Should AI Engineers Still Read Code in 2026? The Z/L Continuum — Alex Volkov, ThursdAI

AI Engineer · 21 min · 353 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

1. 00:01[music]
2. 00:12Two talks at the AI Engineer Europe.
3. 00:15One guy is saying, "Code is free."
4. 00:17and deleted his ID.
5. 00:19And the other one is saying, "Read every effing line of code."
6. 00:24So, should AI engineers still read code their agents output I named this the Desolation of Apollo Continuum.
7. 00:34And you guys probably have argued about this in Slack.
8. 00:37You probably talked about this in the hallway track.
9. 00:39So, let's talk about this here.
10. 00:40Because code got cheap, attention didn't.
11. 00:46As you may know, back in December 2025, something big changed.
12. 00:52AI engineering has changed forever, and it broke its own trendline.
13. 00:56Actually, Swyx, the organizer of AI Engineer, is collecting evidence to that single moment in time at the website called wtfhappened2025.com.
14. 01:05I recommend you go and check it out.
15. 01:07It's really, really funny.
16. 01:08Uh this is just one example from METR, the machine evaluation center.
17. 01:13And it shows that models, for the first time, started completing tasks that would take engineers over 16 hours to do.
18. 01:21And in fact, we've gone way up the curve, way up the trendline after that.
19. 01:26This is the backdrop to everything that AI engineering is experiencing.
20. 01:31Because we don't write code anymore.
21. 01:34Most of us, at least.
22. 01:36I want to see one Can you guys give me a raise of hands if you still handcraft and write code, most of your code?
23. 01:41Anybody here?
24. 01:42Most of your code is written by hand?
25. 01:44Amazing.
26. 01:45This is the Token Maxing track, after all.
27. 01:48I I think the one person here who still writes code is maybe a little shy of raising their hand.
28. 01:52That's okay.
29. 01:52Cuz we don't type code anymore.
30. 01:53We're not handcrafters.
31. 01:55We supervise.
32. 01:56I like to say we babysit agents.
33. 01:59And the greatest example for this obviously is Boris's journey.
34. 02:03You guys know Boris, the creator of Cloud Code, uh Anthropic.
35. 02:06100% of his code is written and authored by Cloud Code at this point.
36. 02:10And he didn't stop being engineer, he moved up the layer.
37. 02:14He still ships 20 to 30 PRs, maybe more.
38. 02:18And recently he talked about he deleted his ID.
39. 02:21I found it really funny.
40. 02:22Just just no reason to just hand type code anymore.
41. 02:26In fact, 80% of Anthropic's code is now AI written and this is this stat is at least a few months old.
42. 02:32It's likely more right now.
43. 02:36And he's not the only one.
44. 02:37Some of you have seen this chart from GitHub.
45. 02:41Some of you have maybe remembered this chart while GitHub was down for you.
46. 02:44Uh the reason is GitHub is on track to to to get 14 billion commits this year.
47. 02:51All of 2025 all of yes we yet last year was 1 billion.
48. 02:55They're 14x the number of commits.
49. 02:58They're seeing 14x the number of commits, which is insane and most of this is AI assisted.
50. 03:03And it's a lot of code.
51. 03:06And so the AI engineering has changed forever and I want to tell you about AI Engineer World's Fair.
52. 03:12I been to every single one and I'll tell you about this later.
53. 03:15And AI Engineer is a great place to get the zeitgeist of where our career is going and how is it changing, okay?
54. 03:22This one obviously is 3x bigger than last year.
55. 03:25This is just one of the rooms.
56. 03:26There's like a bunch of rooms.
57. 03:277,000 people I think we clocked in.
58. 03:2936 tracks.
59. 03:31And if you want to know what happens in AI engineering, you kind of have to be here.
60. 03:35So this would be a little bit of meta talk.
61. 03:37So one of the guys at AI Engineer EU talked about code as ship.
