YouTube2Text

Merge Is the Bottleneck - Prasanna Sankar, Vorflux — Transcript

by AI Engineer · 3,463 words · 542 segments · language en · Watch on YouTube

Full transcript

  1. 0:05Hey guys,
  2. 0:06my name is Prasanna and I really
  3. 0:08appreciate the opportunity to speak in
  4. 0:10front of all of you guys. Today,
  5. 0:12AI has really accelerated leverage and
  6. 0:15you know, this is a story of leverage.
  7. 0:17I'll start with my own personal
  8. 0:18background.
  9. 0:19I'm a software engineer,
  10. 0:21was ranked number one coder in India
  11. 0:24when I was 18 years old during my
  12. 0:26college days.
  13. 0:27I won [clears throat] all these
  14. 0:28programming contests like top coder. I
  15. 0:30was ranked number one there.
  16. 0:33I was two times world finalist in both
  17. 0:34Google Code Jam and ACM ICPC. These are
  18. 0:39the you know, the top programming
  19. 0:40contests in the world and did really
  20. 0:42well there, represented India
  21. 0:45quite a few times. Now
  22. 0:47[clears throat]
  23. 0:48the models have gotten way better than
  24. 0:50any of us. They perfect score these
  25. 0:52programming contests. So it's an
  26. 0:54interesting time to be alive and you
  27. 0:56know, super exciting time. I want to
  28. 0:58talk a little bit about you know, my own
  29. 1:00background, right? Like my own journey.
  30. 1:01After college, I kind of like worked at
  31. 1:03Microsoft for a year, quickly realized
  32. 1:05you know, the bottleneck to
  33. 1:08impact on the real world or my own
  34. 1:11personal leverage
  35. 1:13was other people
  36. 1:15like working in big teams.
  37. 1:17[clears throat] And then there was the
  38. 1:18the alternate path. I came across Paul
  39. 1:20Graham's blogs.
  40. 1:22Silicon Valley was happening.
  41. 1:24Internet companies were getting a lot
  42. 1:26cheaper to start. So a lot more people
  43. 1:28like myself
  44. 1:30were able to access it.
  45. 1:32And Paul Graham's blogs gave me a path
  46. 1:34towards living a life of higher
  47. 1:36leverage. So I moved to the valley
  48. 1:38sometime around 2009 to
  49. 1:41you know, start building companies. At
  50. 1:43this point, I'm a three times YC
  51. 1:45founder.
  52. 1:47[clears throat] The most successful
  53. 1:48company that I founded was a company
  54. 1:50called Rippling, which is worth more
  55. 1:51than 10 billion today. More recently, I
  56. 1:53started a company called Warflex. It's a
  57. 1:55it's a part of the latest YC batch as
  58. 1:57well. We've raised around 15 million
  59. 1:58dollars.
  60. 1:59And it's a AI software engineer
  61. 2:03or AI coworker, the highest leverage AI
  62. 2:06coworker. So the way I got that is like,
  63. 2:08you know, post founding Rippling,
  64. 2:11I was just like building things for
  65. 2:12myself and I was building things of like
  66. 2:15higher and higher leverage. Something
  67. 2:17interesting happened in in in my world
  68. 2:19with the launch of
  69. 2:21Opus 4.5. I think it was like almost a
  70. 2:23year ago at this point. It was a magical
  71. 2:26moment.
  72. 2:27Software went from autocomplete in in
  73. 2:29the IDE
  74. 2:31to
  75. 2:32[clears throat]
  76. 2:33agents working by themselves for a few
  77. 2:35minutes at a time. So you could suddenly
  78. 2:37sort of move from completely skip the
  79. 2:39IDE and move into the terminal. That was
  80. 2:41seminal, you know, it was naturally
  81. 2:43obvious that, you know, that was just
  82. 2:44going to take off. Those few minutes are
  83. 2:46going to get longer and longer and
  84. 2:48longer.
  85. 2:49So
  86. 2:50[clears throat] you know, at the same
  87. 2:51time, these models started winning all
  88. 2:53these programming contests and it was
  89. 2:54kind of like fairly clear to me like
  90. 2:56they were better than me in just like
  91. 2:59raw reasoning.
