YT YouTube2Text

YouTube2Text → Transcripts → How To Build The Future: Aravind Srinivas

How To Build The Future: Aravind Srinivas — Transcript

by Y Combinator · 6,860 words · 928 segments · language zh-CN · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00we release the ability to ask follow-up
0:02questions that double the engagement
0:04time on the site and also increase the
0:06number of questions every day so I was
0:08like okay there's something here it's
0:09not worth killing and pivoting to
0:11Enterprise it was not like I want to go
0:13and kill Google like that sort of a
0:15motivation it was more like what is an
0:17idea of that scale and ambition is
0:19something like this today my view of
0:22perplexity is a more intelligent Google
0:25search that's really useful in certain
0:27scenarios what do you want me to think
0:29it it in 3 or 4
0:32[Music]
0:36years welcome back to another episode of
0:38how to build the future today we're
0:40joined by Arvin shavas co-founder and
0:43CEO of perplexity which in less than
0:45three years has grown to more than a N9
0:47billion valuation thanks for joining us
0:49thank you for having me David how did
0:51you get into this world I was pretty
0:53interested in AI deep learning research
0:56uh that's actually what got me into the
0:58US I was an underr in India came here to
1:01the US for doing my PhD here at Berkeley
1:03life really changed when I got to do an
1:05internship at open aai and Ilia s was
1:08there I still remember the day I first
1:10met him and I was very prepared and had
1:13all these fancy ideas that I thought
1:15were very interesting and he listened
1:17for 5 minutes and said all this resource
1:19is useless feels really bad to hear that
1:22so I got used to you know hearing the
1:25right things even if they're
1:26uncomfortable and then he told me the
1:28only thing that matters is he drew two
1:29cir circles one big circle called it
1:32unsupervised learning and then inside he
1:35said reinforcement learning another
1:36Circle and he said this is Agi every
1:40other research doesn't matter this was
1:42around the time when they were building
1:44gpt1 okay they didn't even call it
1:47gpt1 uh when I saw the research I went
1:50back to Berkeley and said I was working
1:52a lot on RL that was the rage at the
1:54time because of alpha go and deep mine
1:57right but that was kind of like chasing
1:59the friend so I went back to my
2:01professor and said hey we have to
2:02actually go and study unsupervised and
2:04generative models and generative AI so
2:07then I got into that and uh did more
2:10internships at Google and during my
2:12Google internship I stumbled upon this
2:14book called in the Plex so I would
2:17launch jobs during the day training runs
2:19and then um go and read these books in
2:21the library because interns don't have
2:23any other thing to do right and we feel
2:25amazing that oh like these guys actually
2:27Were Once Upon a Time grad students like
2:29me and now I'm working as an intern in
2:32their offices and reading reading the
2:33book book it's it feels nice it would be
2:36amazing to start a company like that in
2:39future where there's a lot of research
2:42there's a lot of like AI at the same
2:44time it's very grounded in product
2:46building uh it's very difficult to do
2:48that uh and I spent a lot of time
2:50thinking about it I even spoke to IIA
2:52about it and like where we said there
2:54are probably only two problems where uh
2:57you can work on AI and also build
2:59product at the same same time one is
3:01like search and the other is sub driving
3:03car because all your product roll outs
3:06are becoming data points for improving
3:08the underlying AI in the product and
3:10that'll make the product even better and
3:13that'll lead to more users and more
3:15usage will lead to more data points and
3:16it'll become a flywheel and uh it should
3:19also be on the AI completeness path it's
3:23sort of a buzzword to say this but
3:25basically what it means is better AI
3:27should keep making your products better
3:29so that that way you can keep working on
3:31your company until AI is solved right
3:34and once it's Sol okay sure we'll worry
3:35about all those you know imp but your
3:37company gets better as AI gets better as
3:39opposed to your company gets run over by
3:41somebody else exactly so searge is like
3:43one of those problems yeah so you're at
3:45this moment where you kind of have this
3:47realization that you want to start a
3:48company how did you get the kind of
3:51activation energy to quit your great job
3:53at open Ai and go do that how did you
3:55find your co-founders I came across this
3:57blog that um one of the
4:00former YC Partners Daniel gross wrote
4:02but it was like how to build the next
4:04Google um and I think basically the core
4:07idea is like you could do so much more
4:10with better query reformulation so you
4:13take a query and you just add some
4:15suffixes uh so if someone's looking for
4:17reviews of a movie uh just suffix site
