If You're in Computer Science, Watch This Now — Transcript

Full transcript

1. 0:00If you're in computer science, watch
2. 0:01this now. If your plan is just to go to
3. 0:03lectures, get a good GPA, and wait for a
4. 0:05six-figure drop at graduation, you are
5. 0:07going to get completely cooked.
6. 0:08>> Wait, what?
7. 0:09>> Universities are still teaching like
8. 0:10it's 2015. If you don't lock in right
9. 0:12now, you're spending four years and tens
10. 0:14of thousands of dollars just to graduate
11. 0:16with the exact same skill set as a free
12. 0:17chatbot.
13. 0:18>> I turned it down just to do this.
14. 0:20>> Here is why the traditional CS degree is
15. 0:22failing students right now and how to
16. 0:23actually get cracked at programming. The
17. 0:25biggest lie in university computer
18. 0:27science is the auto-grader bubble. Think
19. 0:29about how a normal college programming
20. 0:31assignment works. The professor hands
21. 0:32you a pre-built skeleton repository
22. 0:34where the architecture is already
23. 0:35designed, the built system is already
24. 0:37configured, and 90% of the boilerplate
25. 0:39is already written. Your entire job is
26. 0:41to fill in a 50-line function and submit
27. 0:43it to GradeScope at 11:58 p.m. until the
28. 0:45unit tests turn green.
29. 0:46>> Green giant.
30. 0:48>> And the second the semester ends, you
31. 0:49throw that code in the trash and never
32. 0:51look at it again. Stanford economist
33. 0:52Eric Brynjolfsson pointed out the exact
34. 0:54problem with this. Clean, isolated
35. 0:56textbook coding problems are the exact
36. 0:58kind of book learning that large
37. 1:00language models have completely
38. 1:01memorized. If your homework assignment
39. 1:03can be graded by a simple input-output
40. 1:05script, an AI can already one-shot it.
41. 1:07Even worse, many university classes
42. 1:10teach computer science in a vacuum that
43. 1:11ignores physical hardware. In data
44. 1:13structures and algorithms, you're taught
45. 1:15big O notation where traversing a linked
46. 1:17list of a million items and traversing a
47. 1:19flat array of a million items are both
48. 1:21considered linear time or big O of n. So
49. 1:23on a chalkboard, they look identical.
50. 1:25Try doing that on a real CPU. A flat
51. 1:27array stores elements side by side in
52. 1:29contiguous RAM, so when the processor
53. 1:31fetches 164-byte cache line, it pulls
54. 1:33the next several items into its fast L1
55. 1:36cache for free. A linked list scatters
56. 1:38nodes randomly across the heap, forcing
57. 1:40the CPU to follow a pointer and stall
58. 1:42for hundreds of clock cycles on almost
59. 1:44every single hop. On paper, they have
60. 1:45the same time complexity. In production,
61. 1:47one is 20 times slower than the other.
62. 1:49The second problem is that universities
63. 1:51teach you theoretical computer science,
64. 1:53but almost never teach you how to
65. 1:54actually operate a computer. Even MIT
66. 1:56admitted this. Three instructors at MIT
67. 1:58noticed that brilliant who could prove
68. 2:00compiler theorems and build neural
69. 2:02networks on a whiteboard had almost zero
70. 2:04proficiency with the actual tools of
71. 2:06software engineering. They saw MIT
72. 2:08students holding down the arrow key for
73. 2:1030 seconds to scroll to the bottom of a
74. 2:11file, manually renaming variables across
75. 2:1420 files by hand, and fixing a Git merge
76. 2:16conflict by copying their code to the
77. 2:18desktop, deleting the entire repository,
78. 2:20and cloning it all over again. MIT
79. 2:22literally had to create a special class
80. 2:24called the Missing Semester of Your CS
81. 2:26Education just to fill this gap. In a
82. 2:28four-year degree, almost nobody sits you
83. 2:30down and teaches you how to pipe data
84. 2:31streams through the Unix shell, how to
85. 2:33attach an interactive debugger, to
86. 2:35inspect corrupted stack frames, or how
87. 2:37to read a CPU flame graph to find a
88. 2:39performance bottleneck. You won't find
89. 2:41classes on how to untangle a messy Git
90. 2:43merge conflict, resolve a cascading
91. 2:45database deadlock in production, write a
92. 2:47clean Docker file, or navigate a
93. 2:49million-line legacy codebase without
94. 2:50documentation. Nobody teaches you that
95. 2:52Git isn't a mysterious cloud backup
96. 2:55button. It's a content-addressed,
97. 2:57directed acyclic graph where every
98. 2:58commit is an immutable snapshot pointing
99. 3:00to a tree of hashed binary blobs. Once
100. 3:03you understand the graph, you can
101. 3:04rebase, bisect a regression across 500
102. 3:07commits in 30 seconds, and cherry-pick
103. 3:09changes without panicking. Until you
104. 3:11learn your tools, you're trying to build
105. 3:13software with one hand tied behind your
106. 3:15back. Let's take a quick break to talk
107. 3:16about today's sponsor, Ingest. When
108. 3:19you're building AI agents or workflows
109. 3:21that run for minutes, hours, or weeks, a
110. 3:23single HTTP request won't cut it. APIs
111. 3:26time out and deploys restart your
112. 3:27servers mid-run. Normally, fixing that
113. 3:29means babysitting worker fleets, queues,
114. 3:32and hand-rolled idempotency. Ingest puts
115. 3:34durable execution right inside your
116. 3:36TypeScript, Python, or Go codebase. You
117. 3:39break your workflow into steps, and
118. 3:40Ingest saves state as each one finishes.
