Coding Isn't About Code Anymore — Transcript

Full transcript

1. 0:00I've been spending a lot more time
2. 0:01talking to engineers outside of my
3. 0:03little bubble in San Francisco recently.
4. 0:06It's a combination of the events I've
5. 0:07been going to, the courses I've been
6. 0:09helping with, and all sorts of other
7. 0:10things. I've been having a lot more of
8. 0:12these conversations about what it means
9. 0:14to code with AI and what even is an
10. 0:16engineer. Today, I've been thinking
11. 0:18about this a lot, and I've been kind of
12. 0:19wanting to do a video on it for a while,
13. 0:21but I'll be real with you guys. I know
14. 0:23what my comment section looks like and I
15. 0:24know how much a lot of you will hate
16. 0:26this video. So, honestly, I'm not doing
17. 0:29it for those people. I'm probably just
18. 0:31going to ignore the comments on this one
19. 0:32entirely because I I fear what they're
20. 0:35going to look like. But I'm not making
21. 0:37this video for the people who are going
22. 0:38to be leaving the nasty comments that
23. 0:39Theo's not a real engineer. He's just a
24. 0:41JavaScript dev that thinks AI is so
25. 0:43smart. It probably is smarter than him.
26. 0:45This video is just it. We can ignore
27. 0:48those people. This video is for
28. 0:49developers who are actually trying to
29. 0:51ship, take advantage of these
30. 0:52technologies, and figure out what their
31. 0:54role is over time. Because if you
32. 0:56haven't changed how you code over the
33. 0:58last year as a result of these tools,
34. 1:00you might not actually have a job in the
35. 1:02next year or two. We are at that point
36. 1:03now. I've danced around it for years and
37. 1:05I was wrong. Everything has changed. And
38. 1:08most importantly, the way that we work
39. 1:10and what our roles are has changed. It's
40. 1:12changed on a really deep fundamental
41. 1:14level. The simplest way I can put this
42. 1:17is that we're not really engineers
43. 1:19anymore. So if we're not engineers, what
44. 1:21are we? Are we tech leads? Are we
45. 1:23product managers? Are we designers? Are
46. 1:25we visionaries? What's our role? I'm
47. 1:28going to make the argument that our role
48. 1:30as gardeners. The best thing we can do
49. 1:32now is maintain a healthy garden. This
50. 1:35is going to take a lot of explanation.
51. 1:38And I'm going to do my best to explain
52. 1:40it. But if I am wrong about this whole
53. 1:42engineers are becoming gardeners thing,
54. 1:44then I guess I'm just an influencer. And
55. 1:46I'm not a very good influencer if I
56. 1:48don't take a quick break for today's
57. 1:49sponsor. I don't know if you guys have
58. 1:50noticed, but it seems like a lot of
59. 1:52stuff is being hacked recently. That's
60. 1:54because as good as AI is at writing
61. 1:56code, it's also pretty good at writing
62. 1:57exploits and finding these types of
63. 1:59things. There's a problem, though. The
64. 2:01models we use every day don't have the
65. 2:03capability to really scan and tell us
66. 2:05what could be insecure about our code.
67. 2:07But they're more than happy to write
68. 2:08insecure code. That's why we need to be
69. 2:10actively monitoring every change as it
70. 2:12goes in, understanding the blast radius
71. 2:13of the changes that we're making, and
72. 2:15also ideally give us the ability to deep
73. 2:17scan our work to find things that might
74. 2:19have been missed from the past. Wouldn't
75. 2:20it be nice if there was someone who
76. 2:22could do that? Good news. Today's
77. 2:23sponsor is Code Rabbit and their new
78. 2:25Code Rabbit rabbit secure product is
79. 2:26here for just this. They can do deep
80. 2:28scans to find things that you might have
81. 2:30missed in the past. And they can scan
82. 2:31every PR as they come in, making sure
83. 2:34they are as secure as possible. Back in
84. 2:36my day, an exploit being discovered
85. 2:38wouldn't be exploited for two or more
86. 2:40years relatively often. Now, as soon as
87. 2:43the patch is live on the repo, we see
88. 2:45people exploiting it immediately. Best
89. 2:47thing you can do to prevent this is to
90. 2:49make your code secure before it even
91. 2:50goes out. When you start shipping
92. 2:52faster, it's much easier to let things
93. 2:53slip through the cracks. That's what
94. 2:55Codepcure is here to help prevent.
95. 2:57Understanding the reachability,
96. 2:58exploitability, and the blast radius of
97. 3:00every change has never been more
98. 3:02important. You can get started now with
99. 3:0314 days free at soyv.link/codabrabbit.
100. 3:06This is one of those videos that's going
101. 3:08to have a lot of layers to it and get a
102. 3:10little bit chaotic. So, sorry in advance
103. 3:12to you and also to my poor editor who's
104. 3:15going to have to figure out how to turn
105. 3:16this into a somewhat cohesive video. I
106. 3:18already see people in my chat who are
107. 3:20concerned that this is going to be a sad
108. 3:23video and I really don't want it to be
109. 3:26that. I'm going to start backwards here
110. 3:28a bit because I want people to stop
111. 3:30being scared of this part and instead be
112. 3:33excited about this one. So, let's fixate
113. 3:36first on this bit. The idea of
114. 3:39engineers. I don't know about y'all, but
115. 3:42the thing I love about engineering is
116. 3:44the feeling of finding the right pieces,
117. 3:48assembling them the right way, and
118. 3:50creating something that is really
119. 3:52meaningful and impactful. I love the
120. 3:55Lego building block aspect of things.
121. 3:58And most importantly, I love how you can
122. 4:01change the size of the blocks, so to
123. 4:03speak. Like with Legos, you always have
124. 4:05those same low-level building blocks.
125. 4:08Unlike with Legos where you just have
126. 4:10these single blocks you are assembling
127. 4:12in different ways. In engineering, you
128. 4:14have different levels of primitives. The
129. 4:16building blocks could be the actual
130. 4:18silicon in the chip or you can abstract
131. 4:20higher and be living in software writing
132. 4:22assembly or you can abstract higher to a
133. 4:24language like C or you can start using
134. 4:26runtimes or packages or other tools. You
135. 4:29can build your code in a way where the
136. 4:30lines are composable or the functions
137. 4:32are composable or the modules are
138. 4:34composable. The flexibility in where you
139. 4:38break things apart and the dynamic
140. 4:40nature of how we can assemble them
141. 4:42together is just so fun. I love
142. 4:46it. That's why I love writing software.
