1:51to the profession. And I want you to
48. 1:53notice the timing there because it
49. 1:55matters for everything that comes after.
50. 1:57That panic started before a single major
51. 2:01tech layoff had happened. The fear
52. 2:03arrived first. The evidence came later
53. 2:05if at all. Then chart GPT launched and
54. 2:07the conversation left the developer
55. 2:09bubble entirely. Now anybody could open
56. 2:12a chat window, type in plain English and
57. 2:15get working code back. And at almost
58. 2:17exactly the same moment, the tech job
59. 2:19market fell off a cliff. More than
60. 2:2293,000 tech jobs were cut that year. The
61. 2:24year after that, it was over 260,000.
62. 2:27Hiring freezes became standard. New grad
63. 2:30roles just disappeared. Now, here is the
64. 2:33thing that got lost, and it's [music]
65. 2:34the single most important detail in this
66. 2:37entire story. That crash had almost
67. 2:40nothing to do with AI. Companies had
68. 2:42massively overhired during the pandemic.
69. 2:45Interest rates went up, cheap money went
70. 2:47away, and the correction came due. AI
71. 2:50was not good enough to replace anybody
72. 2:52yet. The tech market was already broken
73. 2:55before AI ever walked into the room. But
74. 2:57it walked in at the perfect moment. Two
75. 3:00years ago, a startup released a product
76. 3:02called Devon and called it the world's
77. 3:05first AI software engineer. The demo
78. 3:07showed it planning a project, writing
79. 3:09the code, and fixing its own errors.
80. 3:12Around the same time, Google said
81. 3:13roughly 25% of his code was AI, and
82. 3:16Microsoft put its number around 30%. and
83. 3:19executives did a very simple piece of
84. 3:21math. An AI subscription cost $20 a
85. 3:24month. A software engineer cost $200,000
86. 3:26a year. If that tool gets even 40%
87. 3:28better over the next 18 months, why
88. 3:30would you keep the engineer? They did
89. 3:32not need AI to be correct. They just
90. 3:34needed to believe it would improve fast
91. 3:36enough to justify cutting now. So last
92. 3:39year, they started saying it out loud.
93. 3:42Meta CEO said AI would function as a
94. 3:44mid-level engineer and write a lot of
95. 3:46the code in their applications.
96. 3:48Salesforce CEO said they were not hiring
97. 3:50any more software engineers that year
98. 3:52because of the productivity boost from
99. 3:54their own AI agent. And by March of this
100. 3:57year, AI became the single most cited
101. 3:59reason for job cuts in the United States
102. 4:02for the first time ever. That's how you
103. 4:04get from an autocomplete tool to
104. 4:06hundreds of thousands of people out of
105. 4:08work in 5 years. Except here is the part
106. 4:11that reframes all of it. Deutsch Bank
107. 4:13analysts put a name on what was actually
108. 4:16happening. They warned that a defining
109. 4:18feature of this year would be companies
110. 4:20blaming AI for redundancies that were
111. 4:23being driven by completely different
112. 4:25things. They called it AI redundancy
113. 4:28washing. The CEO of OpenAI of all people
114. 4:31admitted the same thing that there is
115. 4:33some AI washing going on where companies
116. 4:35blame AI for layoffs they were going to
117. 4:38do anyway. [music] Think about why
118. 4:40that's such an attractive move for an
119. 4:42executive. If you announce layoffs
120. 4:44because you overhired and the business
121. 4:46is soft, your stock goes down and you
122. 4:48look like you cannot run a company. If
123. 4:50you announce the exact same layoffs and
124. 4:52say it's because you are becoming an AI
125. 4:55first organization, your stock goes up
126. 4:57and you look like a visionary. Same
127. 5:00people losing the same jobs, completely
128. 5:02different press release and investors
129. 5:05rewarded the second version every single
130. 5:07time. So the number that everyone quoted
131. 5:09as proof of AI job apocalypse, the count
132. 5:12of layoffs attributed to AI was never a
133. 5:15measurement of what AI could do. It was
134. 5:18a measurement of what companies wanted
135. 5:20to tell shareholders. That's the lie.