119. 3:43If a call fails or you deploy mid-run,
120. 3:45your agent doesn't lose progress or
121. 3:47waste tokens starting over. [music] It
122. 3:49just retries the failed step and keeps
123. 3:51going. You can pause runs for days
124. 3:53waiting on human approval at zero
125. 3:54compute and cancel them automatically
126. 3:56when a user replies. You also get one
127. 3:59line per tenant concurrency and rate
128. 4:00limits to protect your model quotas. And
129. 4:02since execution and observability share
130. 4:04the same layer, tracing, replay, live
131. 4:07experiments, and outcome-based evals
132. 4:09come built-in without a second vendor.
133. 4:11Hit the link in the description to build
134. 4:12for the long run with Ingest. Thank you
135. 4:14to Ingest for sponsoring. Now back to
136. 4:16the video. Once CS students realize
137. 4:18their classes aren't enough, they
138. 4:20usually fall into the next trap, the
139. 4:22clone resume. Open a hundred new grad
140. 4:24resumes right now, and 90 of them have
141. 4:26the exact same three projects, a full
142. 4:28stack to-do app, a React movie
143. 4:30dashboard, and a wrapper around the
144. 4:32OpenAI API with a Tailwind landing page.
145. 4:35Hiring managers skip right past these
146. 4:37because when you build a basic web
147. 4:38wrapper, 90% of the actual computer
148. 4:40science is hidden inside third-party
149. 4:42packages. Calling an API and rendering a
150. 4:44JSON array doesn't prove you can
151. 4:46engineer systems. It just proves you can
152. 4:48read a quick start guide. If you want
153. 4:50your resume to stop a senior engineer in
154. 4:52their tracks, delete the five tutorial
155. 4:54clones and build one infrastructure tool
156. 4:55from scratch. Instead of importing a web
157. 4:58framework, open a raw TCP socket in C,
158. 5:00Rust, or Go and write your own
159. 5:02multi-threaded HTTP server. Parse the
160. 5:04raw byte stream yourself, handle
161. 5:06keep-alive connections, and implement
162. 5:08gzip compression from the specification.
163. 5:10Or build a toy in-memory Redis clone
164. 5:13that parses the Redis wire protocol over
165. 5:15a non-blocking event loop and handles
166. 5:17concurrent client connections and key
167. 5:18expirations. Or write your own Unix
168. 5:20shell that parses command tokens, forks,
169. 5:22child processes, and wires up operating
170. 5:24system file descriptors for input and
171. 5:26output redirection. When you build the
172. 5:28infrastructure layer yourself, you
173. 5:29aren't competing with people who prompt
174. 5:31engineered a weekend web app anymore.
175. 5:33You walk into interviews able to talk
176. 5:35about socket buffers, race conditions,
177. 5:37and memory allocation like someone who
178. 5:39has already been working in the industry
179. 5:40for years. So if you're in school right
180. 5:42now, here's how you actually lock in and
181. 5:45play your cards to win. First, stop
182. 5:47optimizing your schedule for an easy 4.0
183. 5:49GPA. Nobody at a tech company is going
184. 5:51to ask you what you got in your easy
185. 5:52elective classes. Take the brutal
186. 5:54upper-level systems courses that scare
187. 5:56everyone else away, operating systems,
188. 5:59compilers, computer networks, and
189. 6:00database internals. The classes where
190. 6:02you lose sleep writing a virtual memory
191. 6:04page table manager, an instruction
192. 6:06parser, or a B+ tree storage engine are
193. 6:09the exact classes that build the mental
194. 6:11model AI can't replace. Second, break
195. 6:13out of the classroom early. A single
196. 6:15summer internship, a merged pull request
197. 6:17to a real open source systems project,
198. 6:20or building software that real students
199. 6:21on your campus actually use is worth a
200. 6:23million times more than 50 isolated
201. 6:25homework assignments because real
202. 6:27engineering only starts when
203. 6:28requirements change, users send
204. 6:30malformed data or try to DDoS your
205. 6:32service, and you have to maintain code
206. 6:34you wrote 6 months ago. A computer
207. 6:35science degree is not a golden ticket
208. 6:37anymore. It's like a gym membership. If
209. 6:39you just pay tuition, copy paste your
210. 6:41way through the autograder, abuse
211. 6:42ChatGPT for 4 years and do the bare
212. 6:44minimum, you'll graduate with a piece of
213. 6:46paper and zero leverage. But if you use
214. 6:48these 4 years to master your tools, take
215. 6:50the hardest systems classes in the
216. 6:52catalog, get internships at top tech
217. 6:54companies, and build real infrastructure
218. 6:56from the ground up, there has never been
219. 6:58an easier time to lap the entire field.