4:20colon rodent tomatoes.com if someone's
4:22looking for uh reviews of some uh new
4:25Gadget uh do site call in that
4:27corresponding subreddit you can get away
4:29with a lot of these suffixes and like
4:32our special strings to like filter
4:34results and and already make Google uh
4:37so much better even with the existing
4:39Goog Google ranking uh I'm not even
4:42talking about the ads problem just
4:44simple ranking and then you can do more
4:46sophisticated things of like classifying
4:49queries and he was talking about how
4:51llms could automatically figure out
4:53these suffixes and I was pretty
4:55interested in that okay that that seemed
4:57like okay maybe uh generative AI might
4:59be um as in like I used to just call it
5:02llm so General models could be a better
5:04way to build search engines too I also
5:07was pretty interested in like trying to
5:08do something agent like when deep mine
5:11had this um Android environment that
5:14they built where like they they kind of
5:15wanted to prototype a mobile app using
5:18agent uh that knows when to use what
5:20apps and control the apps um that's when
5:23I I spoke to my uh co-founder in CTO uh
5:25his name is Dennis we had written the
5:27same paper a day apart so we we knew
5:29each other and he was a visiting student
5:30in my lab and we used to talk about I
5:32with brainstorm ideas of how like we can
5:34build agents to control the Android
5:36environment so we we were definitely
5:38chatting about doing a lot of things but
5:40never concretely about any company or
5:42product the first thing that anyone
5:44would tell you is like why why you work
5:47on this of course Google's going to do
5:49this right it's not even like you go
5:51build a better Google Docs uh Google
5:54will eventually do it because it's a
5:56secondary thing for them so companies
5:58like notion can still be funded this is
6:00their core crownwell so why would you
6:03even try I think the reason it actually
6:06made sense this is again after launching
6:09the product we realized this not before
6:12um so there's like some benefit to being
6:13ignorant ignorance is bliss uh is that
6:17if people stop clicking on links the ad
6:19economy kind of dies now you can there's
6:22a lot of you know Nuance to this but
6:24that core Insight was only uh realized
6:27by us after launching so so once we
6:29realized that I thought okay we were on
6:31to something and that kind of took us
6:34last two years yeah walk us through like
6:35the first iterations of your
6:37experimentation like I know you did a
6:39bunch of demos that were very dissimilar
6:41from perplexity yeah so I was like adash
6:43is enough to go and pitch to The in
6:46first seed investor of ours elot Gil
6:48that like hey like um you know I want to
6:51disrupt Google um but I kind of want to
6:54do it from pixels from a glass and I
6:57think that's way you know you're not
6:59competing with people typing on the
7:01search bar they just seeing even at that
7:03point you like knew in your mind I want
7:05to go after Google or like yeah yeah it
7:08was not like I want to go and kill
7:10Google like that sort of a motivation it
7:12was more like what is an idea of that
7:14scale and ambition is something like
7:16this it was also around the time when
7:17multimo models were slowly beginning to
7:19work so I thought like if you were on
7:21the trajectory of improving technology
7:23you could build something pretty amazing
7:25my investor rightfully said like not to
7:27work on it in the beginning so we
7:29focused more on searching over like
7:30specific verticals or data sets or
7:32databases tables actually and we were an
7:35Enterprise kind of focused company
7:37except like nobody wanted to give us
7:40their data I remember I used to you know
7:43hustle for calls with like bitbook or
7:45crunch Bas as because I kind of wanted
7:47to build a demo that would first make
7:49sense to an investor and that way we can
7:51keep you know raising some capital and
7:53then actually hiring good people and
7:55then go and like do the real thing and
7:57so crunch base add all this data which
7:59will as but they they just don't want to
8:00give it to us and so um next best step
8:04Twitter yeah Twitter pre Elon CEO
8:07moments uh academic access was allowed
8:10legal so we um um built a database of
8:13Twitter we organized it in the form of
8:15tables we Tred to do it with the openai
8:18Codex modeles this was even pre gbt
8:213.5 we we wrote a lot of templates oh
8:24for these kind of queries these are
8:25example sequels and then the model would
8:27do kind of rag they pull queries in the
8:30templates and then write the actual SQL
8:33based on the template sqls and that was
8:36the only way to get it to work reliably
8:38and then we had a lot of call backs in
8:40case errors happen it'll automatically
8:42correct it uh and then it will go and
8:44query the database and then retrieve the
8:47records it was very nice and it was a
8:49chat UI you could chat you could