62. 03:42The other one talked about uh we should read every line of code.
63. 03:46Let's listen to them for just a second, okay?
64. 03:48This is Ryan LePopulaire from OpenAI.
65. 03:50I I think he's he made it here, but uh this is Ryan LeFebvre from Open AI.
66. 03:54The models at this point are good enough where they're isomorphic they can write code.
67. 03:58Can you guys
68. 03:59Do you mind?
69. 04:00These tools are high quality that solve real user problems in real code bases.
70. 04:07Code is free.
71. 04:08It's free for you to produce, free to refactor, and it is not a thing to get hung up on anymore.
72. 04:17Humans no longer need to concern themselves with implementation.
73. 04:20The important thing is not the code, but the prompt and the guardrails that got you there.
74. 04:25You can just simply say, "Do not produce slop."
75. 04:28Don't accept slop.
76. 04:29You won't get slop in your code base.
77. 04:30But, to do that requires taking short-term velocity hits in order to back up or double-click into a task to figure out what it is the agents are struggling with.
78. 04:41So, this is Ryan LeFebvre, okay?
79. 04:43He came up on stage at the engineer, and he opened with like, "Hey, I'm a talking billionaire, and I want you to be as well."
80. 04:50In fact, the talking billionaire lounge that's in front of the leadership track that you guys see, that's because of him.
81. 04:54He came up with this concept.
82. 04:55And he got the golden card and everything.
83. 04:57On the other side, the same conference, the other side, Mario Zechner, creator of Pi.
84. 05:02Slow the down.
85. 05:04Everything's broken.
86. 05:07And then there's people that say, "Our product's been 100% built by agents."
87. 05:11Yes, we know it sucks now.
88. 05:12Congratulations.
89. 05:15[applause]
90. 05:21Agents are actually compounding booboos, which is my word for errors, with zero learning and no bottlenecks and delayed pain.
91. 05:28The delayed pain is for you.
92. 05:30Those are my most beloved people.
93. 05:32I don't even read the code anymore.
94. 05:34Congratulations.
95. 05:35Something is broken, and your users are screaming, so who you going to call?
96. 05:39Not yourself, because you haven't read the code.
97. 05:41Non-critical code, sure, write slop ahead.
98. 05:44Critical code, read every line.
99. 05:46So, two folks, same conference, day after day, talking about the one anxiety that we all feel.
100. 05:52Should we all still be reading code in 2026?
101. 05:54By the way, these two folks are the number six and number seven most watched YouTube videos from AI engineer from all time.
102. 06:00So, they're obviously representing something that we're feeling we're talking about.
103. 06:04And this is being the leadership track, something that you folks that report to you are talking about, okay?
104. 06:08Should they still be reading code?
105. 06:09And what's the what's the level quality?
106. 06:11So, they name the same anxiety from both ends.
107. 06:13Uh at this point, I probably should introduce Uh hey, I'm Alex Volkov.
108. 06:17I'm the host of Thursday AI podcast.
109. 06:19It's a podcast and a newsletter.
110. 06:20We go live every week to talk about AI.
111. 06:22For the past 3 and 1/2 years, we've been tracking every change in the AI engineering, every release from every lab, every model.
112. 06:28And I'm also an AI evangelist with Weights & Biases and Core Weave.
113. 06:32Um what also should I tell you about myself?
114. 06:34That I've been covering AI engineers specifically since the first one in 2023.
115. 06:38And oh boy, has it changed.
116. 06:40And so, you can treat this as a dispatch from the front line because all of these people now are my friends.
117. 06:44And we constantly talk about this in the speakers' room in the hallway track.
118. 06:50I couldn't stop thinking about that tension.
119. 06:52I couldn't stop thinking about that kind of disparity between the two folks, okay?
120. 06:57And I put them both on a line, Zeshner from one end, Lapopolo on the other end.
121. 07:01I called it a continuum.
122. 07:03And I basically started asking people, "Hey, where are you on this line?
