Everything we knew about software has changed — Theo Browne, @t3dotgg ​ — Transcript

Full transcript

1. 0:01[music]
2. 0:17>> Hello, hello. Fantastic to see you guys
3. 0:19here. I still can't believe they're
4. 0:21letting me take a stage at something
5. 0:23like this, a YouTuber apparently, but
6. 0:26can't wait to share a bit about how I've
7. 0:27been thinking because if I'm being real,
8. 0:31kind of going through some AI psychosis.
9. 0:33Who here would classify how they feel
10. 0:34right now as some form of AI psychosis?
11. 0:37I want to see some hands.
12. 0:39The Those who don't have their hands up
13. 0:41yet, don't worry. We'll get you there by
14. 0:42the end of this talk if I do everything
15. 0:44right.
16. 0:45In order to talk about this, I want to
17. 0:46start with a bit of a personal journey
18. 0:49of my own. And I'm going to go through
19. 0:50this the way anybody does in modern
20. 0:52timelines,
21. 0:53with the models.
22. 0:54Who here used Sonnet 3.5 when it was the
23. 0:57creme de la creme, the cream of the crop
24. 0:59model available to us?
25. 1:01It was unbelievably better than what we
26. 1:03had used before, right? Like having used
27. 1:05all of these different models and trying
28. 1:07them in tools, Sonnet 3.5 was a big
29. 1:09moment for me because it felt like these
30. 1:11models could suddenly complete much more
31. 1:13end-to-end tasks, like actually get real
32. 1:16work done that takes multiple steps.
33. 1:18And then we got Opus 4.5. Who felt or
34. 1:21I'll go a different way differently
35. 1:23here. Who didn't feel a big jump when
36. 1:25they switched over to Opus 4.5?
37. 1:28That's a relief. There were not too many
38. 1:29hands because Opus 4.5 is probably when
39. 1:31my psychosis started in November and
40. 1:33December of last year. Having a model
41. 1:35that couldn't just write the code and
42. 1:37call tools, but could go way further. A
43. 1:39model that could test the work and
44. 1:41actually get it into a good state and
45. 1:43complete tasks that take hours instead
46. 1:45of minutes. It was unbelievable.
47. 1:48And then we got Mythos.
48. 1:49Who here has had a chance to play with
49. 1:51Mythos and Fable so far?
50. 1:54We agree it's a pretty damn good model,
51. 1:55right?
52. 1:57But why? It's not just better at coding.
53. 1:59If you handed a prompt that you would
54. 2:01have handed to these other models
55. 2:02before, it's not going to feel that
56. 2:04different.
57. 2:05I think of these almost as eras now,
58. 2:07where Sonnet 3.5 is the tool call era.
59. 2:09Not that it was the first model that
60. 2:10could do tool calls, rather it was the
61. 2:12first one that did them consistently and
62. 2:14reliably enough in context of a code
63. 2:17base where you could use this for
64. 2:18day-to-day coding work.
65. 2:20Then we got Opus 4.5, which was able to
66. 2:22do much longer running tasks without
67. 2:24losing track of what it's working on.
68. 2:26It's no longer, "Okay, build step one."
69. 2:28and then it does it. Then you say,
70. 2:29"Okay, can you build this next part?"
71. 2:30and then the next part. You can just
72. 2:31tell it what you want, and it could
73. 2:33figure it out a lot of the time.
74. 2:35Mythos is another jump
75. 2:37to orchestration.
76. 2:39It feels to me like it's the first model
77. 2:40that doesn't just understand your code
78. 2:42base, but it understands itself. And it
79. 2:44knows how to spawn additional models and
80. 2:46break up work in a way where it could be
81. 2:48completed more reliably and then
82. 2:50verified afterwards. And if you tell the
83. 2:52model to do that, it will just do it.
84. 2:53You don't need some custom tooling, some
85. 2:55custom system, some fancy software
86. 2:57factory. You just need to prompt it to
87. 2:59go a little further. I think you'll be
88. 3:00surprised how far it can go.
89. 3:03What I'm trying to say here is we need
90. 3:04to go bigger.
91. 3:06You're not going to see the benefits
92. 3:08going forward if you're not pushing the
93. 3:09model further, you're not pushing
94. 3:10yourself further with what you're
95. 3:11building. Most of the Jira tickets I
96. 3:13closed in my previous job could be
97. 3:15trivially solved with a model like Opus
98. 3:174.5.