8:51Converse you could plot and this was
8:53like the first real like product or demo
8:55that you guys watch yeah yeah we did it
8:57very fast like took only a month to do
8:58this um because there like three people
9:01only but that energy in the beginning is
9:03insane and they showed it to a bunch of
9:05people and they all allowed it main
9:07there are two reasons one something like
9:08that never existed before like uh you
9:11could never actually goarch exactly to
9:14this day even today right and then uh
9:17also people lowed I finding all these I
9:20think I think the social search of like
9:23knowing who other people are following
9:25whose streets are they liking whose
9:26streets are they not liking who did they
9:28unfollow this week mhm you know those
9:30kind of diffs it's all like funny so you
9:33launched this Twitter search thing yeah
9:35how did you transition from that to now
9:38what we all know as perplexity yeah so
9:39we had that right and then we were
9:41trying to do something similar to that
9:43for many different databases GitHub like
9:47if coders could go and search about Reb
9:49host or LinkedIn if you could like
9:52almost be a recruiter just say but even
9:54now it's pretty hard I want to say like
9:56I want all the people who worked uh who
9:58who have been C Founders you know who
10:01and and who also worked at C C or D
10:03startup because they would know what it
10:05means to be Scrappy it's very difficult
10:07to like using the LinkedIn UI to do that
10:09you cannot do that right for whatever
10:11reasons people don't want to give their
10:12data their pay wall like you know such
10:15technology if it exists we'll be
10:17creating like way more value but some it
10:20doesn't exist for many other reasons we
10:22were beginning to see how like even with
10:25the capability of the models at that
10:27time in 2022
10:29um pre 3.5 turbo uh things were actually
10:33pretty reliable to the extent where like
10:36people would like use this and find
10:38Value I actually read this um polyram
10:41tweet I think uh like where if you try
10:43to often when you when you figure out
10:46the better solution when you try to Sol
10:48solve a harder version of it but you end
10:50up with a simpler Solution that's more
10:51General and scalable that's what we
10:54realized like okay like there's one way
10:55of doing these things where we go to
10:57each of these domains and like try to
10:59build an index of it and put it into
11:01specific formats like tables and then
11:04have the
11:05llm like like read that uh in a
11:08structured language SQL or you could do
11:11the other way where you just keep it
11:12unstructured and expect the llm to do
11:15most of the work at the inference time
11:17at at at the time of the query and and
11:19don't do all this work in the indexing
11:21time and clearly we knew that if if
11:24second is where the world is headed uh
11:27where the models will get smarter and
11:28smarter
11:30it gives you an advantage to build it
11:32that way because it's it's more General
11:35and uh you also stand a chance against
11:37the Legacy system that Google has built
11:40which is a lot more in the first style
11:42so we thought okay we would try to build
11:44a more General solution and then we
11:45prototype this thing one weekend
11:47actually actually John Schulman's theme
11:50had already published this thing called
11:51Web GPT at the time so I was pretty
11:53aware of it opena I even had a bot when
11:56I worked there called the truth bot
11:58which John John bill with this steam
12:00okay where you could ask it a question
12:02and it'll go and search the web and then
12:04it'll give you an answer uh with some
12:06with some sources and um it was very
12:09slow and it was built with the 175b gp3
12:13model so incredibly slow and inefficient
12:15it was more agentic like it would
12:17actually be like an RL agent that
12:19decides if it wants to click on a link
12:21and browse it scroll okay this is very
12:23slow so what we tried is a very simple
12:26uh heuristic version but much faster
12:29which is okay you always take the topk
12:31links that a search API provides you you
12:33always only take the summary Snippets
12:35that the index is already cached so
12:38there's no scrolling there's no clicking
12:40and you always feed all those links into
12:42the prompt so there's no selection ask
12:45you to write a summary with sources in
12:47like the academic format and that's it
12:50when these models were're getting to a
12:52point like 3.5 turbo sort of models were
12:54beginning to come um this actually
12:57started working much better yeah
13:00instruction following capability
13:01increased enough that you didn't have to
13:03do it very very uh rigorously got so so
13:07you kind of did like the the dumb
13:09approach um betting on the fact that the
13:12the AI would get good enough that would
13:14make all of
13:15that uh right timing I would say one
13:18year ago and John and his team tried