123. 07:08Where are you a Zeshner?
124. 07:09Do you still read every line of code?
125. 07:11Are you a Lapopolo?
126. 07:12Do you just YOLO and don't even look at code and think agents are good enough, etc.?"
127. 07:17And I I got the framing wrong.
128. 07:20But, I'll tell you about this in just a second, okay?
129. 07:22So, before this, I want you to be kind of honest with yourself.
130. 07:24And again, if you don't write code, or let me say this, if you don't babysit your own agents, but you you have reports that babysit agents for you, uh think about them when you answer this, okay?
131. 07:33And be honest.
132. 07:34On the ZL continuum, where are you?
133. 07:39And let's take um let's take by vote of hands, who here has committed code that they've never looked at before?
134. 07:48Amazing, love that.
135. 07:50Who here still reads every line of code or at least critical code?
136. 07:54I see one cowboy over there.
137. 07:55I love that, man.
138. 07:56I'm going to talk to you afterwards, okay?
139. 07:57I want to understand exactly why you do this.
140. 07:59Um, and so who's right?
141. 08:03Let's talk about who's right.
142. 08:04Let's talk about where we are right now, and we start with Ryan Lepopolo.
143. 08:07If you get to meet Ryan over here, he's very AGI built.
144. 08:11I think even within OpenAI, the AGI organization, Ryan is kind of like the more AGI built person.
145. 08:17If you have a chance to go downstairs and grab the AGI pills that 2X prescribed, I think Ryan had all of them.
146. 08:23He works at OpenAI, where he sits code is literally free.
147. 08:28So are tokens.
148. 08:29It's really We renamed Ryan, do you guys know the dash dash YOLO in Codex?
149. 08:34It's kind of like the skip dangerous permissions in cloud code.
150. 08:36So we renamed the Ryan Lepopolo YOLO Popolo.
151. 08:41He's okay with it, by the way, I asked him.
152. 08:43So if we check his kind of side, the the folks like him against the data, they're actually right.
153. 08:48The optimists are right, at least about output.
154. 08:50This is from Faros AI, I think I'm not the only speaker at this conference who sites this essay.
155. 08:56It's sorry, this survey.
156. 08:57It's new from April 2026.
157. 08:59I think it's one of the best kind of evidence of where we're going that we can now site, okay?
158. 09:0422,000 engineers were surveyed about code.
159. 09:07They call this the acceleration whiplash.
160. 09:10And they're talking about my favorite stat on here, and you can read this yourself, 861% increase in code deletion per PR.
161. 09:18So us together with AI agents, we love deleting code.
162. 09:22Anthropic also said that they're shipping eight times more code per quarter than in 2025.
163. 09:29But is it all good code?
164. 09:33Okay, let's play a game.
165. 09:34If you know the answer, you let me have my moment on here on stage, okay?
166. 09:38But if you don't know the answer, let's guess.
167. 09:41Whose status page is this?
168. 09:45I think I I hear a few answers.
169. 09:47I think most of us guessed it.
170. 09:48This is Claude.
171. 09:50Uh in fact, as you can see on the on the right, it was down when I took the screenshot.
172. 09:55It was really funny.
173. 09:56Uh this Anthropic is the company that probably uses the most AI-generated code, and their status page looks like a Christmas tree.
174. 10:04Now, I'm not here to dunk on Anthropic.
175. 10:06Uh they did an incredible job back on stage uh talking about Claude and etc. Um this may be due to scale, this may be due to other factors.
176. 10:13Uh I'm not here to dunk on them, but
177. 10:15[snorts]
178. 10:15it just goes to show that they're not the only ones like this.
179. 10:17Obviously, GitHub is famously also suffers from a little bit of golf.
180. 10:21Output does not mean stability, okay?
181. 10:23So, maybe this is a good example of what?
182. 10:27Same essay, 31% increase in PRs merged with no review at all, human or agentic.
