IITM ACR - LLS - TALK BY SHRI. ARAVIND SRINIVAS -FOUNDING STORY ANDJOURNEY OF PERPLEXITY
8K views|6 months ago
💫 Short Summary
Aravind Sash, CEO of Perplex City, discusses building an AI-native search engine providing precise answers. The journey involves pitching ideas to venture capitalists, transitioning to conversational search, and focusing on user interaction. The team emphasizes factual accuracy, iterative improvements, and scalability. The speaker reflects on AI research, startup challenges, and personal growth. Key takeaways include focusing on strengths, team collaboration, perseverance, and embracing entrepreneurship. The company monetizes through subscriptions for upgraded AI models. The discussion also explores the importance of survival in business, AI advancements, and the interplay between rationality and irrationality in decision-making for startups.
✨ Highlights
📊 Transcript
✦
Overview of Aravind Sash and Perplex City.
00:09Perplex City is an AI company that specializes in creating an AI-native search engine for direct user query responses.
The company has experienced significant growth since its launch in December 2022, serving half a billion queries in 2023 with 10 million monthly users.
Aravind Sash, the founder and CEO, has a strong educational background and valuable experience from companies like Google Brain and DeepMind.
The company's approach focuses on delivering clear answers to user questions by cutting through the noise.
✦
Importance of precise answers from search engines based on high-quality sources.
06:26Need for a tool like Wikipedia with personalized content tailored to user preferences.
Introduction of a conversational answer engine for natural interaction and on-demand knowledge.
Goal of scaling the tool to serve millions, potentially billions of people.
Emphasis on sustainable growth through revenue reinvestment for continuous improvement.
✦
Challenges of pitching ambitious ideas to venture capitalists and importance of risk-adjusted concepts.
09:27Shift from aiming to build a Google competitor to focusing on natural language to SQL for searching databases.
Journey of learning to build products and startups, leading to the development of a productivity tool.
Drawbacks of chatbots without search grounding, leading to the creation of a tool grounded in web data.
✦
The success of a conversational search tool led to the launch of OpenAI's GPT chat in late November.
14:06Users found the tool more useful than Google for quick answers, leading to frequent usage.
The tool eliminated the need for Discord servers and bots in AI products.
Lessons learned included prioritizing privacy in search queries and positioning the tool as an answer engine rather than a chatbot.
The team quickly launched their product, leveraging available resources and focusing on scaling usage.
✦
Development of search product with conversational chat UI.
15:49Emphasis on follow-up questions and detailed prompts for users.
Unique user interface influenced by academic backgrounds, with citations to show information sources.
Commitment to rigorous writing with peer-reviewed articles as citations.
✦
Importance of backing up statements with facts and journalistic rigor.
21:57Process of writing accurate, concise answers using web search results with unbiased tone.
Emphasis on iterative improvements and overcoming doubts when launching new projects.
Addressing scalability challenges, business models, and potential funding issues for startups in AI/ML field.
Encouragement to have conviction in language models and search index improvements to reduce errors over time.
✦
The decreasing cost of hardware is leading to more efficient models and lower costs for serving users.
22:55Startups can monetize their services by saving people valuable time.
Disrupting existing systems through a venture or company can be challenging but lead to success.
Aim for big, bold bets in order to achieve success.
Internalize the power law of nature, where the majority of work is done with a minority of effort, to improve task and team dynamics efficiency.
✦
Importance of Iteration and Working on Challenging Projects
27:45Aim for 'just enough' rather than perfection and continuously iterate on ideas.
Startups should focus on addressing the 'longtail' over time and validate ideas and resources.
Startup valuation is dominated by a few successful companies like Apple, Google, and Microsoft.
Creating outlier events through a few key decisions can lead to significant outcomes.
✦
Importance of Collaboration in Startups and Problem-Solving.
29:50Being contrarian and correct is crucial for making impactful contributions.
User feedback on perplexity has transformed work processes for students and teachers, saving time and energy.
Perplexity is gaining popularity as a search tool over Google due to its user-friendly and efficient results.
Focusing on the right tasks helps in staying motivated and making progress, despite skepticism.
✦
Emphasis on believing in oneself and following one's heart despite others' opinions.
33:45Influence of Steve Jobs and Larry Page in pursuing visions and achieving success.
Highlight on entrepreneurial spirit and autonomy in thinking from first principles.
Google founders, PhD students at Stanford, turned PageRank into a trillion-dollar company through innovation and calculated risks.
Success through hiring talented individuals and taking risks in business ventures.
✦
Reflections on AI Research Journey
36:54The speaker discusses his early work on generative models in AI research before they gained popularity.
Starting a company is challenging, requiring a founder's energy for success.