13:20like the models were just so so much uh
13:23worse that like if you tried the dumb
13:25approach it just wouldn't work and so
13:27therefore you would conclude that you
13:28need a smarter approach okay but then
13:30when the modeles began to be much better
13:32instruction following the dumb approach
13:35actually works and that fixes a core
13:38product ux problem of latency you are
13:40used to uh like like links appearing
13:43instantly on a traditional search right
13:46even then by the way the first version
13:48we launched which is the answer version
13:50uh took 7 seconds or something to um
13:53because we didn't even have this concept
13:54of streaming answers we we would wait
13:57till the entire answer was generated we
13:59couldn't control the verbosity so
14:00sometimes the answer would be very very
14:02big we even had to hardcode a prompt
14:04saying only write five sentences or
14:06something like that or 80 words to keep
14:08it fast yeah exactly okay so you
14:10launched this when was the first moment
14:13that you thought like oh I'm on to
14:15something here so we tweeted it okay I I
14:18I was while writing the tweet I was like
14:20um you know people are going to ridicule
14:22it it's going to make mistakes blah blah
14:24blah first moment of virality came and
14:26one uh annoyed uh like intellectual
14:30academic came search for herself uh it
14:34said she it gave a biography in the past
14:37tense and she's like I'm still alive
14:39what the hell but actually what happened
14:42was there was a person with the exact
14:44same name and
14:46spelling uh who died and LM thought she
14:50died and she gave a it gave a past tense
14:52okay actually thought that was pretty
14:54clever reasoning on the modotto except
14:57it's not even higher order to know that
14:58they're different people so then that
15:01got us a lot of attention people were
15:02beginning to start thinking okay look
15:03the sources thing is good but can we
15:06really trust the answers these things
15:07are uh saying and then um that got into
15:12into this trend of people uh like
15:13searching for themselves this is
15:15something that keeps happening time and
15:17again with all consumer products when I
15:19got a chance to speak to Mike ker uh
15:21that vacation uh he said the same that
15:24even though you can click on your own
15:26profile icon and go go back to your
15:28photos people always love to go to their
15:31profile on Instagram by typing their
15:33username on the search bar it's such a
15:36human habit yeah so we a lot of people
15:39start putting their Twitter handles or
15:41social like usernames and then it would
15:43Mash all their activity across the
15:45internet including stuff they did in the
15:48childhood like many years ago and then
15:51give like this interesting summaries and
15:52they would screenshot it and share it
15:54yeah so I thought there was something
15:55there there something driving it that
15:57you but I still wasn't sure yeah and
15:59then we release the ability to ask
16:01follow-up questions that double the
16:03engagement time on the site and also
16:05increase the number of questions every
16:07day and number of people number of
16:09questions every day was increasing
16:10exponentially so I was like okay this is
16:13there's something here it's not worth
16:14killing and pivoting to Enterprise you
16:16have this like initial momentum and you
16:19you said earlier it wasn't until
16:20hindsight that you had the idea that
16:22like oh we actually have a chance of
16:23competing with somebody like a Google um
16:26when did that realization happen in this
16:27journey how' that go down so I never
16:29really thought about the Google
16:31competition in a serious way to be very
16:33honest um because I knew that like the
16:37they cannot make this exact product on
16:40the Google homepage it's so hard to know
16:43when a query is purely informational or
16:45not and then the Google search page is
16:47already like so cluttered that's the
16:49answer box the knowledge panel uh
16:52there's some ads there's some links
16:53there's you know like perspectives from
16:56socials all these social cards it's all
16:59too much information so that it's
17:01clearly like feels like you know fast
17:03food and like healthy meal difference
17:06for using Google and perplexity on even
17:08informational queries I was more worried
17:10about like Microsoft in the beginning uh
17:13because they were launching Bing chat in
17:15fact on the day we agreed to uh have a
17:18term sheet like hand Shook on a term
17:20sheet uh with with with with one of the
17:23Venture Capital investors Nea here uh in
17:25in San Road after like one week of
17:28torturous pictures and we just having
17:30like uh uh like coffee and then the
17:34Verge leak screenshots of Bing chat and
17:37um I was like okay like uh there's this
17:3930-day due diligence period right and
17:42one of the other investors would give me
17:43a term sheet he just increased it to 45