99. 3:18My previous work would not benefit from
100. 3:21a model like Mythos.
101. 3:22If the models are going to keep getting
102. 3:23better, and at this point I'm confident
103. 3:25saying we they I was wrong when I
104. 3:27claimed that we were hitting a wall
105. 3:28before.
106. 3:29The models are getting better faster
107. 3:30than we are. So, we can't necessarily
108. 3:32get better. So, instead we have to go
109. 3:34bigger.
110. 3:35In order to do that, we have to get over
111. 3:36ourselves. This was really hard for me
112. 3:39as someone who spent a long time writing
113. 3:41software. Who here has written code for
114. 3:44more than 10 years? I want to see hands.
115. 3:46That's the majority of the people here.
116. 3:49I don't want to think about how long
117. 3:50I've been writing code for. But I have
118. 3:52been building up all of these strong
119. 3:53opinions since I started. I was using
120. 3:56GNU screen and eventually tmux back in
121. 3:58the day. I learned how to use those
122. 4:00tools and SSH and get before I even
123. 4:01wrote code. And those have all been
124. 4:03really ingrained in my workflow.
125. 4:06I think back to the old days in a weird
126. 4:09way. Hear me out. Let me talk about iOS
127. 4:11for a second.
128. 4:12Who owned an iPhone back when they
129. 4:14looked like this? iOS 6 or earlier?
130. 4:17How have you guys written code for 10
131. 4:18years when a fourth of you are that old?
132. 4:21I'm confused. Hopefully, you're all
133. 4:23Android people or something.
134. 4:25This is how iPhone apps used to look.
135. 4:28You might notice it's different from how
136. 4:29they look now. It looks less like an app
137. 4:31and more like somebody took a picture of
138. 4:33a compass and put it in the phone.
139. 4:35This is how apps used to look.
140. 4:38But, now they look like this.
141. 4:40And most people look at that and they're
142. 4:42like, "Oh, that's an obvious downgrade.
143. 4:44That's so much worse. Why would Apple
144. 4:46ever do that? This is the downfall of
145. 4:47Apple and the beauty of their design."
146. 4:49I'm going to fight you guys on that.
147. 4:51iOS 7 was Apple moving away from trying
148. 4:55to convince you that these devices can
149. 4:57replace the old tools we used to rely
150. 5:00on. The compass had to look like a
151. 5:01compass because the compass had to
152. 5:03replace the physical compass that you
153. 5:06relied on.
154. 5:07The books app had to look like a
155. 5:09bookshelf with real pages that turned
156. 5:11because I had to convince you it was a
157. 5:12reasonable alternative to buying a book
158. 5:15and reading the paper version.
159. 5:17Apps had to be designed to convince you
160. 5:19to use them, not to be useful.
161. 5:22And iOS 7 represents the shift to not
162. 5:26focusing on convincing you anymore.
163. 5:28Apple won. By that point, everyone knew
164. 5:30their iPhone could do all of these
165. 5:31different things. The point of iOS 7 was
166. 5:34to stop convincing and start embracing.
167. 5:37Start making a better interface. And
168. 5:38this interface, as much as we might not
169. 5:39like how it looks, it's so much more
170. 5:41useful. You have clear indications of
171. 5:43the difference between what direction
172. 5:44you're you're on and where you're
173. 5:46currently pointing. That red block is a
174. 5:47super nice way to know that you're not
175. 5:49in the direction you intend.
176. 5:51The current direction you're facing is
177. 5:53way clearer, too, with the giant 228 at
178. 5:55the bottom. You just get way more info
179. 5:58here than you did before. It's so much
180. 6:00clearer. Even if we don't like it
181. 6:02because it's not the thing we're used
182. 6:04to,
183. 6:05we got over it.
184. 6:08We're currently in our skeuomorphic
185. 6:09phase as software developers.
186. 6:11Skeuomorphism is this design aesthetic
187. 6:14trying to represent the way things used
188. 6:16to look, the physical goods that we
189. 6:18relied on, and try to make them digital.
190. 6:21We're doing this right now with
191. 6:22software.