Despite initial failures in research projects and startup ideas, perseverance is key to eventual success.
The speaker shares personal experiences of failures leading to unexpected opportunities and encourages embracing the journey, not just the outcomes.
✦
Key Highlights for Success in Entrepreneurship
41:31Focus on developing your natural strengths and skills to excel in your endeavors.
Surround yourself with people you trust and respect, while being mindful of your weaknesses and strengths.
Build a strong team by recruiting talented individuals, as they are crucial for achieving success.
Use your profile and experience to secure funding when raising capital, even if the initial amounts are lower than expected.
Start with a minimal viable product and continue to build and improve upon it.
Embrace the entrepreneurial journey, be prepared for rejection, and stay focused on your goals.
✦
Importance of raw energy and continuous improvement in early stage startups.
43:49Unique qualities of startups including speed, risk-seeking attitude, and agility.
Importance of authoritative sources and unbiased approach in search engine ranking algorithms.
Explanation of the logo design and business model of the company.
Monetization through subscriptions for upgraded AI models and features.
✦
Journey into AI and Importance of Python Programming
49:44Challenges faced during a machine learning class were overcome with exposure to Python and related libraries in college.
Proper attribution and driving traffic to relevant sources are emphasized when dealing with copyright issues related to internet data.
Strategies for maintaining product value while managing traffic volume are discussed in the segment.
✦
Importance of Soft Skills in AI Journey.
52:53Participated in ML contest and took a class to enter AI field.
Split responsibilities between tech and business based on skill sets and marketability.
Emphasized the importance of marketing and making correct technical decisions.
Acknowledged CTO's superior programming skills and value of soft skills in business and tech roles.
✦
Importance of long-term survival in the business world for 5-10 years to plan for success.
55:16Mention of AI advancements like Char gbt and concerns about job security.
Advice for college students to use AI tools to increase productivity while retaining human judgment.
Compliments to the speaker for academic achievements.
Caution to consider the impact of messaging on young audiences.
✦
Importance of applying 80% accuracy in software products and considering implications on human life.
59:33Bottleneck for AI agents due to hardware limitations, stressing the need for advancements in algorithms.
Potential of AI-native consumer devices in creating new experiences and future of robotics discussed.
Challenges in achieving human-like dexterity and common sense in robotics highlighted.
Mass market adoption of AI technologies may be limited until breakthroughs like the 'CH GPT of Robotics' occur.
✦
Key highlights of GPT Builder discussion.
01:04:08Importance of accuracy, speed, and distribution in the industry for staying ahead.
Competitive edge of GPT Builder in terms of speed and accuracy.
Plans to differentiate the product further from competitors like Bard and Chach PT.
Discussion on the sustainability and potential impact of GPT Builder on the economy of GPTs.
✦
The importance of balancing rationality and irrationality in decision-making for startups.
01:08:03Rational thinkers are encouraged to embrace their irrational side to take risks and pursue entrepreneurial endeavors.
Rationality may lead to questioning and doubt, but incorporating irrationality can drive action and innovation.
Pain tolerance is crucial in entrepreneurship, and rational individuals with a high pain threshold can leverage their irrational side for risky but rewarding pursuits.
00:00artificial intelligence Grace our stage
00:02today we are privileged to have with us
00:05a distinguished individual who is
00:07shaping the AI industry in India and
00:09Beyond Dr sh Dr irin Shas the founder
00:13and CEO of
00:15perplexity now I will invite Mr kaviraj
00:19n CEO office of institutional
00:21advancement to introduce Our
00:27Guest good evening everyone uh welcome
00:30to this evening's leadership lecture
00:32Series where arind will be talking to us
00:35about his journey so far so for the sake
00:38of formality and all of that let me
00:41arind is not a stranger to any of you
00:43but let me introduce his uh and read out
00:45his profile aravind SAS is a founder and
00:48CEO of perplex City a company that seeks
00:51to build an AI native sech that directly
00:54answers your questions rather than
00:56provide you
00:57links since it's launched in December
01:002022 perplex city has grown to serving
01:03half a billion queries in 2023 and 10
01:06million monthly users and has raised 100
01:09million in uh funding from the likes of
01:11Jeff Bezos Jeff Dean Yan M Leon Nat
01:16fredman and top Venture firms like
01:18bessma Ventures ivp Nea at all aravind
01:22got his btech and mtech from I medras in
01:25computer science and a PhD from UC
01:27Berkeley and has worked as a research in
01:30open AI Google brain and Deep Mind
01:33publishing research papers on supervised
01:36and reinforcement learning so we look
01:39forward to very interesting and exciting
01:41session uh this evening uh and welcome
01:44arind uh to do the honors
01:47[Applause]
01:58yeah
02:01okay can you hear me okay thank you
02:05everybody for coming
02:07here uh and always like you know it's