  92. 3:00Especially, you know, but then the the
  93. 3:03the harnesses that won these programming
  94. 3:06contests were not a vanilla Claude code
  95. 3:09or a Codex harness. They spent a lot
  96. 3:12more on inference time compute, a lot
  97. 3:15more inference time compute to like, you
  98. 3:16know, access higher order intelligence.
  99. 3:19So that's what I wanted to build for
  100. 3:20myself, you know, the higher leverage
  101. 3:22version which spends a bit more tokens,
  102. 3:24but you know, saves a lot of time, you
  103. 3:27know, gives me like a lot more
  104. 3:28intelligence.
  105. 3:30So I started building Worx Flex for
  106. 3:32myself
  107. 3:34and, you know, the first obvious thing
  108. 3:36is like, you know, once agents start
  109. 3:37running beyond a few minutes into into
  110. 3:41hours at a time, especially if you're
  111. 3:43able to solve a bunch of these
  112. 3:44bottlenecks that that I'm going to talk
  113. 3:46about,
  114. 3:47you really want them running on the
  115. 3:49cloud. You don't want them running on
  116. 3:50your machine because, you know,
  117. 3:53that's where they belong, a task that
  118. 3:56runs hours at a time.
  119. 3:57So [snorts] I think that's the first big
  120. 3:59bottleneck that people need to solve,
  121. 4:01like, you know, cloud agents need to
  122. 4:02solve to get
  123. 4:04people moving from the desktop to to the
  124. 4:06to the cloud is a perfect environment
  125. 4:09setup. Like today cloud agents run like
  126. 4:12most cloud agents run blind. At best
  127. 4:14they run the one service that you're
  128. 4:17trying to change, but really like for
  129. 4:19for like good testing the reason why
  130. 4:22people sort of like do all this stuff on
  131. 4:24their local host is because
  132. 4:26there are a lot of other dependencies
  133. 4:28and you want to be able to run you know,
  134. 4:30you want to be able to almost like
  135. 4:31mirror the local host environment on the
  136. 4:33cloud to truly transport someone. So we
  137. 4:36spent a lot of tokens and money and time
  138. 4:39solving the machine setup bottleneck
  139. 4:42where you can just like onboard and we
  140. 4:44start a bunch of sub agents to like go
  141. 4:46set up raw EC two instance
  142. 4:49with your whole environment, wire up the
  143. 4:51ports,
  144. 4:53figure out the machine secrets, solve
  145. 4:55authentication, solve seed data and a
  146. 4:57bunch of these things, spend tokens
  147. 4:59upfront migrating the user to the cloud.
  148. 5:02I think that's really important. That's
  149. 5:03something that most cloud agents miss.
  150. 5:05So that's that's like one of the major
  151. 5:07bottlenecks to moving from desktop to
  152. 5:08the cloud.
  153. 5:10Once you solve that, what you realize is
  154. 5:13like if you you know, when I was
  155. 5:15profiling how people were spending their
  156. 5:17time including myself, a lot of the time
  157. 5:18was being spent on the planning phase
  158. 5:21and
  159. 5:23you know, [clears throat]
  160. 5:24like
  161. 5:25back then the planning agents
  162. 5:29were not spending as much tokens.
  163. 5:32The agents weren't spending as much
  164. 5:33tokens on the planning phase. You know,
  165. 5:35they usually come back within a minute
  166. 5:37or two and you were iterating a lot back
  167. 5:39and forth, back and forth on the
  168. 5:40planning phase. And you know, obviously
  169. 5:42if human time is being spent there,
  170. 5:44that's where I want to throw more
  171. 5:45tokens, right? So what you know,
  172. 5:46workflows we sort of scaled up
  173. 5:50inference time computed the planning
  174. 5:51stage. We launched a bunch of parallel
  175. 5:53explorer sub agents which sort of built
  176. 5:57built a design or an architecture of
  177. 5:59what needs to be done
  178. 6:00and you know, we sort of came up with a
  179. 6:03plan of plans which are like a lot more
  180. 6:05detailed and you know and then sort of
  181. 6:09produced a summary out of it to help
  182. 6:12solve the time spent by humans on the
  183. 6:15planning phase, right? So that that was
  184. 6:18like a major bottleneck that we were
  185. 6:19trying to solve. Then comes
  186. 6:21orchestration. People usually often like
  187. 6:24break down like a large task into a
  188. 6:25bunch of small tasks and like work hard
  189. 6:28trying to land each one of them.