143. 4:48That's why I love building. That's why I
144. 4:49love all of the stuff that we've been
145. 4:51doing forever. And what I have been
146. 4:53scared of is losing that. And I know a
147. 4:55lot of people seem to have a fear that
148. 4:58that feeling of like tinkering and
149. 5:00finding the right pieces and getting
150. 5:01them together the right way is going to
151. 5:03be gone because when you start using AI
152. 5:05it is a lot of what we have loved about
153. 5:08software dev when you start going down
154. 5:10this rabbit hole just isn't there
155. 5:12anymore and that sucks. But the good
156. 5:14news that I hope you can get out of this
157. 5:17video is that when you get through it,
158. 5:19when you come out the other end, the
159. 5:21types of problems you'll be solving, the
160. 5:23types of tools you'll be assembling, the
161. 5:24types of building blocks you'll be
162. 5:26building and building with are so much
163. 5:28more fun than I've ever had building
164. 5:30software before. Like, I genuinely am
165. 5:33just unable like I've been going to bed
166. 5:35at 6 a.m. at least once or twice a week
167. 5:36recently because I get so sucked in to
168. 5:38the stuff that I'm building. It's just
169. 5:40so fun and rewarding. And I'm not doing
170. 5:43this out of fear or FOMO or worry that
171. 5:45I'll be left behind. I'm doing this
172. 5:47because I love it right now. So, if
173. 5:49you're coming in here fearing what's
174. 5:51happening, I hope you can open your mind
175. 5:54a little bit and just tone down that
176. 5:56fear slightly because I felt it a bit
177. 5:58too. I am lucky that my career is not
178. 6:00just software dev. So, I was able to get
179. 6:02through it faster because it didn't feel
180. 6:03like my world was crumbling. It just
181. 6:04felt like one of my hobbies was. But
182. 6:06then I found a deeper love for it on the
183. 6:08other side. Let me now explain how I got
184. 6:11here. I don't know how to make this
185. 6:13video in any way other than it being
186. 6:15really personal. So, I think I'm just
187. 6:18going to do that. I think the best place
188. 6:20to explain why I think this way starts
189. 6:23back with my time at Twitch. For those
190. 6:25who don't know, I was working at Twitch
191. 6:27from 2017 to 2021. I hopped around Teams
192. 6:30a decent bit when I was there, but I
193. 6:33want to focus on a specific
194. 6:35cross-section here, which is roughly
195. 6:37from 2019 to 2021. There's a couple
196. 6:40reasons I want to focus on this area,
197. 6:42but the simplest way to put it is from
198. 6:442017 to 2018, I did pretty well. From 18
199. 6:47to 19, I almost got fired because I just
200. 6:50did not like the things I was working
201. 6:52on. When the team I joined for at Twitch
202. 6:55folded and the exec who pulled me into
203. 6:57Twitch retired, I was just not happy.
204. 7:00The main reason I survived this window
205. 7:03where I was this unhappy was because my
206. 7:06new manager who I got thrown under with
207. 7:08a team shakeup internally had taken a
208. 7:10liking to me. Knew I was a decentish.
209. 7:13Still not great but like motivated. He
210. 7:16was a front-end guy. I was a deep
211. 7:17systems backend guy mostly doing like
212. 7:19video infrastructure and elixir. And I
213. 7:22had to move to a different video team
214. 7:24the VOD team and work in Go and was just
215. 7:26miserable. He saw how miserable I was.
216. 7:28He pulled me aside and told me I should
217. 7:30go try out this new React TypeScript
218. 7:33thing that we were starting to use more
219. 7:34internally. I didn't like it much at
220. 7:36first, but it made some sense. But then
221. 7:39hooks happened, which was when React
222. 7:41added more functional programming
223. 7:43patterns, and I got, for lack of a
224. 7:45non-pun way you're describing it, hooked
225. 7:48quite quickly because I was already a
226. 7:50huge functional programming nerd due to
227. 7:51Elixir. And the mental model for hooks
228. 7:54clicked so much better for me. And I
229. 7:56went from feeling like I didn't belong
230. 7:58when I joined in 2017 to like mid 2017,
231. 8:01mid 2018 being like, "Oh I
232. 8:03understand where everything is in my
233. 8:04head. I know how all these parts come
234. 8:06together. I understand the system. I can
235. 8:09change anything in the system and know
236. 8:10what will happen." And that feeling was
237. 8:12the first time I had that like a thing
238. 8:14that mattered. I understood the outside
239. 8:17and inside of and I could reason about
240. 8:19how things would behave when changes
241. 8:21were made. I did not get that same
242. 8:23intuition especially as I was joining a
243. 8:25team that already had a gigantic
244. 8:26codebase in a language I was not
245. 8:28familiar with and were they were
246. 8:29building in a way I was not used to at
247. 8:31all. I was really struggling to get that
248. 8:33same grasp around the VOD platform code.
249. 8:37So when React Hooks happened and at the
250. 8:39same time Twitch was starting to move
251. 8:41away from the old decrepit Ember plus
252. 8:44Rails codebase for the site and towards
253. 8:46a Go plus GraphQL plus React TypeScript
254. 8:49codebase. That was a huge shift for me
255. 8:52and I got to be there when the thing
256. 8:54started to be made and I got to be the
257. 8:56person advocating for this new way of
258. 8:59building that I got to like see a bit
259. 9:02earlier than others did like why it was
260. 9:04valuable because most of the people who
261. 9:06were part of this huge rewrite had been
262. 9:08writing React code for 5 years had been
263. 9:09writing web code for 10 years and I had
264. 9:11just moved to web code like a few weeks
265. 9:14prior with a bunch of weird niche
266. 9:16functional programming experience and I
267. 9:18was just so genuinely exced excited that
268. 9:21they pulled me under their wing for the
269. 9:22team running the like core web product.
270. 9:26I should give a little bit of background
271. 9:28on like how teams are structured at
272. 9:30Twitch. Most of the teams are feature
273. 9:32teams. So, right now I'm live and I have
274. 9:35a bunch of people in chat. The chat team
275. 9:37is a team. So, there's a team that is
276. 9:39chat. Their job is running all of the
277. 9:42chat related things. Then there's a team
278. 9:43that is the creator dashboard. The
279. 9:45creator dashboard is the view that I was
280. 9:47just in here where I could see
281. 9:49everything going on in my stream on
282. 9:50Twitch. Then one of my personal
283. 9:51favorites is the mod tools team. This is
284. 9:53the team that built a lot of the
285. 9:54moderation platform stuff that we used
286. 9:56at Twitch to allow mods to keep our chat
287. 9:59safe. Things like the mod view platform
288. 10:01that I built and talked about in a bunch
289. 10:02of other videos. These are just like,
290. 10:04funny enough, the three teams I was on
291. 10:06for like core product in my time at
292. 10:08Twitch. And all of these teams are
293. 10:10focused on like one specific surface
294. 10:13area that they own all of. So the mod
295. 10:16tools team owns the backend infra, the
296. 10:18database, the client, the web code, and
297. 10:22all of the things people interface with.
298. 10:24The chat team owns the backend, the
299. 10:26protocols, the way that messages go to
300. 10:29and from the chat, and of course like
301. 10:31all of the backend info side, as well as
302. 10:32all of like the emoji stuff and the
303. 10:35important like features built into T or
304. 10:37the chat inside of Twitch. Creator
305. 10:39dashboard is, as you expect, very
306. 10:41similar, but this is one of the two main
307. 10:43types of product and engine teams at
308. 10:45Twitch. There's a very hard separation
309. 10:48between them. I'll call this side
310. 10:50product. I'm going to call the other
311. 10:52side platforms because while mod tools,
312. 10:55chat, and creator dashboard, they owned
313. 10:58their portions of the website and their
314. 10:59portions of the back end, their portions
315. 11:01of the database, we still needed a team
316. 11:03to make sure the APIs didn't fall over
317. 11:05and could scale the way Twitch did. And
318. 11:07it's not the responsibility of each
319. 11:09product team to make sure that every
320. 11:11single thing they do scales and plugs in
321. 11:12perfectly with everything else. That
322. 11:14would just be insane if every single
323. 11:15team had to build their own scalable
324. 11:17infra for every single thing. It would
325. 11:19be hell. So we end up also with teams
326. 11:21like browser clients. Browser clients
327. 11:23was the team that managed the repo for
328. 11:25the website, the repo for the electron
329. 11:26app that was still being maintained, the
330. 11:28repo for all of the different web
331. 11:30product stuff that we had at Twitch,
332. 11:32like the smaller mobile site that we
333. 11:33were pushing in other countries at the
334. 11:34time. Then we had the GraphQL Edge team.