02:10all great to be back I'm actually back
02:12to India after 5 years so it feels
02:16really good to come back to IIT and give
02:18a talk and I bumped into my professor
02:21here professor ravindran and like he was
02:23like okay so came here why don't you
02:25give a talk and okay sure you know and
02:27then that's how this happened uh so I'll
02:30be talking about my company
02:32perplexity uh just kind of how it came
02:35together and some lessons that you know
02:38I've generally learned over this period
02:40And like I feel could be useful to some
02:42of you here so for those of you who
02:46don't know perplexity uh I feel like
02:48this video would communicate to you what
02:49we are trying to
02:58do
03:07[Music]
03:13[Music]
03:32[Music]
03:39[Music]
03:57so right now like this is the current in
03:59in cumbent right like you ask something
04:02I'm let's say you're just looking for
04:03headphones uh instead of just getting
04:06the you know like something that you
04:08really want you're actually filled with
04:10more clutter you're not sure exactly
04:13which one to click on like you're going
04:15to
04:15scroll uh spend some time like seeing
04:19things and you're still more confused
04:20than when you started off with uh
04:24instead like we built this product where
04:27it just tries to cut through the noise
04:29and gets exactly what you wanted like
04:31clear answer to your question and with
04:34references so for for example a query
04:36like this you don't even know exactly
04:38what is the best but you still want to
04:40know what is the best so ideally you
04:42want an AI where you just say hey best
04:45headphones and then uh it want it should
04:48come back and ask you clarifying
04:50questions like it should ask you okay
04:52what's your budget cuz some you don't
04:54have to specify it beforehand the AI
04:56should work for you and then it should
04:58say what what what's your primary reason
05:01for using it and uh get more input from
05:04you help you expand what you originally
05:06wanted in your head but you didn't
05:08translate it well to an actual search
05:10query and just give you the answer
05:13precisely based on like you know high
05:15quality sources on the internet that's
05:17ultimately what search should do for you
05:20uh and the current incumbent does not
05:22have any incentives to do that because
05:25they make money of you clicking on all
05:27of these links but what you really want
05:30is to just get to the answer right and
05:32that's what we are trying to
05:34build and uh this this particular user
05:38you know on Twitter said this very well
05:41uh how to best think about perplexities
05:43like Wikipedia and chat GPT having a
05:46baby uh but the data is coming from the
05:49whole of internet not just from uh
05:52Wikipedia now Wikipedia is something
05:54like all of us have learned a lot from
05:57right like it's it's it's like a amazing
06:00tool especially you can even say it's a
06:02nerds Paradise so it'll be great if we
06:05can like recreate something like
06:07Wikipedia but not in a static version
06:10for everybody that is it's not like one
06:13version for all of you uh depending on
06:15the level of depth you want or the
06:17breadth you want you can personalize
06:18what you get and that for that you
06:20actually need an AI to do that for you
06:23and today we have the right tool for
06:25that and that's
06:26perplexity so our mission here is to to
06:30build the best conversational answer
06:33engine uh and I call it answer instead
06:35of search engine because right you don't
06:37you don't need to type in keywords
06:39anymore you can talk to it like how you
06:41would talk to your friend you let's say
06:43you're before an exam you're preparing
06:45for something you're not going to your
06:47friend and asking like a topic right
06:49you're actually asking hey about this
06:50particular thing what what what how does
06:52this work you can actually now start
06:54talking to computers like that and get
06:57knowledge at demand uh and we want to
07:00scale this for all people like build
07:03their infrastructure needed to serve
07:04millions of people and over time
07:07hopefully a billion uh today we already
07:09serve millions of people and by doing
07:13this like we want to make sure that we
07:15can grow sustainably so build a
07:18business and generate Revenue through
07:20both the platform and the infrastructure
07:23and take all that money and reinvest
07:24into making the AI better and better
07:27because we want the product to keep
07:29getting better for every individual user
07:32as more people use the product and AI
07:35can do that because as it gets more data
07:37it just keeps getting
07:39smarter now how did we arrive to where
07:42we are today it's very
07:44nonlinear
07:46so the if you want the answer already
07:49it's like we kind of built something
07:51that we wanted first and we found it
07:54very useful and we saw that people
07:56around us started finding it useful and
07:58therefore like we put it out and
08:00everybody started using it and I I
08:02strongly believe in this a lot of people
08:04start companies and they try to think
08:08for an end user end customer that's one
08:11way of building a company it's it's
08:13certainly like you know nothing to
08:14criticize about that uh it's worked out
08:16for many
08:17people but I personally believe that um
08:21unless you care that is you the founder
08:25or the group of Founders really care
08:26about that particular problem it's very