136. 5:22Not that AI is useless, but that the
137. 5:24layoff numbers were ever evidence of
138. 5:27AI's capability in the first [music]
139. 5:28place. Quick thing before we keep going.
140. 5:30Drop a comment below whether your
141. 5:33company has actually replaced anyone
142. 5:35with AI or whether it just started
143. 5:38saying the word AI a lot in the
144. 5:40meetings. I read every comment on this
145. 5:42and the gap between those two answers is
146. 5:44usually the whole story. All right. So
147. 5:46the cover story is one thing but plenty
148. 5:49of companies genuinely tried to do the
149. 5:51replacement. So why didn't it hold?
150. 5:53Three reasons. The first one is the
151. 5:55productivity gap. There is a study from
152. 5:57a research group called MER matter that
153. 6:01got passed around everywhere. They took
154. 6:03[music] 16 experienced open-source
155. 6:05developers, gave them 246 real tasks in
156. 6:08code bases they knew well and randomly
157. 6:11allowed or blocked AI tools. The
158. 6:13developers predicted AI would make them
159. 6:1524% faster. It made them 19% slower. And
160. 6:19the most uncomfortable part was not the
161. 6:21slowdown. It was that afterward those
162. 6:24developers still believed AI had sped
163. 6:27them up by 20%. They were wrong about
164. 6:29their own experience by nearly 40
165. 6:32percentage points. Now I want to be
166. 6:34straight with you here because a lot of
167. 6:36videos scored that number and stopped.
168. 6:38Matter went back and revised their own
169. 6:40methodology earlier this year. They
170. 6:42found that developers who benefit most
171. 6:44from AI were declining to participate in
172. 6:46the no AI condition which skewed the
173. 6:49sample. [music] Their newer cohort
174. 6:51showed roughly a 4% slowdown instead of
175. 6:5419% and the updated position is that AI
176. 6:57probably does provide some productivity
177. 6:59benefit. Now maybe the honest version is
178. 7:01not AI makes you slower. The honest
179. 7:03version is that the perception gap is
180. 7:06real and enormous. The measured gains
181. 7:09are far smaller than anyone claimed and
182. 7:12organizations were making $200,000
183. 7:15staffing decisions based on a feeling of
184. 7:18speed that nobody had actually put a
185. 7:20clock on. The second reason is the day
186. 7:23two problem. AI is genuinely excellent
187. 7:25when a project is brand new and the goal
188. 7:28is simple. Boiler plate tests,
189. 7:30prototypes, a first draft in seconds.
190. 7:33That part is real. But most software
191. 7:35doesn't stay new. It gets updated,
192. 7:37connects to other systems and inherited
193. 7:39by people who did not write it. Git
194. 7:41clear analyzed large volumes of code
195. 7:43changes and found that duplicated code
196. 7:46climbed sharply as AI [music] tools
197. 7:48spread while refactoring dropped off.
198. 7:50Teams were producing more code and more
199. 7:52of it had to be cleaned up shortly
200. 7:54after. And in Stack Overflow's developer
201. 7:57survey, the most common complaint about
202. 7:59AI output was that it was almost right
203. 8:02but not quite. Almost right is worse
204. 8:05than obviously broken. When something
205. 8:07fails immediately, you catch it. When it
206. 8:09works 95% of the time, the mistake sits
207. 8:12quietly in production until a customer
208. 8:15loses money or data gets exposed. So, a
209. 8:17company generates a dashboard with AI,
210. 8:19everything looks great and then 6 months
211. 8:22later, the billing system changes and
212. 8:24half the numbers are wrong. Nobody
213. 8:26understands how the data flows through
214. 8:28it. Every AI suggested fix breaks
215. 8:31something [music] else. And now you are
216. 8:32hiring an experienced engineer to
217. 8:35untangle a codebase that no human ever
218. 8:38fully understood. The cheap prototype
219. 8:41turned into an expensive cleanup job.
220. 8:43The third reason is the one that
221. 8:45surprised everybody and that's cost.