192. 6:23We're pretending our terminals are the
193. 6:25ultimate interface when they're not even
194. 6:26good interfaces. And I'm saying this as
195. 6:28someone who loves their terminal deeply.
196. 6:30Natural language has no place in a
197. 6:32terminal, but we pretend it does because
198. 6:34the terminal's familiar. It's what we
199. 6:36It's what we're used to. It's what we
200. 6:37love. It's where we like to think of
201. 6:39ourselves when we're thinking about
202. 6:40coding.
203. 6:41Who here's an aspiring Vim user that
204. 6:43like wishes you could use Vim, or even
205. 6:46does?
206. 6:47I know we've all had that Vim phase
207. 6:48where we all tried.
208. 6:51This is just how we are as devs. We care
209. 6:52so deeply about these things. We care so
210. 6:54deeply about our tools, the systems we
211. 6:56pick, the frameworks, the languages.
212. 6:58We like to think it all matters.
213. 7:00And we've blinded ourselves in this. We
214. 7:03think things that just don't make sense
215. 7:05when you take one step back. Like, why
216. 7:07can't we commit our environment files?
217. 7:10It sounds stupid when I put it on a
218. 7:11slide like this, but I want you to
219. 7:12really think about this for a second.
220. 7:14When I have a team of engineers that are
221. 7:16working on a project, why do I have to
222. 7:18build another system to share this
223. 7:20specific file, but all the other files
224. 7:22can go and get just fine?
225. 7:24It's dumb. It's just how Git was built
226. 7:26because it was built for a very specific
227. 7:27thing, and then it took over our
228. 7:28industry, and it took over our brains,
229. 7:30and we aren't letting go of that.
230. 7:33There's a lot of these things in our
231. 7:35heads that we have to start fighting. We
232. 7:38have to take the step back and think,
233. 7:40is this how we do things cuz it's right,
234. 7:42or is this how we do things cuz it's
235. 7:44just how we've always done it?
236. 7:46And as you start to think more about
237. 7:47this, you'll realize there are so many
238. 7:49things that we do this with.
239. 7:51Like, why do we qualify ourselves by the
240. 7:53languages we know?
241. 7:55I used to think this was a junior thing.
242. 7:56Like, I could can't tell you how many
243. 7:57times I had a junior engineer I was
244. 7:58talking to who was like, "Oh, you're a
245. 7:59coder? What languages do you write? I
246. 8:01write JavaScript."
247. 8:02I thought this was a junior problem, but
248. 8:04then you talk to senior engineers who
249. 8:05are like, "Oh, he writes JavaScript.
250. 8:06He's not a real developer."
251. 8:08We care too much. We pride ourselves in
252. 8:11these things. They're our identity.
253. 8:12These weird facts, these weird choices,
254. 8:15these things that feel essential
255. 8:18just don't matter that much anymore. And
256. 8:20they didn't then, and they matter less
257. 8:22now. We got away with it because it was
258. 8:23so hard to find engineers that we could
259. 8:26just tell the company what we were
260. 8:27doing, and they couldn't really say no
261. 8:28cuz the alternative is spend 6 months
262. 8:30trying to hire someone else. They're not
263. 8:31going to do that.
264. 8:33And along that note,
265. 8:35why are we so scared of deleting code? I
266. 8:37cannot tell you how many times I've been
267. 8:38in a conversation with someone
268. 8:41where the solution is to just delete it
269. 8:43and reset, but we have such a bad sunk
270. 8:46cost mindset in this industry. We care
271. 8:49so much about the code we wrote, and we
272. 8:51care so much about it still being there
273. 8:54that I feel bad working with my team
274. 8:56sometimes when somebody files a PR that
275. 8:57isn't quite the right solution, but they
276. 8:59spent a week or two on it. Like, who
277. 9:01here has guilt merged a PR before? Where
278. 9:03you just felt bad because somebody put a
279. 9:04lot of work in, it's not quite the right
280. 9:06thing, but you merge it anyways cuz the
281. 9:08alternative was a conversation you
282. 9:09didn't want to have.
283. 9:12Why do we do this to ourselves? One of
284. 9:14the nice things about ages you don't
285. 9:15have to feel bad when you shut down
286. 9:17their work,
287. 9:18but we we just care too much. It's the
288. 9:20point I'm trying to make. And the things
289. 9:22we care about are not necessarily the
290. 9:23things that matter anymore, and I hope
291. 9:25we can finally start to challenge some
292. 9:27of these.