183. 10:34Don't do this.
184. 10:35I beg of you.
185. 10:35Don't It's it's We'll talk about how to fix this in a second.
186. 10:39So, when you ship this fast and this much, something gives, and usually it's quality.
187. 10:44So, maybe Mario's right, yeah?
188. 10:46Maybe the bill does come due in production.
189. 10:49Same study, 242% increase in incident per PR.
190. 10:54This kind of scary.
191. 10:55The second study is also scary.
192. 10:56Bugs per developer is up six times than 2025.
193. 11:02So, even Anthropic can see this.
194. 11:06I don't know if you guys read the RSI essay they posted, the recursive self-improvement, where they talk about, "Hey, what does the future hold?"
195. 11:14They outline two scenarios.
196. 11:15One of them says, "Maybe the acceleration will stop, and we're going to get used to this."
197. 11:21They all They say, "That's actually not likely to happen.
198. 11:23We just added this uh eventuality for clarity.
199. 11:26We don't think that's likely to happen.
200. 11:27What we think is going to happen is uh engineers and companies 10Xing to 100x-ing to 1,000x-ing their output and productivity, and then they say this, "We, as we began to push more code around the organization,
201. 11:41human code review has become a new bottleneck."
202. 11:44They're citing Amdahl's law that shows that if you have an explosion of productivity in one area, another area going to going to get blocked.
203. 11:52And nobody removes the human in these organizations.
204. 11:58In fact, careers in Anthropic and careers in OpenAI, they're still hiring humans.
205. 12:01So, nobody's removing the human, and they're both saying that human code review is still a concern.
206. 12:08And here's my mea culpa.
207. 12:09I promise you I'll tell you where I got it wrong the framing.
208. 12:12My mea culpa is the continuum is real.
209. 12:13The Zeal continuum is real, but it's not about the people.
210. 12:17It's about the tasks.
211. 12:20The continuum is real.
212. 12:21It's not about the people.
213. 12:22It's about the task.
214. 12:24Same engineer could be a Ryan LeCompte on Apollo on one piece of code, and has to be Mario Zechner and read every line of other pieces of code.
215. 12:34Different task just need different proof.
216. 12:39If we look at them closely, I obviously character- characterized them.
217. 12:44I practiced this word multiple times and I still got it wrong.
218. 12:46Characterized them.
219. 12:47Their character on both ends for the Zeal continuum.
220. 12:51But, if you look at them closely and what they're saying closely, they're actually not that different.
221. 12:56Ryan's mechanism is moving attention up the layer.
222. 12:58He's saying humans are unreliable at catching repeated mistakes of the same type.
223. 13:03Repeatedly catching the mistakes of the same type.
224. 13:05So, when they you do catch a mistake during the PR review, write the documentation, the linter, and the reviewer needs to remember this once, so the system will catch this type of bugs.
225. 13:15He's not saying don't inspect your code.
226. 13:17He's saying inspect the system, not every line.
227. 13:20Mario from the other end is saying route by task.
228. 13:23If it's not critical, let it rip.
229. 13:25He said it.
230. 13:27And if it's critical, you read every line.
231. 13:29How do you know what's critical?
232. 13:32Well, his answer is easy.
233. 13:33You read the effing code.
234. 13:34Uh my answer to add to this is also you ask your clankers.
235. 13:38They're great at looking at a large repository and saying you and telling you, "Hey, this line is actually critical.
236. 13:44You should look at this area.
237. 13:45This primitives over here are critical."
238. 13:47So, you ask your clanker.
239. 13:49So, they agree more than I can have given credit for.
240. 13:53And so, I think at the beginning of this, the wrong question is should I still be reading code in 2026?
241. 14:00I think the better question right now for all of us is what proof does this specific change need?
242. 14:06What proof does this specific change need?
243. 14:10And so, I took Mario on the left, obviously.
244. 14:12I took Ryan on the right.