335. 11:36We referred to the GraphQL API as the
336. 11:38edge because it was just the
337. 11:40front-facing thing that the user APIs
338. 11:42would interface with. Like when you're
339. 11:44doing things on Twitch, the graphical
340. 11:45edge is what hits and then we would
341. 11:47federate out to the smaller services the
342. 11:49different teams owned. And of course
343. 11:51there's like the data database platforms
344. 11:53team, the video infrastructure platform
345. 11:54team. There were different teams for the
346. 11:56different platforms within Twitch. And
347. 11:58this meant that a relationship had to be
348. 12:00built between the product teams and the
349. 12:03platform that the product teams were
350. 12:04building on. This ended up being one of
351. 12:07my favorite parts of working at a
352. 12:09company like Twitch is that I got to
353. 12:11spend most of my time on product. I was
354. 12:13never at any point formally on a
355. 12:15platform team. But if you ask the
356. 12:17platform teams, they would possibly say
357. 12:19otherwise depending on who you're
358. 12:20talking to because I spent so much time
359. 12:23hanging out with the platform teams. For
360. 12:25example, with browser clients, the guys
361. 12:26who ran the repo for the Twitch site, I
362. 12:29was the only non browser clients or
363. 12:32platform team member that was in their
364. 12:34weekly planning meetings because I just
365. 12:38cared a lot about the platform. This all
366. 12:40started because of a conversation I had
367. 12:42early on when I was working on the site.
368. 12:44And I was reading through the docs for
369. 12:46this new Twitch rewrite in React. And I
370. 12:49was trying my best to take what I
371. 12:50learned from the docs and apply it to my
372. 12:52code. And I made a PR that was not very
373. 12:54good. And I remember a guy named Derek
374. 12:57jumped on it and like came over to my
375. 12:59desk and asked me just questions on like
376. 13:01why I wrote it this way. And I showed
377. 13:02him the pages in the docs that led me
378. 13:04there and the examples in the codebase
379. 13:06that I was referencing. He thanked me
380. 13:07for that. said it was super super
381. 13:09helpful. Ran back to his desk and within
382. 13:1130 minutes had two PRs up. One getting
383. 13:14rid of the junk in the docs that misled
384. 13:16me and the other rewriting all the code
385. 13:18in the code base that confused me. And
386. 13:21that's when I realized how important
387. 13:23that work was. That if you weren't
388. 13:26setting up your team for success, then
389. 13:29that would just constantly be lowering
390. 13:31the bar. And every time something even
391. 13:34gets to the point where it's a poll
392. 13:35request that isn't good, that shows an
393. 13:38opportunity to improve things to make it
394. 13:40so the next person doesn't make those
395. 13:43mistakes. And if you don't take those
396. 13:44opportunities as soon as you see them,
397. 13:46the quality of everything around you is
398. 13:48going to erode so fast. And this is the
399. 13:51first time I'd like really seen that in
400. 13:54action. I like this framing from Twitch
401. 13:56chat. Thank you, Thus. Hopefully I
402. 13:59pronounced that right. Do not fix the
403. 14:01symptom, fix the cause. Exactly. And I
404. 14:04ended up really loving this. There was a
405. 14:07combination of other fun things like
406. 14:08when I was on the safety platform/safety
407. 14:11tooling team internally, not for mod
408. 14:13tools yet, but for internal safety
409. 14:15stuff, we were kind of running our own
410. 14:16like clone of the Twitch infrastructure
411. 14:19so that it was easy to hop between the
412. 14:21internal tools and external tools
413. 14:22without having your mind get all like
414. 14:24screwed up because everything was so
415. 14:25different. But our repo was smaller and
416. 14:27only had like 300 users instead of the
417. 14:29like millions that the core Twitch site
418. 14:30had. So I was able to experiment with
419. 14:32ideas that we weren't sure were ready
420. 14:34yet for the main Twitch repo and then
421. 14:36bring them to the browser client team
422. 14:37being like, "Yo, we dog fooded these
423. 14:39four ideas. These two kind of worked.
424. 14:41This one seems worth it. We should
425. 14:43figure out how to do this in the main
426. 14:44Twitch site." And they loved that so
427. 14:46much. Like to this day, a lot of my
428. 14:47friends from Twitch are still just
429. 14:49people from the browser clients team. I
430. 14:51absolutely love that team. They offered
431. 14:53multiple times to let me switch to their
432. 14:54team, but I just liked building product
433. 14:57too much to do it. I love building
434. 14:59things people actually use and play with
435. 15:01and interact with that make their lives
436. 15:04better in some small way. And I don't
437. 15:06like the platform side unless it's
438. 15:08making a product better. And I don't
439. 15:09like the product side unless the
440. 15:11platform is good and also honestly in
441. 15:13continuing to improve as well. And it's
442. 15:15this relationship between making product
443. 15:17that solves real problems for people and
444. 15:19building platforms that allow those
445. 15:21products to grow and succeed that made
446. 15:24this relationship just so fun and
447. 15:26enticing for me. So huge shout out to
448. 15:28the browser clients team in particular
449. 15:30at Twitch for really flipping how I
450. 15:32think about engineering. It was one of
451. 15:33the like the moments in my life that
452. 15:35formed the way I work now. And I have
453. 15:38been thinking back on this moment a lot
454. 15:40recently because I've been trying to
455. 15:42figure out why I was able to make the
456. 15:44move to working with agents so quickly.
457. 15:46I already see the comment below this
458. 15:48video. The reason Theo could make the
459. 15:50move so fast is because he was writing
460. 15:51shitty JavaScript slop so he doesn't
461. 15:53care about the quality of the code. I
462. 15:55hope that what I'm describing here shows
463. 15:57exactly the opposite. that the quality
464. 16:00of the code is as much a communication
465. 16:03and like organization problem as it is
466. 16:07like the code that you're writing and
467. 16:08the quality of the engineers. And
468. 16:10ideally, if all goes well in the world,
469. 16:13an okay engineer and a phenomenal
470. 16:15engineer can both write good enough code
471. 16:18to merge where it will not degrade the
472. 16:20quality of your codebase. Before we go
473. 16:22to this next section, I want to call out
474. 16:23a few of the things I didn't say at any
475. 16:26point so far. I did not talk about
476. 16:28becoming a manager. I did not talk about
477. 16:30super intense code review processes. I
478. 16:33did not talk about building a bunch of
479. 16:35guard rails and unit tests to prevent
480. 16:37things from going certain ways. I didn't
481. 16:38talk about code coverage. I didn't talk
482. 16:40about any of that All I've talked
483. 16:42about is this relationship between
484. 16:44setting up a platform that good product
485. 16:47can grow out of and setting up product
486. 16:50to give feedback to the platform so that
487. 16:52they can grow together. So what happened
488. 16:55next? What happened next is that I quit
489. 16:57my job at Twitch. And this was a big
490. 17:00forcing function for how I worked. Both
491. 17:03because I'm now out of a job, but also
492. 17:06just cuz it it changed why I cared. Not
493. 17:10just what I cared about, but more
494. 17:11importantly the why. Because a lot of
495. 17:13the reasons I would care about the
496. 17:15platform the way I did before is that it
497. 17:17would set up my co-workers to do better
498. 17:19things. It would set up other teams to
499. 17:21make better product. And I got to evolve
500. 17:23everyone around me by making these
501. 17:25changes. Everything I had done at Twitch
502. 17:27was trying to make it easier for all of
503. 17:28us to ship good stuff. It was not
504. 17:31necessarily focused on faster, although
505. 17:33that was a nice consequence of things. I
506. 17:35was regularly pulled in to try and get
507. 17:37timelines back on track when road maps
508. 17:39started to get blown up, but I really
509. 17:41just wanted to focus on like building
510. 17:43the platform out so it was easy to do
511. 17:45the right thing and there was no
512. 17:47examples of how to do the wrong thing.
