How To Build The Future: Aravind Srinivas

00:34:39
https://www.youtube.com/watch?v=SP7Ua8FKZN4

الملخص

TLDRArvin Shavas, co-founder and CEO of Perplexity, shares insights into the inception and growth of the company, emphasizing its goal to build an intelligent search engine that exceeds traditional models like Google. He recounts his early experiences in AI and how they led to recognizing the opportunity in enhancing search functionalities through AI. Perplexity aims to create a user-centric product that allows for an end-to-end experience in information retrieval while adapting to the evolving AI landscape. As the company navigates the competitive space against giants like Google, Shavas stresses the importance of user engagement and innovative monetization strategies.

الوجبات الجاهزة

  • 🚀 Perplexity is a growing AI-based search engine with over a billion-dollar valuation within three years.
  • 👁️ Emphasis on user experience is key, aiming to make search simple and relevant without clutter.
  • 📈 Engagement doubled after introducing follow-up questions, showcasing strong user interaction.
  • 🤖 Shavas emphasizes the importance of evolving AI models to enhance functionality and speed.
  • 💡 Future ambitions include facilitating end-to-end user tasks and shifting from traditional advertising models.

الجدول الزمني

  • 00:00:00 - 00:05:00

    The introduction of follow-up questions has significantly increased user engagement on the site, highlighting the potential of the platform. The visionary ambition behind the project reflects aspirations similar to those who sought to innovate at a scale akin to Google.

  • 00:05:00 - 00:10:00

    Arvin Shavas recounts his journey into the AI world, rooted in deep learning research, culminating in pivotal internships at OpenAI. His formative experiences underscore the importance of foundational research in shaping future entrepreneurial pursuits.

  • 00:10:00 - 00:15:00

    A focus on developing AI-powered search capabilities led to the inception of Perplexity. Initial challenges included skepticism about competing with tech giants like Google, yet a unique approach towards querying and information retrieval led to a path forward, marking the beginning of a distinct venture.

  • 00:15:00 - 00:20:00

    Early product iterations involved leveraging Twitter data to create a conversational search experience, allowing users to engage with their social media insights in innovative ways. This initial MVP laid the groundwork for understanding user needs in information retrieval.

  • 00:20:00 - 00:25:00

    Transitioning from the Twitter-centric demo to Perplexity involved recognizing the potential of unstructured data. The team aimed to capitalize on evolving generative AI models, believing that improving AI could lead to a more fluid user experience without heavy indexing.

  • 00:25:00 - 00:34:39

    The competitive landscape solidified an understanding that while giants like Google posed challenges, the unique appeal of Perplexity lay in its user-centered design and adaptability. The intention is to evolve not just as a search tool but as a comprehensive solution that integrates various user needs, from shopping to booking, all while maintaining high engagement.

اعرض المزيد

الخريطة الذهنية

فيديو أسئلة وأجوبة

  • What is Perplexity?

    Perplexity is an AI-driven search engine designed to deliver intelligent and accurate search results tailored to user needs.

  • How has Perplexity grown since its inception?

    Perplexity has rapidly escalated to over a billion-dollar valuation within three years, focusing on enhancing user engagement through follow-up questions.

  • What differentiates Perplexity from Google?

    Perplexity aims to provide a more focused and user-friendly search experience, emphasizing simplicity and immediate relevance without the clutter of ads.

  • What challenges does Perplexity face in competing with Google?

    Perplexity must address the complexities of integrating monetization with delivering quality search results while competing against a well-established platform.

  • What is the future vision for Perplexity?

    The vision includes evolving into a comprehensive platform that not only answers queries but also facilitates user tasks, shifting away from current ad-centric models.

عرض المزيد من ملخصات الفيديو

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