17:45days you know you could see the diff
17:47yeah right it was done sneakily and I
17:50knew why clearly he also text what do
17:53you think about this thing okay okay I
17:55get it I get it getting a little
17:56sheepish and then the other person I
17:58hand with like he text me the night
18:00saying hey do you have time for a call
18:02tomorrow okay like clearly like this is
18:05it right so I told my co-founder look
18:07maybe they're going to uh back out or
18:10ask us to Pivot so um maybe we should
18:14just try to sell the company and like
18:15get it done you know this is not going
18:17to go anywhere the person actually who
18:19hand shook said look I'm not going to
18:21ask you to Pivot I'm not going to ask
18:22you to like do anything different uh you
18:26guys keep going and uh we already word
18:28is word and I I was like damn that's
18:30that's pretty impressive and then the
18:31next week actually Google also releases
18:34a Blog from Sundar saying they're
18:36announcing something called The Bard
18:37with just screenshots so we knew that
18:39like this is going to get pretty big and
18:41competitive but we were like look it's
18:44at the end um Microsoft was never really
18:48good at consumer products for a long
18:49long time you can't suddenly change that
18:52uh so they actually messed up the
18:53opportunity in my opinion totally Google
18:55obviously I knew that they're going to
18:57have their own problems challenges so I
18:59felt like there was space for someone
19:01else here yeah having spent almost a
19:03decade at Google myself um I see a lot
19:06of the culture of the early days of
19:08Google like the things I've learned
19:09about Larry or about SAR and I see a lot
19:12of that in the way that you have built
19:13your product like there's a lot of
19:15attention to detail feels like you are
19:16the primary user of the product yourself
19:19like is that a thing that you
19:21deliberately tried to do yeah I did I
19:23did deliberately try to do it uh one
19:25thing that Larry said is like you know
19:27we we uh I keep reminding everyone in
19:30our company about it the user is never
19:32wrong so even today while testing a new
19:35feature um it didn't work uh but there
19:39was some ambiguity in the query so the
19:42person I was talking to the engineer and
19:44say Hey you know this is not good what
19:46else could theyi have done here and you
19:50know what they should have done it
19:51should have come and clarified to me
19:53right and and asked me hey I'm not sure
19:55either it's this or this which one did
19:58you actually want
19:59and then I should have clarified and
20:00then it should have gone and done
20:01instead of saying I don't know that is
20:04the user is never wrong principle the
20:06other way of Designing products is like
20:08make the user be a better prompt
20:10engineer mhm blame the user and tell
20:13them to be a better prompt engineer
20:14teach them educate them get them to do
20:15it the way that the product wants to do
20:17it yeah exactly enterprise software is
20:19more like second kind yeah but magical
20:22consumer products are more the first
20:23kind age right like in Google why should
20:26uh Google have handled typos they need
20:28it right we should have all been great
20:30at English it's like Larry says he was
20:32never good at spelling and that's why I
20:34think the true story is YC partner Paul
20:37buite he was just annoyed by it and he's
20:38like someone should build that yeah
20:40exactly and spell check corrector it's
20:42all there similarly Auto suggest why is
20:45it there like easier right similarly uh
20:48cached results I was even reading
20:50somewhere where Larry wanted the
20:51homepage to have the a simulation of the
20:54weather outside your home so that you
20:56don't even need to type the weather
20:58query is this already there so I was
21:00very influenced by that style of design
21:02like including like subtle things like
21:04Chrome search bar if you've already gone
21:06to a site it's already there you just
21:08have to click enter after typing the
21:09first two letters so that influenced me
21:12to like make sure we have the cursor
21:15ready to type on the search bar you
21:17don't need to take your mouse and place
21:19it there it sounds like your your main
21:20metric that you care about is number of
21:22queries per day which is exactly what
21:24Google did I think in the early days
21:25right it's hard to grow that uh
21:29uh without like retention in the long
21:31run you cannot just uh pay for a user
21:34and get that number up user could
21:36install your app and maybe you can even
21:38game it where when they install as one
21:40query automatically submitted but a
21:42repeat query doesn't need to be
21:43submitted yeah I think the only counter
21:45example which I don't think is happening
21:46in your case is the product is not
21:48serving their needs and so they need to
21:51issue a bunch of queries to get what
21:52they want which is kind of the opposite