513. 17:49But then I went from working at a
514. 17:50company with 1500 employees to a company
515. 17:52with one me. And my priorities
516. 17:56necessarily had to shift because when I
517. 17:59started paying, it was just me. And now
518. 18:02I don't have a platform client team to
519. 18:05rely on for making sure our Webpack
520. 18:07bundle works properly across all the
521. 18:09different ways it can be used. I don't
522. 18:11have a back-end video infrastructure and
523. 18:12platform team. I don't have a GraphQL
524. 18:14edge distribution team. I didn't have
525. 18:17the teams who owned these things
526. 18:18anymore. And my first instinct was to
527. 18:21try and replicate the things those teams
528. 18:23did and to try and build out all of the
529. 18:25stuff I love. And I was spinning up
530. 18:27GraphQL backends with Elixir powering
531. 18:29them and spinning up crazy complex
532. 18:32client bundles where I was going through
533. 18:34and manually managing the way the
534. 18:35bundles were split and stuff like that.
535. 18:36I was able to make a product that kind
536. 18:37of worked, but I ended up just doing all
537. 18:40of the jobs of all of these different
538. 18:42people that no longer were my
539. 18:43co-workers. And I realized that I had to
540. 18:46change the way I was thinking about
541. 18:48these things. I still wanted to set up
542. 18:50everything in my codebase for success so
543. 18:52that the right thing was easy and the
544. 18:54wrong thing was hard. But I had to make
545. 18:56all of the things faster and easier to
546. 18:58do and less work to maintain. And then a
547. 19:01thing happened that the people who have
548. 19:03been here for a while have heard of. T3
549. 19:06stack. The T3 stack happened because I
550. 19:09needed the confidence of the platforms I
551. 19:11was used to without the friction that
552. 19:14was necessary to build them. And I
553. 19:16created this set of tools that to be
554. 19:19clear, I didn't really create the tools.
555. 19:20My innovation here wasn't particularly
556. 19:23innovative. I took three sets of tools
557. 19:25people were using well and assembled
558. 19:27them into one. And it was Nex.js
559. 19:29TypeScript client with Nex.js as the
560. 19:32server alongside of course React on the
561. 19:35client. I brought in Tailwind because I
562. 19:37started finally to like Tailwind around
563. 19:38this point. The more important pieces
564. 19:39were how the front end and back end
565. 19:41worked together though. The back end was
566. 19:43using Prisma for the database because it
567. 19:45was full stack type safety from the
568. 19:47database to the back end. And then I
569. 19:49used a new tool called TRPC that I'd
570. 19:51fallen in love with that made it easy to
571. 19:53define functions on the back end in
572. 19:54Typescript and then call them on the
573. 19:56front end without having to define a
574. 19:57bunch of contracts between the two. You
575. 19:59still got a contract out of it, but the
576. 20:01contract was TypeScript. wasn't another
577. 20:03language like GraphQL or another weird
578. 20:05standard like Open API or something. And
579. 20:07this was the thing that blew me up
580. 20:09initially on YouTube when I started
581. 20:11telling people about my stack cuz I made
582. 20:13this before I'd even started doing
583. 20:14YouTube videos. And my first video was
584. 20:16actually just me streaming trying to
585. 20:18explain it to people I'd met on Twitter
586. 20:19when I was a nobody. And yes, chat, you
587. 20:21remembered correctly. I learned TRRPC
588. 20:24because Tanner Lindsley gave it to me as
589. 20:25homework because he saw Fred Shot, the
590. 20:28creator of Astro, and Ryan Carnegie, the
591. 20:31creator of SolidJS, were following me on
592. 20:32Twitter. So, when I requested to join
593. 20:34one of his Twitter spaces back in the
594. 20:36day, he let me up. We talked about some
595. 20:38weird proposals I had for making it
596. 20:39easier to make these contracts between
597. 20:40back end and front end. He brought up
598. 20:42TRPC. I went and explored it and fell in
599. 20:44love immediately. Alex, the main creator
600. 20:47maintainer of TRPC, is still a good
601. 20:48friend. This is the community that we
602. 20:50got Julius from, who is now the T3 code
603. 20:52guy. So much happened because of this T3
604. 20:54stack moment. And that moment happened
605. 20:57because I loved how good it felt to have
606. 21:00a platform you understood and could
607. 21:03build on top of confidently. And most of
608. 21:06T3 stack came out of my product. I was
609. 21:08building the best stack I could within
610. 21:11Ping to make the Ping product as easy
611. 21:13and like able to be maintained as
612. 21:16possible while I also was doing my
613. 21:17YouTube channel and was also doing all
614. 21:20of the like fundraising for what became
615. 21:22T3 tools which is now T3 Code and T3
616. 21:24Chat getting into Y Combinator doing all
617. 21:26the business side there doing all the
618. 21:27sales marketing and everything I
619. 21:29couldn't be spending 10 to 15 hours a
620. 21:32day coding like I wanted to so I was
621. 21:34putting all this effort in to make sure
622. 21:36that when I did have time to code it was
623. 21:37very clear where the time should go and
624. 21:40that I was probably 50/50 split between
625. 21:41my time spent doing platform code making
626. 21:44it easier to work on Ping to the actual
627. 21:46product code making the product itself
628. 21:48better. This sounds insane when you
629. 21:50think about it that like half the poll
630. 21:51requests I was filing as a solo dev did
631. 21:54not affect the users at all. They did
632. 21:56not see any difference from half the
633. 21:57poll requests I was filing and merging.
634. 21:59But those poll requests made it so the
635. 22:01other half were way easier to land. And
636. 22:03that was the back and forth, the the
637. 22:05pendulum swing between the two that I've
638. 22:08been doing my whole career since.