222. 8:48Uber rolled out an AI coding tool across
223. 8:50roughly 5,000 engineers. Adoption went
224. 8:52from about 32% to 84% within a couple of
225. 8:56months. Around 70% of committed code was
226. 8:59coming from those tools. total success
227. 9:01by every adoption metric, right? And by
228. 9:03April, the company's allocated AI spend
229. 9:06for the entire year was gone. Here is
230. 9:09the mechanics underneath that. And it's
231. 9:11[music] counterintuitive. The blended
232. 9:13cost of AI dropped about 67% in 12
233. 9:16months from around $18.40 per million
234. 9:18tokens down to $6.70. Prices fell by
235. 9:212/3. And over that same stretch, more
236. 9:24than 73% of enterprises blew past their
237. 9:28original AI cost projections. Anyway,
238. 9:31the reason is agents. A chat message is
239. 9:33one round trip. You ask, it answers, the
240. 9:36meter clicks once. An agent is a loop.
241. 9:39It reads the codebase, drafts a [music]
242. 9:41plan, makes an edit, runs a test, fails,
243. 9:44rereads everything, and tries again.
244. 9:46estimates put a single agentic task
245. 9:48somewhere between five and 30 times the
246. 9:51tokens of a comparable chat exchange. So
247. 9:53you have the price per unit collapsing
248. 9:56and the bill going up at the same time.
249. 9:58That's not a technology failure. That's
250. 10:00the company discovering it swapped a
251. 10:02predictable $200,000 salary for an
252. 10:05unpredictable and uncapped variable
253. 10:07cost. If you want to check any of these
254. 10:09numbers yourself rather than taking my
255. 10:11word for it, check the description and
256. 10:13the pinned comment below for helpful
257. 10:15resources on this topic. The underlying
258. 10:17studies and reports are all there. So
259. 10:20you can read the primary data if you're
260. 10:22interested. Which brings us to the
261. 10:24reversal. The most famous example is the
262. 10:27fintech company whose CEO went on stage
263. 10:30and said their AI assistant was doing
264. 10:32their work of 700 customer service
265. 10:34agents. It became the single most cited
266. 10:37proof that replacement had arrived. Then
267. 10:40customer complaints started coming in.
268. 10:42Service quality dropped and that same
269. 10:44CEO eventually admitted the company had
270. 10:47gone too far and started hiring humans
271. 10:50back. And that turned into a pattern the
272. 10:53industry now calls the AI boomerang.
273. 10:56Forester found 55% of employers regret
274. 10:59the AIdriven cuts they made. Gartner is
275. 11:01forecasting that by next year half of
276. 11:04the companies that reduced [music]
277. 11:05headcount citing AI will be hiring
278. 11:08people back for similar work usually
279. 11:10under a different job title. IBM's HR
280. 11:12assistant handles 94% of routine
281. 11:15requests and the remaining 6% the ones
282. 11:18requiring actual judgment still needs a
283. 11:20person which is part of why IBM said it
284. 11:23plans to triple entry-level hiring
285. 11:26across [music] his US business. There is
286. 11:28even a rough shape to how it goes.
287. 11:30announce AI will do the job. Cut the
288. 11:32team 6 to 12 months pass. AI handles
289. 11:35about 60% of the work and is completely
290. 11:37stuck on the other 40%. Rehire often at
291. 11:4020 to 35% higher pay because now you are
292. 11:43paying to get back the exact
293. 11:45institutional knowledge you threw away.
294. 11:48And here is the part that matters. If
295. 11:50you are job hunting right now, almost
296. 11:52none of this is being announced. No
297. 11:54company is going to publish a press
298. 11:56release saying they overestimated AI and
299. 12:00want their people back. That makes them
300. 12:02look weak and investors [music] punish
301. 12:04it. So the rehiring runs through the
302. 12:06recruiter outreach, alumni networks,
303. 12:09referrals from former managers and newly
304. 12:11created roles slotted into different
305. 12:14teams. It is a real hiring wave that is
306. 12:16almost invisible from the outside.