21:53of like Larry's approach on Google was
21:56you should be on Google as short as
21:57possible cuz trying to get you somewhere
21:59else to solve your so that's not
22:01happening I mean sure I'm sure there are
22:03some errors and stuff but most of the
22:06followup queries actually we see are
22:07like completely irrelevant to the first
22:09query because they just want to keep
22:10continuing the session or questions that
22:13they never even knew they wanted to ask
22:15but they want to keep asking so so I
22:17presume your team has grown a bunch you
22:19raised a bunch of money um how do you
22:22manage the team how do you operate your
22:24team on a week to week or or cycle to
22:26cycle basis with that you know number of
22:29queries per day is our primary metric so
22:32every All Hands we start with that
22:33number okay I don't believe in this um
22:36putting a TV and having the metric you
22:38know being seen every day because I
22:40think that's also distracting but I I do
22:43think like it it makes sense to take a
22:44look every week see the weekly growth
22:46rates um see the monthly growth rates
22:49and like if something declined then
22:52discuss about it figure out ways to
22:54actually freak out if something declines
22:56we do and something grows was like look
22:59into why where so we are very data
23:01driven and we shareed across the company
23:05actually I've been trying to share to
23:07the users too so that they feel like uh
23:10you know it's it's an something that's
23:12actually happening right in front of
23:13their eyes and and they want to be part
23:15of it there's no hierarchy like if if
23:18there is some bug to be fixed if I know
23:20some particular person's working on it I
23:22can go and talk to the person directly
23:24nobody else feels threatened because I'm
23:26going and talking to that person there
23:27is no feeling that because I'm raising a
23:29bug uh uh it's like oh they are going to
23:32be fired or something uh because I
23:34raised like 50 bugs a day so you know
23:37like it's more like they understand okay
23:40this it's important for the product to
23:42feel uh great and if it doesn't feel
23:45good for ourselves then the user is also
23:47not going to feel that in fact we have
23:49way more incentive to go use our own
23:51product but the user doesn't so the
23:53standards for the user should be even
23:54higher so always feel like a user I
23:57think that culture is there company I
23:59love that and and did you intentionally
24:01select for that when you were hiring
24:03like people who were very product
24:04Centric and in the details I wouldn't
24:06say I I explicitly had that as a
24:08Criterion but I look for people who
24:10cared about doing good work if you don't
24:13care and you're just reading it as a job
24:15then it's very hard for you to get
24:17excited about things and I think so much
24:19of it feeds off of the founders and like
24:21your culture your DNA and sounds like
24:23you're that type of person that obsesses
24:25over the details and you're just going
24:27to naturally want to hire people who
24:28share that trait yeah I I do get pissed
24:31off if answers are wrong and I do get
24:33pissed off if people on Twitter are
24:34saying like perplexity is degrading or
24:36like you know but a lot of the things
24:38this some things are actually not true
24:41but I do try to see you know leave aside
24:44the cynicism um even if it was someone
24:47who's like a hater y but if there was
24:50something true there and I want to still
24:52know yeah I I love seeing you engage on
24:55Twitter with customers is that the
24:57primary way that you talk to users or
24:59are there lots of other ways that you
25:01are talking so I mainly use x Twitter
25:04people are just like super like brutally
25:07honest there and uh I think in email
25:10people are a lot more polite yeah which
25:12is okay too like I I like both sides but
25:14I think the brutal honesty brings out
25:16the worst bugs and uh things that people
25:19are afraid to say oh and in person is
25:21the worst where you go show someone
25:23something and they're just going to tell
25:24you good things even if they hate it I
25:26kind of like don't like any hey what hey
25:28tell me what do you think yeah you're
25:30always gonna say nice things right
25:32you're gonna grow your company
25:34presumably you're going to need to hire
25:35more people how do you avoid the fate of
25:38becoming a big slow company well it's uh
25:41beginning to happen already a little bit
25:43right we're not as fast as we used to be
25:46I think some of it is not because of
25:47people it's it's also because things
25:50breaking in production people start
25:53losing trust in the product like today
25:56we deployed some change and then someone
25:57got rated that there was some uh front
26:00and Bug somewhere it was actually
26:01something in the backend but people are
26:03just assuming things I think like uh
26:06fast loading all that stuff matters and