639. 22:09Knowing when to tap out on product
640. 22:11because the thing is harder than it
641. 22:13should be. Give it up for a minute, go
642. 22:15back to the platform, make the platform
643. 22:17easier to build on top of. And then once
644. 22:20that is in a good state again, go back
645. 22:21and build the feature. That back and
646. 22:23forth is what leveled me up more than
647. 22:26almost anything in my career is learning
648. 22:28when and how to tap out of one side and
649. 22:31go to the other and to not treat them as
650. 22:33different like roles or teams or things
651. 22:37that you can only do one of but to blur
652. 22:39the lines more. And this is why I was
653. 22:41able to like make all the friends I made
654. 22:43at Twitch because I was the platform
655. 22:45team's favorite product guy and the
656. 22:46product people loved me because I was
657. 22:48their connection to the platform. And I
658. 22:49cannot tell you how many times I was
659. 22:51getting blocked in a product meeting
660. 22:53because the other engineers or some team
661. 22:55leader, someone was convinced there was
662. 22:57no way that the GraphQL Edge team would
663. 23:00let us do this thing that was obviously
664. 23:02the right call. So, I would hop on a
665. 23:04quick meeting with the guy running that
666. 23:05team, Tony, and say, "Hey, I'm probably
667. 23:07not the right person for this because
668. 23:08I'm not the backend engineer on my
669. 23:10team." But the backend engineers are
670. 23:11convinced that this weird recursive data
671. 23:14structure we need for the layout of the
672. 23:16dashboard in here, funny enough, like
673. 23:18cuz everything here is mobile and can be
674. 23:20like built like nested. The best way to
675. 23:23encode this is just a JSON blob. And my
676. 23:26team was insisting that the GraphQL team
677. 23:28would not let us just put a JSON blob in
678. 23:30the GraphQL API and that we had to make
679. 23:32this super complex node system to fetch
680. 23:35them all as a flat array and then
681. 23:37reconstruct it into the dynamic layout.
682. 23:41I knew this was wrong and stupid. I
683. 23:43couldn't convince my team. So I went to
684. 23:45my friend who ran the other team, the
685. 23:48GraphQL Edge team, and I explained the
686. 23:50situation to him. I didn't tell him what
687. 23:51the options were. I just told him what
688. 23:53the problem was. to which he responded
689. 23:55like he interrupted me to say why don't
690. 23:57you just store the JSON blob I asked him
691. 23:58if there's any reason not to he thought
692. 24:00about it for a second said no that seems
693. 24:01like the right call here like okay how
694. 24:04do I convince my team that this is the
695. 24:05right call he said I don't know none of
696. 24:07them ever hit me up and then he vented
697. 24:08to me we he ended up using the rest of
698. 24:10the call to vent to me about how
699. 24:12frustrating it is to be on platforms in
700. 24:15GraphQL land and to feel like nobody on
701. 24:17the other side like no one from product
702. 24:19cares at all about what he's doing and
703. 24:21he's just stuck saying no on shitty PRs
704. 24:22all day and he was so thankful to have
705. 24:25somebody from the front-end teams on the
706. 24:27product side hit him up for a call that
707. 24:29we ended up chatting all the time and to
708. 24:30this day Tony is still a very good
709. 24:31friend he runs a lot of the platform for
710. 24:33Netflix now if I recall love Tony
711. 24:35death but like these are the
712. 24:36conversations that shaped who I am now
713. 24:38was the realization that so many of
714. 24:41these things so many of the pieces of
715. 24:43friction you encounter as a dev are like
716. 24:46one conversation and one small change in
717. 24:48the platform away from not needing to
718. 24:50happen and all of this is such a
719. 24:53waste of our time. Like I put more time
720. 24:56into those conversations trying to just
721. 25:00get a JSON blob stored than I spent just
722. 25:03building the dashboard itself. It was
723. 25:05insane. I was going mad. Those were the
724. 25:08moments that shaped me the most. And one
725. 25:10of the things I was really thankful
726. 25:11about when I left is that I could just
727. 25:14do the thing on both sides. I could
728. 25:16constantly be refining both ends. When I
729. 25:18was working on teams like the creator
730. 25:19dashboard, every conversation was one
731. 25:22around which costs us the least internal
732. 25:24political Is it better for us
733. 25:26to invent our own node graph system in
734. 25:29order to represent some pretty basic
735. 25:31data? Or is it worse to have a
736. 25:34conversation with this team that might
737. 25:35say no, lose the time, and have to do it
738. 25:37this way? Anyways, some of this was
739. 25:38internal politics, but a lot of this was
740. 25:40just the platform and the way it worked
741. 25:42and the lack of understanding one side
742. 25:44had of the other. This is why I got so
743. 25:46deep in like blurring the lines between
744. 25:48them. And I think that's the biggest
745. 25:50part of why I have had so much fun. Now,
746. 25:53fast forward a whole bunch into the
747. 25:55future. I want to talk about instead of
748. 25:57the start of something, I want to talk
749. 25:58about the end of something. I want to
750. 26:00talk about my last big handcoded
751. 26:02overhaul. The last time probably for the
752. 26:05rest of my life where I sat down and
753. 26:08rebuilt something by hand. And the few
754. 26:10times I tried to have agents help, they
755. 26:12just couldn't do it. So, I gave up and
756. 26:14did it all myself. I see people in chat
757. 26:16already guessing what it is. I'm curious
758. 26:18if any of y'all remember what it is.
759. 26:20Someone already got it was T3 chat
760. 26:22convex. I went from super skeptical
761. 26:25about Convex because it at the time to
762. 26:27me seemed like a closed source
763. 26:28alternative to the T3 stack that I
764. 26:30didn't love. Since then, Convex has
765. 26:32become both open- source and way more
766. 26:34mature than anything I had built for
767. 26:35this type of stuff. It is a combination
768. 26:36of a backend, a database, an API layer,
769. 26:39a SDK for the client and the server.
770. 26:42talked about convex a lot in my videos.
771. 26:44Definitely search convex seal. You'll
772. 26:45find a ton of them. Still how I love to
773. 26:47build most like full stack services when
774. 26:49I'm building things. Absolutely love
775. 26:50Convex. Up until I was ready to use
776. 26:53Convex and T3 chat, I was building my
777. 26:54own sync engine by hand and it was
778. 26:57horrible. It worked, but maintaining it
779. 26:59was hell. The T3 chat port from my
780. 27:03custom sync engine to Convex was one of
781. 27:06the hairiest and most difficult
782. 27:09overhauls I've ever done. And for the
783. 27:11most part, when I finally could merge
784. 27:14that like 25 plus thousand line of code
785. 27:17PR where I shifted everything over, most
786. 27:19people didn't feel a difference beyond
787. 27:22like if they were using it on their
788. 27:23phone and their computer at the same
789. 27:24time, it was more likely to keep things
790. 27:25in sync. This is probably the most I've
791. 27:27worked on one big underlying tech change
792. 27:30in my life. I spent a whole month cuz I
793. 27:32I started the month with COVID, funny
794. 27:34enough, and I got the news at the end of
795. 27:36the month that I had to have my hand
796. 27:37reconstructed. So, I knew this was going
797. 27:38to be my last time like sitting in
798. 27:39coding for a while. So, I spent the
799. 27:41whole month on this and it had its rough
800. 27:43edges. We ended up at the time being one
801. 27:45of the biggest convex customers. So, it
802. 27:47caused a lot of like not outages, but
803. 27:49like weird edges breaking for certain
804. 27:51things we were doing a lot of that other
805. 27:53customers weren't. All got smoothed out
806. 27:54really quick by them. But god damn, this
807. 27:56was so much work and it was so much fun
808. 27:59and none of our users could possibly
809. 28:01have really noticed the difference until
810. 28:03it was done and then it was shipped and
811. 28:05then after is when we really got to
812. 28:08start doing the fun The reason I
813. 28:10did this overhaul wasn't because I was
814. 28:12tired of maintaining the custom sync
815. 28:14code, although I was. It wasn't because
816. 28:16the database costs for how I was doing
817. 28:18things were absurd, because they were.