  190. 6:30We are like default sub agent build, you
  191. 6:33know, once a plan is built, you know, it
  192. 6:34has a dependent it comes with a
  193. 6:36dependency graph and sub agents go do
  194. 6:39independent pieces of work, right?
  195. 6:41And then testing is like another major
  196. 6:43bottleneck. If
  197. 6:47[clears throat] if you still need to
  198. 6:48pull stuff to localhost to test like
  199. 6:51like you you can't transition to the
  200. 6:53cloud. So we built
  201. 6:55like
  202. 6:56a detailed testing framework where you
  203. 6:59know
  204. 7:00browser testing happens on the machine,
  205. 7:03end-to-end testing happens and
  206. 7:06you know the output is something that
  207. 7:08the that proves to the user that things
  208. 7:10are working, right? So for me like I
  209. 7:13really enjoy like looking at a video
  210. 7:14which proves that the that the new
  211. 7:18feature actually works or you know,
  212. 7:20sometimes like I really you know, it
  213. 7:22also gives me like a staging link where
  214. 7:25for that particular feature I can click
  215. 7:27around
  216. 7:28because it's anyways hosted on a
  217. 7:30machine, right? So you can just like run
  218. 7:32the whole app on that machine, click
  219. 7:34around, play with it and
  220. 7:38and see if it works. So that's the
  221. 7:40that's the way to like completely remove
  222. 7:42localhost as a requirement for testing.
  223. 7:44Post that code review becomes a big
  224. 7:46bottleneck. Often times people ask
  225. 7:50multiple models to do reviews for you
  226. 7:52and we bake that in inside the harness.
  227. 7:55You can customize the harness and many
  228. 7:57people often ask often build with Claude
  229. 7:59and get GPT to do their review
  230. 8:04and you can have like multiple reviewers
  231. 8:06looking at like different
  232. 8:08These are all sub agents. Um, so you can
  233. 8:12basically customize the harness which
  234. 8:14um solves a lot of the reviewing
  235. 8:16bottlenecks.
  236. 8:17Then comes the merge. Um, often in fast
  237. 8:21moving teams when people are agents are
  238. 8:23opening a lot of PRs in parallel, um you
  239. 8:26need something like a merge queue which
  240. 8:27we built
  241. 8:29which constantly like rebases, keeps
  242. 8:32keeps you mergable with master and like
  243. 8:34lands the PR by itself. So you have a
  244. 8:37bunch of these bottlenecks that needs to
  245. 8:39be solved before you can transport the
  246. 8:41user entirely into into the cloud and
  247. 8:45really enable long running, multi hour
  248. 8:47running cloud agents. So that was over
  249. 8:50flex 1.0 which we launched about
  250. 8:53about a couple of months ago, I believe.
  251. 8:55And that has sort of like really taken
  252. 8:58off and we learned a lot from.
  253. 9:00[clears throat] So the biggest thing
  254. 9:01that we realized was now the bottleneck
  255. 9:04changed. Now that long running agents
  256. 9:06are possible, we were able to run a lot
  257. 9:09more tasks, you know, if desktop coding
  258. 9:11sort of enabled maybe three to five
  259. 9:14things in parallel, long running cloud
  260. 9:16agents started enabling 20 30 things in
  261. 9:18parallel, right? So, you know, obviously
  262. 9:20it sort of provided a lot more leverage
  263. 9:22and now the bottleneck sort of changed
  264. 9:25from writing code to project management
  265. 9:28and you know, making sure like the stuff
  266. 9:32lands and is actually useful
  267. 9:35um in making sure the things that are in
  268. 9:38my head are actually started.
  269. 9:40You know, because I have a limited
  270. 9:41context window, once it sort of like
  271. 9:43exceeds um 10 tasks in parallel, it gets
  272. 9:47like really really hard. So the
  273. 9:49bottleneck moves into project management
  274. 9:51actually.