08:28hard to build something truly unique and
08:31new uh
08:33because if you don't put in the energy
08:35and and at in the beginning when you're
08:38nobody or nobody knows about you uh the
08:40only person who cares the most about
08:42your product is yourself so if you don't
08:44really want to build it well then it's
08:46not going to end up well and uh we were
08:49not working on a search engine company
08:51in the beginning at all in fact like if
08:53you go and Pitch to a venture capitalist
08:56that you want to build a Google
08:58competitor they would not even give you
09:00like you know any money they' be happy
09:02to give you a lot of money for an idea
09:04that might never work but feels very
09:07safe like say you're saying I'm going to
09:09build an AI co-pilot for a particular
09:11narrow profession knowledge worker uh it
09:14feels very risk adjusted for them that
09:16they would put in a lot more money but
09:19if you go and say I'm going to build
09:20this AI NATO search engine they'll be
09:21like no that's that's too competitive
09:23like there there's this one Monopoly
09:25here so obviously we we we were not that
09:27ambitious either to start start with uh
09:30so we started working on a very simple
09:33idea of like natural language to SQL SQL
09:35is basically structured query language
09:38that you would write to query over
09:40spreadsheets like you know your your
09:41Google Sheets or like a bunch of tables
09:44and we were working on this as a way to
09:45search over
09:47databases and um instead of building a
09:50tool where it's just like a SQL editor
09:52and like you're trying to write SQL our
09:55passion for search flowed into the
09:57product that we were kind of viewing it
09:59as search and we saw okay how what if
10:02the internet was just organized as
10:03tables and we could just power search
10:06over that uh kind of like imagine if
10:08Google Maps was a table or Twitter was a
10:11table or LinkedIn was a social graph
10:13organized in the form of tables and we
10:15could just power search over that so we
10:17were we were like kind of you know your
10:20your passion draws you to the ideal
10:23convergence point so we we were working
10:26on search and then while building the
10:28startup we had like very little idea of
10:31how to do most of these things uh like
10:33for example this is you know Johnny is
10:35my co-founder and he's asking about
10:37Vector databases KN and search these are
10:39all like machine learning Concepts um or
10:42like s soql is one type of SQL and here
10:45me I'm asking some qu commands about
10:48servers and Ruby and Bash commands and
10:50things like that we this was like you
10:52know way before chat
10:53GPT uh and we just wanted a tool where
10:56we could ask all these questions because
10:57we were completely in inexperienced in
11:00building products or building startups
11:02so we built this tool for ourselves uh
11:04to be just more productive and then we
11:07realize that if you just have a chatbot
11:11uh that has no search grounding and it's
11:14only answering based on what the
11:16language models already learned it's
11:19going to hallucinate it's going to tell
11:20incorrect things and it's uncontrollable
11:23hallucination where you cannot even
11:24verify if it hallucinated or not so we
11:27had this and this is me asking questions
11:29about like health insurance and why did
11:32this come up because uh we got our first
11:35engineer and he came and asked us hey I
11:38need health insurance okay like I I
11:40don't know anything about health
11:41insurance and I don't I can't like
11:44bother my investors to ask about all
11:45this so and Google doesn't tell me
11:47anything about these kind of questions
11:49because uh insurance is a very highly
11:51paid ad adword category that they have
11:54no incentive to actually give you
11:55answers to any of these so we built this
11:58tool we had this tool and was very
11:59useful so then we said okay why don't we
12:02ground it with the web and prevent
12:05hallucination at least like not
12:07completely prevent but reduce it
12:09dramatically and then we realized like
12:11it could answer questions about what's
12:13going on in the world like for example
12:15that was the time Elon Musk was
12:16completing his Twitter purchase so we
12:18asked questions about like has Elon
12:20completed his Twitter purchase and then
12:24we made it
12:25conversational so one person could start
12:27asking a question another person could
12:30become curious about that and then uh
12:33you know ask a follow-up question so
12:35then search for the first time became
12:37like a conversational experience and a
12:39group learning experience so it was we
12:40definitely knew we were on to something
12:42new it was obviously a side project
12:46right a tool that was meant to be useful
12:48for us to build something else but that
12:51tool itself started becoming the main
12:53part of the story so we were pretty
12:56motivated by our own like you know usage
12:59of the tool because you could track how
13:01many queries were being done per week
13:03and I could see it increasing so we have
13:06this idea okay let's put out a Discord
13:07server and invite a few friends you know
13:11and see if they take the tool seriously
13:13uh when you're initially doing a startup
13:16and you ask friends to try out your
13:18product that's in beta most of the
13:20people don't try they're like hey very
13:22cool and then they don't even open the