818. 28:20It was because at this point in time, we
819. 28:22had scaled up the team a bit. I had
820. 28:23pulled Julius full-time. I had Mark
821. 28:25constantly putting up the fires I was
822. 28:27creating, and I was the only person who
823. 28:29could touch the backend code for the
824. 28:31database stuff because it was the
825. 28:33rightish platform for what we need at
826. 28:35the time, but I was the only person who
827. 28:37understood it. And as a result, I was
828. 28:39blocking my team constantly. And it was
829. 28:42really fragile and breaking all the
830. 28:44time, too. The move to Convex fixed that
831. 28:46and allowed me to not be able to code
832. 28:48for a few months because of my hand and
833. 28:50still continue iterating on top of T3
834. 28:52chat and shipping and improving the
835. 28:54product. What I'm trying to say is I
836. 28:55love these overhauls, the ones where no
837. 28:57one outside can or should feel a
838. 28:59difference, but we know it sets us up to
839. 29:02work much more effectively. and then
840. 29:04feeling, seeing, and experiencing the
841. 29:08increased movement that I and my
842. 29:10teammates get on the product side once
843. 29:13this stuff lands. The thing that was
844. 29:15exciting about this overhaul was
845. 29:17watching how many more PRs Mark and
846. 29:20Julius were putting up and landing as a
847. 29:22result of how much easier they could
848. 29:23move around in the codebase. Hopefully,
849. 29:26you can kind of see where we're going
850. 29:27now. I was writing slop before agents
851. 29:30were. The setup we had for T3 chat
852. 29:33before we had convex was a bunch of
853. 29:36theosync engine slop and it was
854. 29:38constantly causing friction. It was
855. 29:40getting in the way of my team. It just
856. 29:42wasn't good. I was an incredibly
857. 29:44effective chef when it came to cooking
858. 29:45slop. Even at the time, I wouldn't have
859. 29:47defended how it was implemented. I would
860. 29:49just defend why it was implemented. And
861. 29:51this is why, as chat pointed out, I had
862. 29:53the five videos of me switching the
863. 29:55database in T3 chat. It was a great
864. 29:57opportunity for me to try out different
865. 29:58things, learn stuff, and build a
866. 30:00scalable platform at an absurd pace
867. 30:03while also having to run my channel and
868. 30:04my company all at the same time. And it
869. 30:06was such a fun era. I I loved every
870. 30:08moment of it. But the important thing to
871. 30:09know about my slop is that my slop was
872. 30:11blocking my team. The reality is that
873. 30:14where the slop comes from doesn't
874. 30:15matter. Neither does who your team is.
875. 30:17Let me reframe this now. Sonnet 4.5 slop
876. 30:20was blocking Opus 5.5. Do you see where
877. 30:23we're going now? all of the time and
878. 30:26effort I was spending on the right side
879. 30:28of this line doing this platform work,
880. 30:32setting up my team and setting up myself
881. 30:35to be more successful building. That was
882. 30:37always one of the things I loved doing,
883. 30:39but I loved it both because it's fun,
884. 30:41but also because I would see the impact
885. 30:42it had on the other side. It would take
886. 30:44a long time to see that impact. It would
887. 30:46take even longer to do the changes in
888. 30:47the first place, but it was very
889. 30:49satisfying work that really made me feel
890. 30:51like an engineer. This work has gotten
891. 30:5410 times more important than it's ever
892. 30:57been. I remember one of the first like
893. 31:00cool conversations I ever had like where
894. 31:02I felt like I should not be allowed to
895. 31:04do this was when I was tweeting back
896. 31:06when I had like 50 followers about
897. 31:09TypeScript stuff we were doing at Twitch
898. 31:10because our TypeScript times for checks
899. 31:12had gotten to like 15 to 20 minutes and
900. 31:15our like npm rundev command time was at
901. 31:18like the 3 to 5 minute mark just to
902. 31:20start working on a project at a given
903. 31:23time like on a random day and I was so
904. 31:26desperate to fix this that I just
905. 31:28randomly replied to a certain person on
906. 31:31the TypeScript team, Daniel Rosen Bosser
907. 31:33on Twitter, asking for some info on some
908. 31:35stuff he was talking about on Twitter
909. 31:36about performance for TypeScript. He DM'
910. 31:38me, we chatted and I got to intro him to
911. 31:41the browser clients team and somehow be
912. 31:44the one to broker this meeting between
913. 31:46the platform team at Twitch and the
914. 31:47TypeScript team at Typescript when I was
915. 31:5021 years old and felt like I didn't
916. 31:52belong anywhere yet. That was the like,
917. 31:54oh all of these things are
918. 31:56malleable. All of these things can be
919. 31:58changed and improved. All of these
920. 32:00things are made by people no more
921. 32:02special than you or I. And you can just
922. 32:05talk to them about things and work with
923. 32:07them on things and get things changed.
924. 32:09And to this day, like when I just put
925. 32:11out TS Rust, the TypeScript team hit me
926. 32:12up and we did a great talk and I got to
927. 32:14chat with Hellsburg, the creator of
928. 32:16Typescript, about the language that
929. 32:18changed my life. And it's just so
930. 32:19cool. And that this is the stuff
931. 32:21I love. And the reason I bring up the TS
932. 32:23Rust thing is because all of the
933. 32:26that got me where I am now and that I
934. 32:29loved then matters more and is making
935. 32:32more cool opportunities. The same
936. 32:34mindset that led to me replying to Rose
937. 32:36and Waser on Twitter to get this meeting
938. 32:38between the browser clients team at
939. 32:39Twitch and the TypeScript team is the
940. 32:41mindset that led to me building TS Rust.
941. 32:44And the best way I can describe this
942. 32:46mindset is gardening. I use gardening as
943. 32:50the analogy here because I want to
944. 32:52emphasize that we shouldn't be the ones
945. 32:54growing the code. We now need to build
946. 32:57environments where the code can grow in
947. 33:00the right direction, have the right
948. 33:01nutrients, have the right support
949. 33:03systems, and have the right monitoring
950. 33:05and everything else in place that the
951. 33:07likelihood the plant grows successfully
952. 33:08is as high as possible. And the
953. 33:10likelihood that you can recover it
954. 33:11before it becomes a weed and infects
955. 33:13your garden is as high as possible, too.
956. 33:15I know this because I was already doing
957. 33:17it. The time I was spending trying to
958. 33:20improve the type checks at Twitch were
959. 33:23to make it so everybody could work on
960. 33:25the codebase more easily and check
961. 33:28things faster, work on things more
962. 33:29confidently and just grow. And it was a
963. 33:33few months ago when I was working on T3
964. 33:34code, I was seeing how much time was
965. 33:36spent in our CI and how much RAM and CPU
966. 33:39is spent on my machines typeing the
967. 33:41code. Even with the TypeScript Go
968. 33:43rewrite, T3 code is just a massive
969. 33:45project and it took a lot of CPU time to
970. 33:48check. Previously, when you end up in
971. 33:50those scenarios, the amount of effort it
972. 33:52would take to make things better was so
973. 33:55high and the amount of value that people
974. 33:57would get was so low that it was rarely
975. 33:59worth it. For example, when I was at
976. 34:01Twitch working on this giant repo, we
977. 34:03could have rewritten TypeScript and Go
978. 34:05then. We had a bunch of incredible Go
979. 34:06engineers at Twitch. absolutely could
980. 34:08have dedicated like 20 people to
981. 34:10rewriting TypeScript in Go before the
982. 34:12TypeScript team even thought to do it.