307. 12:18[music] Now, I want to do the honest
308. 12:21part because if you walk away thinking
309. 12:22the job story was pure hype, you'll be
310. 12:25just as wrong as the people who bought
311. 12:27the apocalypse. On the surface, the
312. 12:29recovery is real. IT and computer
313. 12:32science postings jumped 14.2%
314. 12:35year-over-year in April, the biggest
315. 12:38annual gain in about 4 years. [music] IT
316. 12:40unemployment fell to 2.9% in June, below
317. 12:43the national rate. Open tech roles held
318. 12:45about 600,000. But look at where these
319. 12:47jobs went. Entrylevel roles slipped from
320. 12:508.1% of the IT job mix to 7.4%. Senior
321. 12:54roles climbed from 38.8% to 43.1%.
322. 12:57Computer science graduates are sitting
323. 12:59at a 6.1% unemployment rate and computer
324. 13:02engineering grads are at 7.5%. Higher
325. 13:06than the national average of all workers
326. 13:08and higher than a lot of majors people
327. 13:11love to make fun of. New grad hiring at
328. 13:14big tech is down more than 50% from
329. 13:16where it was 7 years ago. So the
330. 13:18bloodbath did not hit the workforce. It
331. 13:21hit the on-ramp. Companies are hiring
332. 13:23engineers again. They want experienced
333. 13:26ones who can supervise AI on day one and
334. 13:29they are much less willing to pay
335. 13:31someone to learn on the job when a model
336. 13:34can produce the junior level output for
337. 13:36a few dollars.
338. 13:37>> [music]
339. 13:37>> The work that used to justify a first
340. 13:40job is exactly the work that got
341. 13:42automated. That is genuinely a serious
342. 13:44problem and it's a slower and less
343. 13:46dramatic one than AI took all the jobs
344. 13:49which is probably why it got so much
345. 13:51less coverage. If the entry funnel stays
346. 13:53broken for another few years, the senior
347. 13:56engineers everybody is fighting over
348. 13:58right now simply don't exist in 5 years.
349. 14:00That's the actual crisis and it's the
350. 14:02one IBM was pointing at when they said
351. 14:05the pipeline dries up. So what do you
352. 14:08actually do with this? Three things.
353. 14:09[music] One, get positioned on the
354. 14:11verification side of the work. The
355. 14:13bottleneck moved. Generating code is
356. 14:15cheap. Knowing whether the generated
357. 14:17code is correct, secure, scalable, and
358. 14:20maintainable is what the companies are
359. 14:22paying for. Practically that means you
360. 14:24should be able to read a diff you did
361. 14:27not write and explain what breaks. Take
362. 14:29something an AI built for you,
363. 14:31deliberately try to break it, and get
364. 14:33good at articulating why it failed. That
365. 14:36skill is worth more right now than being
366. 14:38able to produce another 500 line.
367. 14:40[music] Number two, learn the cost side
368. 14:43because almost nobody has model routing,
369. 14:45prompt caching, context discipline,
370. 14:48deciding when a task doesn't justify an
371. 14:50agent at all. Companies are literally
372. 14:52posting roles for this now with terms
373. 14:55like token economics and per request
374. 14:57unit cost in the job description. It's
375. 14:59learnable in a weekend if you already
376. 15:01have a software background. and it maps
377. 15:03directly onto the problem every
378. 15:05engineering or is currently panicking
379. 15:07about. Number three, build the
380. 15:09relationship channel [music] because
381. 15:10that's where the quiet rehiring is
382. 15:13actually happening. Applications are
383. 15:15competing against thousands of others. A
384. 15:17message from a former manager competes
385. 15:20against nobody. Stay in touch with
386. 15:22people you have worked with and do it
387. 15:24before you need something from them.
388. 15:26Talk about what you are building,
389. 15:28[music] not what you need. That's your
390. 15:30actual carrier insurance in a market
391. 15:32where the best openings never get
392. 15:34[music] posted. And if you know somebody
393. 15:36who is making a big career decision
394. 15:38right now based on the AI replacement
395. 15:41narrative, [music] send them this video.
396. 15:42They will probably thank you later.