13:23link you sent so then you know some some
13:26fraction of the people do come and try
13:28and and they try one one or two queries
13:30and they just leave off and like you
13:31know and then the the final fraction is
13:34going to like actually be using it and
13:36you can look at how they use it so what
13:38we realized is like people start whoever
13:41was excited about it were were using it
13:44pretty frequently during the week and
13:47some of them were like this is way
13:49better than Google for me because I I
13:51can just get the answer okay like there
13:53was a point when meta stock was dropping
13:56and like someone could clearly
13:57understand like what was going on on
13:59much better than like read you know
14:00reading five or six news pages and uh
14:04this was like you know beginning of
14:06November last
14:082022 and then obviously the biggest
14:11change in AI in recent history uh open
14:14AI launches chat GPT end of November and
14:18U dispels the myth that you need to be a
14:21Discord server and a Discord bought for
14:23AI products because before chat GPT the
14:26wisdom was mid Journey's the most
14:28successful
14:29AI product and it was a Discord bot so
14:31we have we had to have Discord Bots and
14:34that was the wrong idea like nobody
14:36wanted their search queries to be seen
14:38by others so nobody wants their prompts
14:39to be seen in a public Discord server so
14:42the these are all kind of obvious
14:43lessons in hindsight but at that time it
14:45was not clear so when chbd came out and
14:48was like huge we got the courage to
14:50launch something right after and we also
14:53knew we were pretty differentiated
14:55because we always use the web to give
14:57the answer and we always gave the
14:59sources and we were more like an answer
15:01engine like a search experience compared
15:02to like a chatbot so we launched exactly
15:05a week after uh and at that time it was
15:09literally just daisy chaining GPD 3.5
15:13and Bing together because that's what
15:15you should do as a scrappy startup we
15:16had like you know one or two million
15:18funding uh so your your job is not to
15:21like really worry about whether you own
15:23the the building blocks but just put the
15:26product out and see if it even has
15:29takers and scale the usage and then
15:32raise enough funding to go build the
15:34building blocks yourself so we put it
15:36out and to our surprise a lot of people
15:40liked it uh they were pretty excited
15:42about it and like the search traffic was
15:44sustaining it was not just like one or
15:46two days of traffic during the winter
15:49holidays in December uh we could see
15:51sustained traffic and then we got the
15:54courage to make it conversational we
15:56didn't copy chat GPT and said okay we'll
15:59just use the chat UI and have like you
16:02know a chat bar instead of a search
16:04interface we said okay what is the whole
16:06point of having a chat in the search use
16:09case you would only ask a follow-up
16:11question that that's what that's what is
16:13a chat here and you would only want to
16:15ask a followup if either the first
16:18answer was not your expectations and you
16:20want to iterate or you want to ask
16:23something related to the first question
16:26that you initially did not even want to
16:27ask but you just suddenly got curious
16:30and then for the second thing we want to
16:32prompt you to ask like the right
16:33questions like just like how you know
16:35you see suggested questions in Google
16:37but we want to keep these suggested
16:38questions in the context of the answer
16:40and the past query and that was like one
16:42of the biggest uh reasons why people
16:44started engaging with their product more
16:47in it's are all like very non-trivial
16:49decisions why would you do this but
16:51people started clicking on these and
16:53that gave us the idea that most people
16:55did not even want to ask a question they
16:58they just just wanted to be told what
16:59question to ask so that's why we buil
17:01co-pilot you know the whole point was
17:04remove the burden of prompt engineering
17:06you don't expect people to be amazing
17:09prompt Engineers help them start with
17:12something ambiguous and you tell them
17:14what to like you know pick over time
17:16like to make their prompts more detailed
17:18and more precise so that they get
17:20exactly what they want and this is
17:22Guided by the product principle that the
17:24user is never wrong so don't blame the
17:27user for writing a bad product prompt
17:28blame yourself that you didn't design
17:30the product in a way where they were
17:32guided towards the right prompt so
17:34that's why we will
17:35co-pilot and uh you know for ex and all
17:38these
17:39like why why does perplexity have this
17:42unique user interface
17:44for uh putting sources at the top and
17:47giving you sources it actually really
17:50comes from um my academic background my
17:53co-founders is academic
17:54background uh citations are a very big
17:57thing in academ it's sort of like a
17:59currency here uh everyone feels proud if
18:02the P paper gets cited by other people
18:05and and especially if it get cited by uh
18:08people they
18:09respect uh then they feel very good
18:12about it and so in in in our uh user
18:16interface even though we give the answer
18:18we do show the user exactly like where
18:20the answer is coming from at a granular
18:23level every sentence has a corresponding
18:25citation and U you know in when you're
18:28writing a research paper uh the f one of