26:09not every new engineer uh has the full
26:12context of the code base in their head
26:13the earlier ones to there's some tension
26:15in like moving fast and breaking things
26:17if you do want to like grow to mass
26:19Market usage uh so that's mainly slowing
26:22us down I would say uh and I haven't we
26:25haven't quite figured out like the best
26:27way to do this fast I mean we do have
26:28staging deployment testing AB test and
26:31all that stuff's happening and that's
26:33naturally slowing us down and like
26:34getting things out to production widely
26:37other than that I would
26:38say uh the obsessive detail oriented
26:42people uh they're only that many people
26:45in the world so obviously you cannot
26:47expect engineer number 250 to be like
26:49that M uh but I I I try my best to still
26:53like you know go flag bugs to whoever is
26:55working on whatever new feature uh and I
26:57kind of like know who's working on what
27:00even at the size I still try I think our
27:03co-founders are amazing they they care
27:05and they push that principle when
27:08they're like building their teams so uh
27:11we we we are trying our best I'm not
27:12saying it's us we figured it out cracked
27:14it uh but at least like we're trying to
27:17fight the entropy here I think that's
27:19the only thing you can try right and
27:21it's a uphill battle but if you keep on
27:23it yeah okay let's talk a little bit
27:25about the future um you know I've seen
27:27you you're most recent launches are kind
27:29of like in different directions more
27:32verticalized or more specific around
27:33shopping or or other things where do you
27:35want to take it like today my view of
27:38perplexity is a more intelligent Google
27:41search that's really useful in certain
27:44scenarios where do you what do you want
27:45me to think it of it in three or four
27:48years if you go and research what's the
27:51best sweater to buy or uh which the best
27:54hotel to stay in this location
27:56perplexity will give you a great answer
27:58but where do you actually go and fulfill
28:00the demand you go to Google and who gets
28:03credit for that monetarily Google Google
28:06we we got nothing maybe we get you a pro
28:08subscription but then someone else will
28:10undercut us and give it away for free
28:12with like cheaper models or whatever
28:14they have bigger cash reserves so the
28:16challenge is like you want to be uh one
28:19place where people can have the endtoend
28:21experience they start with a problem in
28:23their mind and they seek your help you
28:27give them the answers
28:28and you also help them fulfill the
28:30action it's difficult because uh people
28:33think like at the if you have an answer
28:35of like oh like what watch does uh Bezos
28:39wear MH I think he wears some omega or
28:42something um I personally thought it
28:44would be amazing if it not only gave the
28:47answer but it also had a product card
28:50for the specific Omega watch with a buy
28:53button and I just click buy and it's
28:54done but there are other people in the
28:57world who think that's an
28:58it's not even an ad right they think
29:01like that company is paying us to do
29:02this so this is where like some of the
29:04tension with uh the early adopters who
29:07love the adree informational
29:10experience with like what is actually
29:13needed to get Mass market and really be
29:15useful on a daily basis comes from and
29:17there are so many other things like
29:19checking the score of a game or quickly
29:21getting to a website uh if you just
29:23wanted to get a docs link of an API or
29:26if you just wanted to go and book flight
29:28on United the answer could just be a
29:30link the answer could be the weather for
29:32tomorrow that's the temperature or
29:34someone's age should be like you going
29:36type like Elon Musk Networth you'll just
29:38get the answer in like less than a
29:39second on Google right in perplexity
29:41it'll pull the right Source maybe it
29:43might be more accurate than Google but
29:46people don't care about like some of
29:47this Minor Details so what you need to
29:50build is this amazing
29:52orchestration of small models uh typical
29:56knowledge graphs uh
29:58widgets llm streaming answers and more
30:01complicated multi-step reasoning answers
30:04but user doesn't care like user is not
30:07going to tell you like when to use what
30:09you decide that AI nobody talks about
30:12like when to use what that that's s of
30:14router that that orchestrator I think
30:16that's the hardest thing to build and
30:18whoever builds that and and and and can
30:20operate that at a scale of billion users
30:23and also knows how to monetize like some
30:25of those queries really well right is
30:28going to be the next Google cuz they'll
30:29have the search bar everything will they
30:31they know exactly what to do they'll go
30:32and ask clarifying questions they it
30:34truly understands the user and also does
30:37tasks for you and and also lets you like