293. 9:29Going to get a little more personal here
294. 9:31by showing some of the ideas I've built
295. 9:33cuz the goal here is that when you guys
296. 9:35go out of this talk,
297. 9:36you have a better mindset for coming up
298. 9:39with ideas that make sense now by
299. 9:41rejecting the things that made sense
300. 9:42before.
301. 9:44These are three of the things that I
302. 9:45have built or are currently working on.
303. 9:47We're going to go from the bottom up.
304. 9:49I built a Reddit scraper because making
305. 9:51good memes is hard and I would rather
306. 9:52just steal them from Reddit and went
307. 9:54pretty well. It was a side project. It
308. 9:56would take me two to three days. Would
309. 9:57just scrape Reddit, top posts on
310. 9:59programming humor,
311. 10:01put them in a nice format for me so I
312. 10:02could go copy paste them onto Twitter.
313. 10:04Zoom for streamers was a startup I went
314. 10:06through Y Combinator with. It was called
315. 10:07Ping. I wanted to make it easier for
316. 10:09live content creators to do high-quality
317. 10:12collaborations in the software they
318. 10:13already used, OBS.
319. 10:15The full stack cloud is
320. 10:17let's just imagine Vercel, but it goes
321. 10:19further each direction. They have off
322. 10:20built in, but they also have databases
323. 10:22built in. These are all things I've
324. 10:23wanted that I benefit from existing,
325. 10:26enough so that I tried to build all of
326. 10:28them.
327. 10:29The bottom two I built in 2021. The top
328. 10:31one I'm working on right now.
329. 10:33These are also kind of tiers, different
330. 10:36levels that we can build at. If I was to
331. 10:39try and categorize them, I would call
332. 10:41the bottom one side project, call the
333. 10:43middle one startup, and call the top one
334. 10:45too big. It just doesn't make sense.
335. 10:48Well, this is how I would have
336. 10:49categorized this even just a year ago.
337. 10:52But things have changed.
338. 10:54Now that the models are bigger,
339. 10:56the tiers have shifted.
340. 10:58Everything is now one tier lower.
341. 11:02And this is a crazy thing for me to
342. 11:04process. The fact that what used to be a
343. 11:06startup is now a side project. In fact,
344. 11:08some of the startups I've talked to even
345. 11:10at an event like this,
346. 11:11their whole startup could have arguably
347. 11:13been a side project or this bottom tier,
348. 11:15which there's a weird gap there. What's
349. 11:17that?
350. 11:18It's the G brain tier. It's a markdown
351. 11:20file.
352. 11:21Do you know how many companies are at
353. 11:22this event where their whole product
354. 11:24could just be a markdown file?
355. 11:27It's insane.
356. 11:28>> [applause]
357. 11:28>> And like, okay, seriously though,
358. 11:32the fact that you can now execute
359. 11:33markdown by just piping it to Codex or
360. 11:36Claude is unbelievable and I think most
361. 11:38of us haven't fully appreciated how
362. 11:40insane that is. I had a service that
363. 11:42would triage all of my PRs, have them
364. 11:44all get reviewed with AI and then help
365. 11:46me prioritize. That service is a
366. 11:48markdown file now. I just literally
367. 11:49wrote like, go to these four GitHub
368. 11:51repos, look at all the open PRs, figure
369. 11:53out what the current status of the work
370. 11:55is,
371. 11:55and then help me prioritize it. And then
372. 11:57when you're done, go update the static
373. 11:59HTML file and send it to S3 and give me
374. 12:01the URL. And every morning at 9:00 a.m.
375. 12:03this runs on a cron and around 9:15 to
376. 12:059:20, my markdown file generates me my
377. 12:07work for the day.
378. 12:09What the hell?
379. 12:11How are we actually here now? I
380. 12:14Try that if you have it, by the way.
381. 12:15You'd be amazed how many of these types
382. 12:17of things can exist that are literally
383. 12:19just a markdown file running on a cron.
384. 12:21But what about Okay, two more things I
385. 12:24want to change about this, though.
386. 12:26First is
387. 12:28the full stack cloud. This is mine.