245. 14:13And then I took a bunch of other uh great AI engineers, friends, some friends, uh many of them speakers at this conference.
246. 14:19I kind of distilled their advice down to a routing table.
247. 14:22And uh they told me, I think Swyx told me on Twitter, "Just give me one slide that I will that people need to take a screenshot of.
248. 14:29It's going to be this slide.
249. 14:29You don't have to read it with me, but in the end, you're welcome to take a picture of this."
250. 14:33This is the your Monday artifact.
251. 14:35Routing the change where the proof needs it.
252. 14:38Routing the change to the proof that it needs.
253. 14:41You read every line of authentication, money movement, permissions, and irreversible data.
254. 14:47You inspect the critical path yourself, and then obviously, you keep going.
255. 14:50Uh decomposing, I think is very important.
256. 14:52The more code is getting written, the more it's hard.
257. 14:55Your eyes are starting to glaze over a a very long pull request.
258. 14:59So, split it into atomic reviewable PRs.
259. 15:01You know who's good at it?
260. 15:02Agents.
261. 15:02They're great at decomposing code.
262. 15:04Ask them to do it.
263. 15:05You verify that doesn't go away.
264. 15:07This has been with us in the in engineering software engineering and the AI engineering doesn't go away.
265. 15:12Traces, evals, shadow mode.
266. 15:14Come talk to me after after this talk.
267. 15:15I don't have enough time, but shadow mode is a really cool one that I learned about preparing this talk.
268. 15:19And then, I think the most important one is separating.
269. 15:21Many people have the same agent that writes the code, also inspects the outputs and writes the tests.
270. 15:26Separating is very important.
271. 15:27If you don't separate, it's kind of like if I came up with an exam and then I took an exam and I scored myself on the exam.
272. 15:33It's not not really productive, right?
273. 15:35And then last one is engineer.
274. 15:37Rails observability rollback.
275. 15:39This is what Ryan Le Popo talks about.
276. 15:41Build the system that builds the system because read spends your attention once, engineer makes the system remember.
277. 15:50Right?
278. 15:52And you might sit in might be sitting there and saying, "Hey, did you hear the news, Alex?
279. 15:56Fable is back.
280. 15:57What about Fable?
281. 15:57What about Mythos?
282. 15:58Is this still relevant at this next scale of capability?"
283. 16:02Because when I coined the Zeal Continuum, it was only 82 days ago.
284. 16:05Mythos has just been announced.
285. 16:07We weren't sure like what was going on.
286. 16:08Only the people in Anthropic got access to it.
287. 16:11And Jared S. Chipar, that was on stage from Anthropic, he said about Mythos and and Fable, "We used to check if Claude is doing the work right?
288. 16:22And with Fable 5, I instead check if Claude is doing the right work."
289. 16:28Let it land for a second.
290. 16:29I don't know if you read the statement.
291. 16:31When I read the statement, I felt like little chills at the back of my neck about the next like level of capability, okay?
292. 16:38We used to check if Claude is doing the work right.
293. 16:40With Fable, we check if Claude is doing the right work.
294. 16:44And our favorite senpai who recently joined Anthropic and is getting unnecessary heat on Twitter said this.
295. 16:50Andrej Karpathy has said, "It's never felt so tempting to stop looking at code at all.
296. 16:56But don't do this in production."
297. 16:58Senpai is great for the sole reason.
298. 17:01Do you guys know the sentence, "This meeting could have been an email?"
299. 17:04So, this presentation could have been Andrej Karpathy's one sentence, okay?
300. 17:07He's naming the anxiety from both ends.
301. 17:11It's never been so tempting to stop looking at code.
302. 17:13Don't do this in production, even with Fable.
303. 17:16And so, if you guys noticed, I have a little thing here.
304. 17:22This.
305. 17:24Uh it's so wide you can see my little laser pointer.
306. 17:28Do you guys see the arrow, the capability drift arrow?
307. 17:30This thing?