  275. 9:52The model advances of the recent months
  276. 9:54also point in this direction. The latest
  277. 9:57gains have come from orchestration. So
  278. 10:01you know, most of the latest models try
  279. 10:03to offload
  280. 10:05um the chunk of heavy lifting into sub
  281. 10:08agents and the main agent, you know,
  282. 10:11wants to act as a orchestrator so that
  283. 10:14you maximally use
  284. 10:15the maximally generate intelligence
  285. 10:18within a context window. So fresh
  286. 10:20context windows
  287. 10:22give a lot more intelligence budget and
  288. 10:25you know, that's the way things are
  289. 10:26pointing.
  290. 10:27So that enables
  291. 10:30a really nice project manager which
  292. 10:32starts and orchestrates a bunch of
  293. 10:33sessions and
  294. 10:36and lands things for you. So that's
  295. 10:38WorxFlex 2.0. WorxFlex 2.0 is basically
  296. 10:42a bunch of autobots. These are 24/7
  297. 10:45employees that work towards a goal.
  298. 10:48The the bottleneck actually changed from
  299. 10:53giving a task and getting getting it
  300. 10:54done to actually even starting the task.
  301. 10:57The one thing that we started noticing
  302. 10:59was there was a lot of incremental
  303. 11:02intent that you could just capture from
  304. 11:04the way things were happening.
  305. 11:06For example, like if a customer call
  306. 11:10ends with a bunch of to-dos, those are
  307. 11:12all intent that I don't need to like
  308. 11:14copy paste and start one at a time.
  309. 11:17So [clears throat]
  310. 11:18like the mere act of even like noticing,
  311. 11:21remembering and starting each one of
  312. 11:23them could just be automated.
  313. 11:26[clears throat] So most teams treat
  314. 11:28agents like a task queue and a task
  315. 11:30queue only moves as fast to the person
  316. 11:32feeding it. But if you actually look at
  317. 11:33what's in the queue, most of it doesn't
  318. 11:36need you. The end to end plus one, that
  319. 11:39intent is already there in the code
  320. 11:41base. The alert or the transcript, it
  321. 11:43doesn't need a genius to start it. You
  322. 11:45can you can sort of figure out from the
  323. 11:47customer transcript, this is what you
  324. 11:49want to start. You can figure out from a
  325. 11:51production alert
  326. 11:53what you need to solve. Those things
  327. 11:55need to be on autopilot. It just needs
  328. 11:57an owner
  329. 11:59and it needs to be like multiplayer and
  330. 12:01collaborative with the rest of the team
  331. 12:03who where the ownership might be
  332. 12:04distributed to make sure all these sort
  333. 12:07of parallel sessions actually land. This
  334. 12:09is the next phase where of WorxFlex
  335. 12:12um and even um coding agents in general
  336. 12:17where things are moving from single
  337. 12:18player to like truly multiplayer and
  338. 12:20moving from um reactive to proactive
  339. 12:24where the intents that are already in
  340. 12:26there are being captured and used. So I
  341. 12:28would say like you know there are two
  342. 12:29kinds of work there is zero truly zero
  343. 12:31to one work that's like the truly
  344. 12:33creative stuff where the customer can't
  345. 12:35really even ask for it it takes taste or
  346. 12:37judgment that's the part that's still
  347. 12:39human
  348. 12:40and you know I I like doing that stuff
  349. 12:43and the part that's going into the auto
  350. 12:45bots is the end to end plus one stuff
  351. 12:48where the intents are sufficiently
  352. 12:50present as a part of the signals that
  353. 12:52the company is getting. So that's the
  354. 12:54new bottleneck and that's what Warflex
  355. 12:552.0 um is all about solving. Now people
  356. 12:59have tried to automate it and have had
  357. 13:01scars um doing something fairly simple.
  358. 13:05So for example you know people have set
  359. 13:07up automations to solve all sentries.
  360. 13:09You tell it to solve all the sentries
  361. 13:10and it goes crazy like a flood of PRs
  362. 13:13just hit you that nobody has times to
  363. 13:15time to merge. It pings you all day like
  364. 13:18one at a time and it's noisy. It often
  365. 13:21just silences the alert and leaves out
  366. 13:23the root cause um and a token bill that
  367. 13:25bursts um while the agent runs for hours
  368. 13:28and something nobody wanted. It's a
  369. 13:30black box you can't trust it with
  370. 13:31anything that matters right?