30:39surf the web in the typical way all
30:41in-one experience you could even argue
30:43maybe nobody will ever be able to build
30:45this because it feels like a daunting
30:47task but I could say whatever Google has
30:50already built is the closest system to
30:52something like this agre so the next
30:55generation of this clearly can be buil
30:57MH you just have to like persever and
30:59work for a decade or two on this problem
31:01if I talk to people at Google they would
31:03say yep that's what we're building in
31:05fact I know they've been saying that for
31:06more than a decade um probably same at
31:09open AI probably the same at anthropic
31:11when you look at the people that you
31:13likely will be competing against in the
31:14next 10 years um what do you think is
31:17the piece that's maybe going to give you
31:19the edge to to win Obsession about the
31:22user and good product taste there's a
31:25lot of these things that require a lot
31:27of domain knowledge out of the list you
31:29mentioned Google is the only company
31:30that actually has the product taste to
31:33do this and and arguably like you know
31:37all the distribution in the world
31:38everything right except the Dilma is
31:40also there it's funnily like you know
31:42it's a search company but it's it's also
31:45an ads company and search is kind of
31:48almost exists in service of the ads comp
31:50yes not in the other way and you could
31:52argue okay that's outside the search
31:56Revenue every quarter which is like
31:58close to 200 billion a year MH there's
32:01still like 100 billion and so other
32:04other places YouTube and Cloud but the
32:08margins are all coming in search right
32:11cloud is only like recently profitable
32:13YouTube is not never going to be a high
32:15margin business because number one they
32:18they don't serve ads on subscription uh
32:21like users and number two like they have
32:23to pay the creators they have to pay the
32:25Media Partners so it's never going to be
32:27as High margins of search so you're
32:29arguing basically the stock price is
32:32going to be their encumbrance correct CU
32:34like ball TR is like like crazy it just
32:38automatically uh you know panics if
32:40search Revenue goes down but search
32:43Revenue has to go around in in a world
32:45where people are just directly talking
32:46to AIS and agents are doing stuff for
32:48them that doesn't mean they're not going
32:50to do anything about it they're still
32:51building Gemini and like the new app the
32:54hypothesis is that like they're not
32:56going to be able to easily put it on the
32:57ore Google where they already have all
32:59the billion users and that's true right
33:01right yeah you're you're arguing that
33:03whoever wins this in the long run will
33:05kind of by definition need to come up
33:07with a new monetization model a new
33:09business model yeah there's like a ton
33:10of other problems to solve like for for
33:13shopping or travel or like all these
33:15things like which which Merchants do you
33:17use or like which hotels do you plug
33:20into or you know who's the middleman
33:22there and who handles the booking and if
33:24a customer wants to cancel stuff Google
33:26actually saw these a lot of these
33:27problems too right they're not just like
33:30oh a page rank or like a map reduce or
33:34um you know all these advances that they
33:35made in like visual like deep learning
33:38and and and like Bird Transformers it's
33:40not just that that is great but they
33:42also did a lot of other boring work of
33:44bringing Google Finance Google shopping
33:47Google flights I feel like perplexity is
33:49better position to do these things than
33:50open air and Tropic because we have it
33:52in our DNA to care about the user and
33:54the product uh we're not just talking
33:56about reasoning and models right we but
33:58we are pretty much familiar with all
34:00those things and we are very much
34:02capable of taking the latest open source
34:04models and serving them ourselves and
34:06fine-tuning them and post training them
34:08and doing evals we're not like AI
34:10illiterate we're not going to like spend
34:12all our band with building data centers
34:14and chips and like trying to just talk
34:16about like breaking the most reason
34:18coding a math benchmarks I think there's
34:19value in that but it's quite autal to
34:22like building the next generation
34:24information experience all right Arvin
34:27thanks so so much for joining us it's
34:28great chatting thank you for having me
34:30again
34:35[Music]

About this transcript

This page contains the full transcript of How To Build The Future: Aravind Srinivas by Y Combinator, generated from the public captions YouTube serves with the video. The transcript has 6,860 words across 928 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 translate the transcript into 100+ languages.

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, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.