18:32the most important things you're you
18:33always keep in mind is you're not
18:35supposed to write any sentence in the
18:37paper for which you cannot like have a
18:40corresponding peer reviewed article as a
18:43citation uh either the experimental
18:46result or someone else's paper should be
18:48able to back up what you write in any
18:50any any part of the paper that's what
18:53makes the writing very rigorous and you
18:55learn this exercise over like writing
18:57many papers in in your PhD or undergrad
19:00research and so this was something that
19:03stuck with me very very well and we
19:06thought okay what if we could bring this
19:08foundational
19:09principle when you're writing research
19:11papers as a core principle baked into an
19:13AI chatbot so tldr don't say anything
19:17you cannot back up with facts another
19:19profession that uses this very
19:21rigorously is journalism right like they
19:23cannot write something that uh unless
19:25it's an opinion piece if it's an actual
19:28story they cannot write something that
19:31they cannot site they need to have their
19:33sources for everything so then this was
19:36actually our first prompt uh write an
19:39accurate concise answer using only the
19:43web search results you must only use the
19:45information from the provided search
19:47results use unbiased and journalistic
19:49tone today's day blah blah blah combine
19:51the search results into coherent answer
19:54and then site the search result using
19:56this bracket dollar Index this is
19:58exactly how you write like academic
20:00papers like if those of you are in AI or
20:02machine learning this a Europe citation
20:04format and uh and then you just use this
20:08as a prompt and tell the model to do
20:10this and it just does it that's thanks
20:13to great AI progress and we were able to
20:15take advantage of that and using this we
20:19launched the first version of perplexity
20:21now it feels so slow like when I look at
20:23this this video is actually sped up when
20:26we first launched uh a friendly investor
20:29told me you should not have it as a sub
20:32you know you should have this not as a
20:34query button you should have it as a
20:36submit button because it's almost like
20:37feels like you're launching a job on a
20:39cluster and the result takes forever to
20:42come it was that slow when we launched
20:44it and now like we are widely
20:47appreciated and like people wonder why
20:49are we so fast when other products like
20:51chbt bar are much slower so that's what
20:54you should really consider you know you
20:57don't have to be per perfect when you
20:58first put it out uh with with several
21:01iterations you can eventually get there
21:04and there is always a LW
21:07tendency and Bill Gates has SP spoken
21:10about this to overestimate the short
21:12term underestimate the long term like
21:15when you first launch something like
21:16this where there's like lot of bugs and
21:18errors like many queries go wrong uh
21:21there's a tendency to think like this is
21:23not going to work like these are all the
21:24things people told us when we started
21:28this a you know this whole answer as the
21:30primary thing is not going to work you
21:32should have the 10 Blue Links and you
21:33should put the answer as a summary at
21:35the top you know it's a step in the
21:37opposite direction it doesn't truly
21:40understand
21:40anything it just like you know
21:42regurgitates what's already there except
21:44it makes
21:45mistakes uh there's it's more expensive
21:48per query how are we ever going to scale
21:50this to billions of queries per day
21:52what's the business model if you cannot
21:54put ads per link why why would a startup
21:57work on this it's meant for like
21:58Microsoft or or even know if Google
22:01decides to do this this Lobby isly get
22:03killed nobody's going to fund this it's
22:05you know search instead of micros
22:07seconds it's macroc seconds so these are
22:09all the things people said and except
22:12you still need to have conviction okay
22:14uh if these language models keep
22:17improving and if your search index also
22:19keeps
22:20improving uh most of the mistakes are
22:23going to decline
22:25exponentially and the cost of Hardware
22:28is going to come down these models are
22:30going to get more efficient so the cost
22:32of serving all these will go down too
22:34and if it gives people back their time
22:38if they don't have to do as much
22:39searching anymore people's time is
22:41fundamentally valuable so you can
22:43monetize that because you're giving them
22:45back and so obviously why would a
22:47startup work on this because incumbents
22:49are going to find it more difficult to
22:51work on this cuz they have to disrupt
22:53what they already have to be able to do
22:55this so we knew all of these from a from
22:58a first principes perspective we knew
22:59that we were doing something right even
23:01though the rest of the world did not get
23:02it so and that's usually a good position
23:04to be in so nevertheless it has to be
23:08said if you want to change something
23:10that's so prominent and
23:13default it it's definitely going to be
23:16painful it's going to be hard and you
23:19have to be irrational to start a venture
23:21back company to do this even now it
23:23might be irrational right like we have
23:25not completely made it yet but
23:28uh this is just going you know trying to
23:30say make it very clear
23:32that it's not meant for the rational