  371. 13:33You know these are all the fears that
  372. 13:36that people end up with when they just
  373. 13:39like point um a fire hose of incoming
  374. 13:43events company events on on an agent and
  375. 13:47or an automation. That doesn't really
  376. 13:49work. People have burnt their fingers.
  377. 13:52What you what
  378. 13:53what you need is a bot which actually
  379. 13:55has common sense. Um think about a new
  380. 13:57hire you don't fully trust yet. You give
  381. 13:59them important things first. You expect
  382. 14:02them to check in a lot early. You give
  383. 14:05them more as their work holds up and
  384. 14:07less when it doesn't. They work inside a
  385. 14:09budget and you can see what they do.
  386. 14:11Nobody hires a person and says solve
  387. 14:13everything.
  388. 14:14That's the bar that we built um for
  389. 14:17Warflex 2.0 and it's the next rung on
  390. 14:19the ladder. 1.0 finishes the work that
  391. 14:22you start.
  392. 14:23it. It gives you autobots, 24/7
  393. 14:26coworkers that live in You give an
  394. 14:27autobot a goal
  395. 14:29and a home channel. It stays plugged
  396. 14:31into your webhooks and starts the work
  397. 14:33itself.
  398. 14:35We even built a directory of opinionated
  399. 14:37bots, right? And you can you know start
  400. 14:40from there and change it. That's the big
  401. 14:42thing, you know people can't imagine
  402. 14:45these bots. So we've built a directory
  403. 14:47and we've sort of made sure
  404. 14:50the ones that we like and works for us,
  405. 14:53it's available out of the box with my
  406. 14:56taste and judgment to you.
  407. 14:59So so one [snorts] of some of the most
  408. 15:00you know I'll just walk you through some
  409. 15:02of the most well used bots. We have an
  410. 15:05on con on call bot that triages sentry
  411. 15:08and page duty alerts and investigates
  412. 15:10before an engineer is looped in. A CI
  413. 15:12bot that keeps speeding up your builds.
  414. 15:15A project manager bot which we started
  415. 15:18using a lot which holds a whole project
  416. 15:20in one channel and reports progress and
  417. 15:22blockers every day.
  418. 15:24This is super cool because it can it can
  419. 15:26make progress on things. It can tag
  420. 15:28people for reviews once
  421. 15:31some work needs to be reviewed. It can
  422. 15:33make your project run on
  423. 15:36on insane speed.
  424. 15:37One of the nice autobots that we have we
  425. 15:40have is like it it watches customer
  426. 15:42calls.
  427. 15:43For example, you know I was rolling out
  428. 15:46our beta to a friendly customer and the
  429. 15:48bot had a bug. It didn't respond back in
  430. 15:50the Slack channel right after. So you
  431. 15:52know normally I would have taken a
  432. 15:53screenshot and tagged Warflex
  433. 15:57to make it make it go work on the bug.
  434. 15:59This time I just you know after the call
  435. 16:01ended the autobot just read the
  436. 16:02transcript and found made to do located
  437. 16:05on the screen at that timestamp because
  438. 16:07you know it also has the you know video
  439. 16:09alongside the screen share alongside.
  440. 16:12And it it checked with me whether it
  441. 16:15wants me to fix it and I said yes. it
  442. 16:18tagged the right engineer who who don't
  443. 16:20the fix and you know came back with the
  444. 16:23PR. The engineer approved it. It it got
  445. 16:26merged.
  446. 16:28It got deployed. It closed the item and
  447. 16:30it even drafted a follow-up to the
  448. 16:32customer. I approved the follow-up and
  449. 16:36it went it went to the customer that it
  450. 16:38was fixed, right? So that that's the
  451. 16:40kind of leverage, you know, you just
  452. 16:41like I think the future is just like
  453. 16:43approve approve approve. A lot of these
  454. 16:44routine stuff is like you you just
  455. 16:46approve it and even that like a bot sort
  456. 16:48of like learns
  457. 16:51how much confidence it has and like
  458. 16:53keeps moving forward and over time it
  459. 16:54gets better and better like an employee.