983. 34:13But that would have been an absurd
984. 34:15undertaking that would not have been
985. 34:16worth our effort or time to unblock a
986. 34:19measly 600 to 800 front-end engineers
987. 34:21from having to wait as long for their
988. 34:23dev servers to start. Now I am rewriting
989. 34:25TypeScript in Rust so that a threeperson
990. 34:27team can move slightly faster. The
991. 34:30costbenefit analysis there was just like
992. 34:34it was really difficult but we still
993. 34:35would do it. I remember one of the
994. 34:37changes we made in order to like more
995. 34:38heavily adopt hooks in the codebase, we
996. 34:41had to move from Apollo GraphQL 2 to
997. 34:43three on the client. This resulted in a
998. 34:45bunch of legacy stuff that Apollo had
999. 34:47removed that we were using heavily
1000. 34:48across Twitch. So what we ended up doing
1001. 34:50is we made one shared PR with the move
1002. 34:53to Apollo 3. We took every single
1003. 34:56TypeScript file that was broken as a
1004. 34:57result, put them in a Google sheet, and
1005. 35:00then would go through the sheet and had
1006. 35:02volunteers take chunks of this sheet
1007. 35:05that was sorted alphabetically and
1008. 35:07assign them to ourselves. And it was our
1009. 35:09job to manually go through each of these
1010. 35:11files and fix the type errors and then
1011. 35:14test and make sure that the code
1012. 35:15actually worked. This was for thousands
1013. 35:17of files across this gigantic codebase.
1014. 35:19We literally had an Excel spreadsheet,
1015. 35:20well Google sheet that we would share to
1016. 35:23figure out who owned what to get all of
1017. 35:25this through cuz that's what it took to
1018. 35:28make a change to our GraphQL client
1019. 35:32connection. This is the one package that
1020. 35:34lets us define the GraphQL syntax to get
1021. 35:36the data. And this took a team of like
1022. 35:38eight people plus a bunch of assistants
1023. 35:40like a month plus to do. That's one
1024. 35:43prompt now. And this is what's
1025. 35:45so cool. Historically, gardening work
1026. 35:48was only really worth it if you were
1027. 35:50shipping at such a scale and the
1028. 35:52friction was so high that you could
1029. 35:54justify the insane investment necessary
1030. 35:57to do it. And this only happened at big
1031. 36:00companies that were architected
1032. 36:02correctly to split between platforms and
1033. 36:04product. There's a lot of companies that
1034. 36:05are right now because they didn't
1035. 36:07separate product and platforms properly.
1036. 36:09And now the product engines are stuck
1037. 36:11doing platform work. they don't
1038. 36:12understand well enough and they just
1039. 36:14cannot keep up or move fast enough. But
1040. 36:16the calculus is very different. Now,
1041. 36:18I've been hinting at the fact that
1042. 36:19agents make it way easier to do these
1043. 36:21types of changes, to make these types of
1044. 36:23wins. That's just one side of the
1045. 36:25equation. The other side is arguably
1046. 36:27even more interesting. All of the things
1047. 36:29I've talked about here that were
1048. 36:30blocking me or slowing me and my
1049. 36:32teammates down were annoying, but they
1050. 36:35would like maybe like best case if I
1051. 36:37could have filed 10 poll requests a week
1052. 36:39with these things being broken, fixing
1053. 36:41them could get me up to like 12 to 15.
1054. 36:43Maybe your agents get blocked much
1055. 36:45harder by these things. If it takes 5
1056. 36:48minutes to spin up the dev server to
1057. 36:51even check if the code works or not,
1058. 36:54your agent is going to be blocked
1059. 36:55insanely deeply by that. Hell, if you're
1060. 36:58using five minute TTLs, you're also
1061. 37:00going to be wasting a ton of tokens
1062. 37:01because you're not going to be hitting
1063. 37:02cache reads when you wait that long to
1064. 37:04validate a change. If your CI is too
1065. 37:07slow or it takes too much RAM and CPU to
1066. 37:09use, you become heavily limited by how
1067. 37:12much you can parallelize the work. If
1068. 37:13your release pipeline is too slow and
1069. 37:16you're scared of doing a release because
1070. 37:18it might take too long or worse, there's
1071. 37:20something bad in it, then the next
1072. 37:21release to fix it takes even longer,
1073. 37:23then you're not going to be hitting
1074. 37:24merge as much. The constant maintenance
1075. 37:27of the platform has never been more
1076. 37:30important because not doing this blocks
1077. 37:32your agents even harder than it blocked
1078. 37:34us. And all the work that you can
1079. 37:36possibly do to trim these pieces of
1080. 37:40friction, to kill these weeds, to get
1081. 37:41these problems out of your
1082. 37:43garden is so important. And there's one
1083. 37:46really big reason why that I've been
1084. 37:48dancing around. Slop grows faster than
1085. 37:51weeds do. If you are not constantly
1086. 37:55maintaining your garden, you will very
1087. 37:57quickly come back to a garden that is
1088. 37:59composed primarily of weeds. This is the
1089. 38:02problem that is destroying so many
1090. 38:05things. And this is what our job
1091. 38:07actually is going to be. We have to
1092. 38:09spend a ton of time making sure we're
1093. 38:10putting the right fertilizer in the
1094. 38:11garden. We have to spend a lot of time
1095. 38:13making sure that once the plants start
1096. 38:14to sprout that we put the right post up
1097. 38:17and like wrap the right material around
1098. 38:20the plant so that it keeps growing
1099. 38:21vertically and doesn't fall over and
1100. 38:23break in a terrible way. And you have to
1101. 38:26be removing the weeds as soon as you
1102. 38:27notice them appearing. Just to give an
1103. 38:30example of the type of work I'm talking
1104. 38:31about here, here's a PR that I was
1105. 38:33particularly proud of. This PR involved
1106. 38:36me deleting around half of all the unit
1107. 38:39tests in T3 code because the more I was
1108. 38:42looking at them and the more I was
1109. 38:44working in the codebase, the more I
1110. 38:46realized how many of these tests are
1111. 38:48useless and more importantly how much
1112. 38:50the useless tests multiply and result in
1113. 38:53more and more unnecessary edge cases
1114. 38:55being tested for no good reason
1115. 38:57whatsoever. And that's why I was able to
1116. 38:58delete 250,000 lines of code from the
1117. 39:01codebase in one PR. And we've had no
1118. 39:03regressions that any of the deleted
1119. 39:04tests would have caught. I think of the
1120. 39:0611,000 tests, I brought back four of
1121. 39:08them total. This is a phenomenal example
1122. 39:11of gardening. This is the type of thing
1123. 39:14that is our job. Now, I just got a
1124. 39:16fantastic follow-up question. How are we
1125. 39:18preventing it from happening again? By
1126. 39:20monitoring it. We can put a lot of
1127. 39:22things and we made some guidance changes
1128. 39:24to the repo to the agents and things
1129. 39:27like that to steer agents away from
1130. 39:29making bad tests. But even just deleting
1131. 39:31all of the slop, killing all the weeds
1132. 39:32was enough to help a lot. But there were
1133. 39:34still bad tests popping up here and
1134. 39:36there. So whenever one of us noticed
1135. 39:38one, we would ask the person who
1136. 39:40contributed it to go find the thread
1137. 39:42where it happened and ask their agent
1138. 39:44why. What led to you adding this test?