23:35thinker to do like a startup it's it's
23:39most likely going to fail right and and
23:41and um a common pattern every successful
23:44entrepreneur is that you kind of
23:46underestimate what what it takes to do
23:48something you think you can do it
23:50because like you thought it was easy but
23:52not not because it's actually easy it's
23:55it's and then you realize it's so hard
23:57but you've already gone so far along
23:59that you don't want to give up anymore
24:01so you just continue to do it and um the
24:05other thing is like why would you decide
24:07to do the hard thing then if it's going
24:08to be painful because
24:1199% most startups fail and so you at
24:15least want okay the 1% chance that I'm
24:17going to be successful I wanted to be
24:21insanely great like I want the outcome
24:23to be really massive uh so that you know
24:27the expected value is high right so it's
24:30like 01 * something really high plus 99
24:34* something really negative so you want
24:37at least the expected value to be like
24:39pretty positive so that's why like you
24:41want to aim for big bold bets and deeply
24:45internalize the power law of nature
24:47right take B bets so what is the power
24:49law of nature basically this is the
24:51graph right uh it's kind of let me give
24:55you some examples because let's say if
24:58you're doing a task uh 80% of the effort
25:02is spent 80% of the task is done with
25:04the first 20% of the effort like say
25:06you're trying to write an email if you
25:08write the first paragraph you already
25:10know what is the you know who's the
25:11recipient right you know write the rough
25:13outline most of the work is already done
25:16but the rest of the work is actually
25:17what takes you more time the edits and
25:19like trying to make sure your tone is
25:20right and like you want to get the
25:22message across but be diplomatic all
25:25these kind of things that takes more
25:26effort right so this is the power law of
25:29nature where like majority of the work
25:31is done with a minority of the time or
25:33if you have a team of 10 people 80% of
25:36the work is done by two people and the
25:38rest of the eight people do the rest of
25:40the 20% of the work it's very hard to
25:42like have a very uniform distributed
25:44allocation here this is just how the
25:46world somehow ends up working all the
25:48time and you can exploit it you don't
25:51have to complain it you can actually
25:52exploit the power love really well if
25:54you make peace with it and internalize
25:56it a lot and uh same thing applies for
26:00shipping a product uh you don't need to
26:02have all the possible features ready to
26:05ship it for example we ship perplexity
26:08without having most of the building
26:09blocks ourselves so you just need to
26:11have The Sweet Spot the Readiness and
26:15then it it's ready to launch and you're
26:16ready to go and after that you can fix
26:18the long tail same thing uh the product
26:20is going to just has to be accurate on
26:2280% of the queries it can be wrong on
26:2420% of the queries to start with but
26:26that's the longtail that you want to
26:28address over time that's why the company
26:29exists and validation of ideas and the
26:33resources you need to have to like see
26:36if an idea works or not or startups and
26:40valuation if you take all the startups
26:41in the world and the cumulative
26:43valuation of all the startups majority
26:46of this valuation is dominated by the
26:48few startups that exist uh majority of
26:51the market if you if you take the public
26:53markets and add the market cap quic most
26:56of the value is going to be generated by
26:58the first 10 or 20 companies like Apple
27:00or Google or
27:01Microsoft and uh this is kind of called
27:04outliers like you know where like
27:07majority of your returns are coming from
27:09few good decisions you made in your life
27:11and that applies to startups too that
27:13applies to almost everything you do like
27:16uh in your life like there are like few
27:18things that made a tremendous amount of
27:20difference and those are the outlier
27:22events and your job is sort of like try
27:25to make more such events happen right
27:28which is kind of fighting the per per
27:31optimality here but you're trying to
27:33like somehow uh understand and like keep
27:36identifying these 802s everywhere so
27:39this is like one of the most important
27:40lessons to learn in a startup like never
27:43try to be aim for
27:45Perfection uh don't try to be like Steve
27:48Jobs uh always trying to like in fact
27:51that's a negative like you know uh image
27:54that has been painted of him in fact the
27:56first iPhone was not perfect it did not
27:58have the App Store it was not like per
28:00perfectly priced it was launched to the
28:02market right so there is a tendency that
28:05you have to be so good to be able to do
28:07something that's I just want to say that
28:09that's that's not true you just you just
28:10need to be just enough but keep
28:13iterating and why work on something
28:16really hard uh it's a lot easier to
28:20attract really talented people to come
28:23work with you if you're working on
28:25something hard and challenging
28:28otherwise like they don't feel like it's
28:29worth their time you know they they
28:32would be better off chasing these
28:34moonshot projects that they get access
28:36to at a larger
28:37company uh it's so this is really like
28:40something that is uh counterintuitive
28:44right you would think someone is willing