308. 17:32When I wrote the continuum, I realized that it it's only a temporary place in time.
309. 17:37Capability increases move us towards the popular.
310. 17:42So, we're going to talk about capability increases as well because the review layer moves.
311. 17:46If yesterday we inspected the outputs and we read the code, and today we inspect the task direction and kind of like directed to the right proof, maybe tomorrow we're inspecting the loops.
312. 17:56Capability drift changes where proof belongs.
313. 17:58It doesn't remove the requirement of proof.
314. 18:02Talking about loops.
315. 18:03Is that the next primitive?
316. 18:05I think most of this conference, I think the zeitgeist for this one is going to be is talking factories and co-factories are real, and is loops is a real thing that I need to be doing at this point.
317. 18:14By raising of hands, who here heard of loops?
318. 18:17Keep your hands up, please, and take them down if you are not running loops right now and you haven't anything about they are.
319. 18:24There's a good perception There's good number of people here who heard about loops, and they started with both these folks, Peter Steinberger, creator of Open Cloud Open Claw, and now is Open AI, and Boris Cherny.
320. 18:36And pretty much within the span of 2 days, both of them start talking about loops that became kind of the zeitgeist.
321. 18:42And loops are moving us from prompting each turn to designing the system that writes the actual prompts.
322. 18:47By the way, do you guys know what's common between these guys and what's different between me and these guys?
323. 18:53Their tokens are free.
324. 18:55So, when they talk about loops and their tokens are free, they're not telling you, "Hey, you should be doing this right now specifically."
325. 19:01But because they work at bigger labs, you can treat them as kind of a lighthouse that's pointing where we're all going.
326. 19:07Kind of like Gretzky skate where the puck is going to to They're going to tell us what all of our enterprises are going to get get up on.
327. 19:15And if it's if it's loops, then let me at least give you a TLDR, okay?
328. 19:19Loops are basically fancy cron jobs that run on a schedule.
329. 19:22But what they do is they discover a task and kind of start writing a prompt for this task from the plan.
330. 19:27They write the plan, they execute, and most importantly for my talk here, they verify themselves, and if it doesn't work, they try again.
331. 19:34So, an agent that loops grades its own work against the goal with less human intervention.
332. 19:40But if the builder grades itself, you didn't remove the review, you hid it.
333. 19:45Okay, this this connects to my routing table.
334. 19:48This comes from Adi Olsmanyu recently at Google.
335. 19:49He's also at this conference, a great engineer.
336. 19:52Uh he said if I if I if I wasn't reviewing the code myself, or relied entirely on automated loops to fix my code, let's say a bug comes up in Jira and my loop picks it up and starts fixing this, my product quality would suffer.
337. 20:05I'd likely end up in a downward spiral, digging myself into a deeper hole.
338. 20:09So again, loops don't remove judgment, but they do raise the stakes on where you put it.
339. 20:15So what about the future, folks?
340. 20:17Nobody knows.
341. 20:21Anthropic did not know that Claude code is going to explode on them and this is going to be a billion-dollar product.
342. 20:26Uh nobody knew that coding agents and Harness are going to be the generalized agent, and now everybody's pursuing them.
343. 20:32Folks at OpenAI with Codex, Elon with with Grok code, uh Google with with anti-gravity.
344. 20:39Model capability is jumping at an insane pace, and what I implore and to tell you here is that flexibility is required.
345. 20:47You need to keep being able to keep up with with the trends.
346. 20:49This is why you stay engineer.
347. 20:51And by the way, I told some folks here about my podcast Thursday and news.
348. 20:55If you want to keep tracking where that line moves, feel free to scan this QR code, join our uh our, you know, newsletter, etc. Um and I'll leave you with this.
349. 21:05Because it's my time, I'll leave you with this.
350. 21:07Not every line in 2026 needs your eyes.
351. 21:11Every system still needs your judgment.
352. 21:13Thank you.
353. 21:15[applause]