388. 12:30Don't do it. Like bit coming soon. Very
389. 12:31excited. I wouldn't want to compete with
390. 12:33me on this one. Trust me. It's going to
391. 12:34be really cool.
392. 12:36But there's still something else.
393. 12:38There's a gap here.
394. 12:40And I'm going to be real with you guys.
395. 12:41I don't know what goes in this gap.
396. 12:43I don't know what too big means anymore.
397. 12:46Is it training your own model from
398. 12:47scratch? Is it building your own
399. 12:49operating system? Is it trying to
400. 12:51compete with NPM and Node directly? I
401. 12:54don't know. I don't know what too big is
402. 12:56right now. And that's scary, but it's
403. 12:57also exciting. It means I need to keep
404. 13:00pushing myself to go bigger than makes
405. 13:01sense
406. 13:03in order to find these limits.
407. 13:06But what does that even mean? What does
408. 13:08it mean to think bigger in this
409. 13:10scenario?
410. 13:12I would argue that bigger's probably the
411. 13:14wrong word for most of how I'm thinking
412. 13:16here.
413. 13:17It's time to think wider.
414. 13:20What I mean by this is spectrum and I'm
415. 13:22sorry I have to do a diagram. If you
416. 13:23watch my videos, you understand.
417. 13:26There's breadth and depth to any piece
418. 13:29of software. The breadth is the range of
419. 13:31things that your software covers.
420. 13:33And the depth is the number of features
421. 13:35in a given area.
422. 13:36Let's look at a company like Vercel.
423. 13:39Vercel does not offer all of the
424. 13:40features that AWS does. They never will.
425. 13:42It doesn't make sense.
426. 13:44But Vercel offers deeper But Vercel
427. 13:46offers deeper features in the space
428. 13:48they're in, which is full-stack
429. 13:50front-end leaning servers.
430. 13:52If you're a front-end developer and
431. 13:54you're not using Vercel, you're feeling
432. 13:55some amount of pain because they're just
433. 13:57further ahead with this. So much so that
434. 13:59even the agents prefer it.
435. 14:00And this was kind of how you had to
436. 14:02build your startups. Because if you were
437. 14:03competing with a company like AWS,
438. 14:06you're never going to have all of the
439. 14:07features they have. You're never going
440. 14:08to cover the range that they cover.
441. 14:11It'd make no sense to try because you
442. 14:14don't have the thousands of engineers
443. 14:15they do doing that.
444. 14:18At least you didn't.
445. 14:21But now things have changed.
446. 14:22All of a sudden, that range is viable in
447. 14:26a way that it never was before. I'm not
448. 14:28saying you can build something as
449. 14:29reliable as RDS. I'm saying that you can
450. 14:31build a database platform into your
451. 14:33product in a day or two of work with
452. 14:35enough prompting and enough effort.
453. 14:37And if you build your stuff right, if
454. 14:39you play your cards correctly, and you
455. 14:40think about things the right way,
456. 14:42you'll realize that you can build enough
457. 14:45across a spectrum of things you care
458. 14:47about to enable most users to at least
459. 14:50start trying the thing.
460. 14:51And when they have features they need in
461. 14:53a given vertical that you don't support,
462. 14:55it's not your problem.
463. 14:56As long as you build it right. Because
464. 14:58they can build the features that are
465. 14:59missing themselves.
466. 15:01If you architect your systems and you
467. 15:02architect your products in such a way
468. 15:04that users can do things that they you
469. 15:06never would have guessed.
470. 15:08Like Slack accidentally did this because
471. 15:10Slack is now the platform people run
472. 15:12their agents in half the time, which is
473. 15:13crazy. Slack sucks. It's not a good
474. 15:14product, but it's the right shape for
475. 15:16people to build the features they want
476. 15:18into it through the somewhat functional
477. 15:21Slack bot APIs.
478. 15:24This is all crazy cuz I'm basically
479. 15:26sitting here telling you like it's time
480. 15:27to compete with Slack. It's time to
481. 15:28build your own AWS. It's time to
482. 15:30challenge Salesforce directly.
483. 15:32It sounds stupid, but I'm going to be
484. 15:34real.
485. 15:35If your idea doesn't feel stupid,
486. 15:37it's cuz your idea's not big enough.
487. 15:40That's all I have to say. Thank you so
488. 15:42much, AIE.