  460. 16:56So
  461. 16:57[clears throat]
  462. 16:58one one thing for this room on why it
  463. 17:00works now, the latest model training
  464. 17:02runs focused on orchestration, a main
  465. 17:04agent that plans and keeps track with
  466. 17:06heavy lifting pushed down to sub agents.
  467. 17:08Most frontier models can now solve a
  468. 17:11well-defined task and their output is
  469. 17:13hard to tell apart. The next level up is
  470. 17:15a manager, an agent that starts sessions
  471. 17:17and keeps track of them. That's an
  472. 17:18autobot. It also runs on what you
  473. 17:20already have. Most teams bring their
  474. 17:23cloud code or codex subscriptions.
  475. 17:25Larger API users can route the well
  476. 17:28specified work to open models and keep
  477. 17:30the frontier models for judgment and
  478. 17:32communication. That way you can save a
  479. 17:33lot of money. Obviously the subscription
  480. 17:36save you a lot of money or if you're not
  481. 17:38eligible for it, the open source models
  482. 17:40save you a lot of money. What we built
  483. 17:42is an agent that decides what to work on
  484. 17:43next inside the limits you set and shows
  485. 17:46you why. Ultimately if you follow
  486. 17:48maximizing the leverage per person to
  487. 17:51its end, you get a company that runs on
  488. 17:54loops. A customer call ends and the
  489. 17:56product gets better. An alert fires and
  490. 17:59someone already knows why. CI slows down
  491. 18:02and a tested fix is waiting for review.
  492. 18:06The people spend their day on zero to
  493. 18:07one and on building the next loop. Our
  494. 18:10not star is a zero employee company. Our
  495. 18:13not star tells you which way to walk.
  496. 18:16It doesn't say that you arrive. 10
  497. 18:18people driving a huge business rounds
  498. 18:20down to
  499. 18:21We're early.
  500. 18:21Autobots will get things wrong. Some
  501. 18:23responsibilities are not ready to hand
  502. 18:25over,
  503. 18:26but we're building the next LEGO block,
  504. 18:29you know, the larger LEGO block which is
  505. 18:32assembled with the underlying model.
  506. 18:34Tesla took ground one intervention at a
  507. 18:36time where the way they solved autopilot
  508. 18:40is by profiling
  509. 18:42which
  510. 18:43which areas or scenarios the model made
  511. 18:46mistake. And that's what we do. We
  512. 18:48profile where humans spend their time
  513. 18:51and what the bottlenecks are and figure
  514. 18:53out how to throw tokens and intelligence
  515. 18:56at the problem. We climb the leverage
  516. 18:59one rung at a time as a new bottleneck
  517. 19:00shows up. We
  518. 19:02try to We are probably in a super
  519. 19:04intelligence takeoff and the
  520. 19:07opportunities to rewrite much of how the
  521. 19:09world runs are like insanely high.
  522. 19:11Companies build record revenue
  523. 19:14in super small amount of time. The
  524. 19:16growth rates are incredible, something
  525. 19:19that I've never seen all my life. But
  526. 19:20also it's like one of the most
  527. 19:22competitive spaces ever.
  528. 19:26The teams that tend to win tend to be
  529. 19:29like really small with
  530. 19:31everyone having a lot of ownership and
  531. 19:33leverage. It's It's not the traditional
  532. 19:35incumbents are not able to adapt and
  533. 19:37like
  534. 19:39and compete.
  535. 19:40There's a land grab and it goes to small
  536. 19:43teams with taste, judgment, and an
  537. 19:46insane amount of leverage per person.
  538. 19:48Own the zero to one and take as much
  539. 19:50ground as you can in this land grab.
  540. 19:52Thank you so much, guys. Really
  541. 19:54appreciate the opportunity speaking with
  542. 19:56all of you guys today. Thank you.

About this transcript

This page contains the full transcript of Merge Is the Bottleneck - Prasanna Sankar, Vorflux by AI Engineer, generated from the public captions YouTube serves with the video. The transcript has 3,463 words across 542 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.