1139. 39:46Why did you think this test was a good
1140. 39:48idea? and we would use the info it would
1141. 39:50give us to go to the root of the
1142. 39:52problem, whatever stem this weed grew
1143. 39:55from and kill that. And this is just
1144. 39:57like one of the types of examples of the
1145. 39:59gardening I'm talking about. If we
1146. 40:01scroll through the PR, we'll see a few
1147. 40:02others. This one in particular is one
1148. 40:04I'm really fond of. I was getting tired
1149. 40:07of the performance regressions I was
1150. 40:09seeing in the wire protocol for how we
1151. 40:11get data from your server on your
1152. 40:13machine where everything is running over
1153. 40:15to the client on your phone or the
1154. 40:17computer that has all of the data that
1155. 40:20you're actually interfacing with. I had
1156. 40:21put a ton of time into making this as
1157. 40:24efficient as possible so that it would
1158. 40:26be as little as like 5 kilobytes to load
1159. 40:28even a big thread. I had put a ton of
1160. 40:31work into this, but it kept regressing.
1161. 40:33After a week of improving the
1162. 40:34performance here, subtle small PRs that
1163. 40:37didn't seem like they were going to
1164. 40:38cause problems here did. So, I built the
1165. 40:41thread transfer impact tool, which is an
1166. 40:43action that runs on every PR that has a
1167. 40:46bunch of real event data from real
1168. 40:48codeex threads and cloud threads, runs
1169. 40:51it through a fully simulated endto-end
1170. 40:53pipe, and checks how much data goes
1171. 40:55through the pipe. and then it would
1172. 40:57print it in every PR and flag a bunch of
1173. 41:00big red X's when anything caused a
1174. 41:02meaningful regression whatsoever. This
1175. 41:04worked incredibly well for a bunch of
1176. 41:06reasons. It puts feedback in the PR that
1177. 41:09a babysitting agent will notice and
1178. 41:11address. It gives the developers and the
1179. 41:14reviewers clear feedback on whether or
1180. 41:16not this is getting regressed or not and
1181. 41:18if these metrics we care about are
1182. 41:19changing. But the most important thing,
1183. 41:21this is really silly. It gives the agent
1184. 41:24the information. This is a thing we care
1185. 41:25about and it should be optimized for. To
1186. 41:27go back to the garden analogy, this is
1187. 41:29like the wood stake you put in next to a
1188. 41:32plant. So, the plant will grow in that
1189. 41:33direction. I put a wood stake in with
1190. 41:36this action. This CI tool, this one
1191. 41:39additional check that actually takes
1192. 41:40like almost no time. I think it takes
1193. 41:42under 45 seconds to run. It might be 30.
1194. 41:44This one small check I added to RCI has
1195. 41:47been an incredibly useful wooden post in
1196. 41:50the garden, guiding the plants to grow
1197. 41:52up the right way and away from the weeds
1198. 41:55and the other things I don't want them
1199. 41:56near in the garden. Do you know how much
1200. 41:58fun it was to do this work? By the way,
1201. 41:59I know I'm a bit of a nerd about this
1202. 42:01protocol stuff, but I really like
1203. 42:03representations of data and finding the
1204. 42:06right set of data to send to and from
1205. 42:08like different platforms, product,
1206. 42:10feature, whatever. It's just so fun. And
1207. 42:13that side of my brain has never been so
1208. 42:16valuable. The vast majority of my work
1209. 42:18for T3 Chat after like month two was
1210. 42:21unblocking my team by improving the
1211. 42:23platform so everyone else could build
1212. 42:25the features that our users needed. Now
1213. 42:27I'm doing that with and for agents. I am
1214. 42:30building up the best possible code base
1215. 42:33to contribute to. That might have some
1216. 42:35parts that are tedious. It might have
1217. 42:36some checks that are strict, but it is
1218. 42:38the best godamn garden I've ever been
1219. 42:40lucky enough to see, much less work in.
1220. 42:43And I am constantly quadruple checking
1221. 42:46my agents to see where they're getting
1222. 42:47stuck, what weeds they're tripping over,
1223. 42:50what fertilizer they're most confused
1224. 42:51by, what posts they need to grow
1225. 42:54against. And if you don't think this is
1226. 42:55engineering, I don't know if you ever
1227. 42:57were an engineer, cuz I've never felt
1228. 42:59more like I was building a platform than
1229. 43:02I do today. And it's just so goddamn
1230. 43:05fun. Even if you don't love this type of
1231. 43:07like maintenance and platform work, I
1232. 43:09think you'll love what happens on the
1233. 43:10other side. Because that was the magic
1234. 43:12that made me go from enjoying doing this
1235. 43:15for the sake of doing it to almost
1236. 43:18feeling bad for any second I'm not
1237. 43:20spending doing this because I've seen
1238. 43:21how much impact it has. This is PRs
1239. 43:24merged per day over time since T3 code
1240. 43:26started. And my face is covering these
1241. 43:29vertical lines because of how vertical
1242. 43:30they are. This is what happens when you
1243. 43:33maintain your garden. Well, we went from
1244. 43:35pretty consistently like 2 to 20 PRs
1245. 43:38merged a day to a bad like weekend day
1246. 43:42where we're busy being 60 merged and
1247. 43:44active days being in the 400 plus range.
1248. 43:47This is because our garden is really
1249. 43:49healthy. This is what happens when you
1250. 43:51make a good garden. Don't tell me
1251. 43:54potato stumbled on the same
1252. 43:56goddamn analogy as me. Why do I learn
1253. 43:59this now? I didn't even see this post. I
1254. 44:02would have liked it if I saw it. Thank
1255. 44:04you, chat. But god damn. Apparently, I'm
1256. 44:06not as unique as I like to think. I
1257. 44:08don't think I'm that unique, but Lauren
1258. 44:09and I are too aligned recently. It's
1259. 44:11dangerous. Anyways, she seems to fully
1260. 44:13agree. Every team needs a gardener.
1261. 44:15Someone quietly watching the stream of
1262. 44:16PR flowing in the codebase, noticing the
1263. 44:18smells. The third is record this week.
1264. 44:20The lint suppressions creeping like ivy
1265. 44:22across your carefully planned garden.
1266. 44:24Steady hand tending the weeds that would
1267. 44:26engulf it if in slop if left unchecked.
1268. 44:29Yep. If I go any longer, this video is
1269. 44:31going to literally go on forever. So, I
1270. 44:33think I am going to stop it here. Let me
1271. 44:36know if you enjoyed this. And if you can
1272. 44:38fight the obnoxious commenters, cuz I
1273. 44:40know they're going to be obnoxious for
1274. 44:41this one. And if you did like this
1275. 44:42enough, let me know if I should do more
1276. 44:45things like this because I would love to
1277. 44:47talk more about the techniques I've been
1278. 44:49using to maintain my garden and make my
1279. 44:51codebases pleasant to work in. Some of
1280. 44:53them are working great. Some of them I
1281. 44:54don't know if they are yet or not, but
1282. 44:56it's just a really fun time to be trying
1283. 44:58to build these ways. I've never enjoyed
1284. 45:00my work more than I do today. And I hope
1285. 45:01some of what I shared here will help you
1286. 45:03enjoy it the same. I know that this
1287. 45:05whole world is changing really fast and
1288. 45:06that it is scary, but if you find the
1289. 45:08magic within it, it is also the most fun
1290. 45:11it's ever been. And if even one of you
1291. 45:13can now see the magic where you might
1292. 45:14have missed it before, then this was
1293. 45:15worth all the shitty comments I'm about
1294. 45:17to see. Until next time, peace nerds.