28:46to take a risk bet on a startup is
28:48already pretty risky so you want to make
28:50sure you're like minimizing their Risk
28:53by working on an idea that feels very
28:55easy and achievable but the reality is
28:59you're not going to get the best people
29:00to work with you then CU they don't
29:02think it's worth their time and
29:04therefore you have to work on hard and
29:05challenging things the similar like you
29:07know for example professors they attract
29:09like the best students to work with them
29:12if they can there's this word called
29:14nerd sniping like you know you're you
29:16give them such a hard problem that they
29:18just can't stop thinking about it and
29:20then you those students all want to work
29:22with the that Professor so that's that
29:25sort of applies to startups too
29:27and what happens is that even though the
29:30original problem is so hard when
29:33talented people are all working together
29:35on hard
29:36problems you often make a lot of
29:38progress maybe the first problem you
29:40wanted to solve is never you know solved
29:44but you somehow end up moving the needle
29:46you solve a lot of intermediate problems
29:49and those intermediate problems solving
29:50them ends up being valuable to the world
29:53and you end up moving the needle and and
29:55creating a lot of economic value value
29:57that
29:58way and U look for so how do you decide
30:02what those things to do are look for a
30:04few things that you know to be true but
30:06the rest of the world doesn't agree with
30:08you right being contrarian and also
30:10being right you don't want to just be
30:12cont you don't want to be this person in
30:13the room that is always disagreeing with
30:16what the rest of the people say just
30:17because it's cool to be that way cuz
30:20then people don't like you they just
30:21like they don't want to hang out with
30:23you so you want to be contrarian but
30:25also you want to be right cuz that is
30:27what signal over noise so we you know
30:31when we were like working on perplexity
30:33we always asked our users it's very
30:35important to talk to them and these are
30:37the kind of things they would tell us it
30:39truly has been like revolutionary in how
30:41I do my work like it has saved me so
30:43much time and energy and I have
30:46personally gotten so much use out of it
30:48both as a student and a teacher because
30:50like the nature of what I'm doing it
30:51like I'm taking my own classes but I'm
30:53also teaching classes and it's so
30:55helpful so I wish that there was like
30:57like you could have like literacy
30:59classes like how to use it in good ways
31:01how to use it in like helpful productive
31:05ways where you're still learning you're
31:06still getting like a good experience but
31:08you're not like having it do all the
31:10work for
31:12you I don't know how long I've been
31:14using it for maybe a couple
31:16months uh probably longer than that I'd
31:19say I wish I could go back and like see
31:21the first thread that I had but
31:25uh I feel like I'm just now reaching the
31:27point where mentally I'm like okay
31:30perplexity like instead of going for
31:31Google you know uh but the biggest thing
31:34I think that it had like kind of has
31:37replaced is like Google's uh and like a
31:41lot of other search engines like they
31:44their uh results are just like so I
31:46don't know like SEO oriented I guess
31:49like and they're like for a lot of
31:52topics they're just like not very
31:53friendly at all
31:55sorry sorry I they said the volume is
31:57not uh enough but I so some reason it's
32:00not okay okay yeah I'll I'll try to just
32:03play one more I would say that flexity
32:07is
32:08definitely getting to the point of like
32:12replacing other like you know like I I
32:15don't go to Google to to like look for
32:18something very often
32:21anymore like I literally say it's what
32:24Google and TP should have been or should
32:26be I just tell them it's like completely
32:28I tell people completely replace Google
32:31and so when people are like saying all
32:33this uh so even though like there's a
32:36lot of
32:37skepticism uh you know you you know that
32:39you're actually working on the right
32:40things and that keeps you going
32:43so when you're you know when you're
32:46doubtful like always look for those
32:49little signs of evidence that what
32:51you're doing is right and if you clearly
32:54know that to be true then just don't
32:56worry about what they rest of the people
32:57are saying right there is value to
32:59listening to the rest of the world too
33:01you know I don't want to give give you
33:03this impression just ignore whatever
33:04everybody else says just keep doing
33:07whatever your heart says that that
33:09that's not
33:10what is right either but if you do have
33:13evidence that what you're doing is right
33:16then definitely ignore what other people
33:18are
33:21saying and the last thing I want to say
33:24is like the Steve Jobs thing right like
33:27not word batam his code but dots connect
33:29in hindsight and like Believe In the
33:31Journey you know this famous speech at
33:34Stanford that that that's really true
33:36and I kind of believe in that a lot like
33:40you know I I this is one of my favorite
33:43movies that had a lot of influence on me
33:45it's not a very popular movie so uh you
33:48can buy it on YouTube and watch it it's
33:50uh it's called Pirates of Silicon Valley
33:52it has the story of Bill Gates and Steve
33:54Jobs and how they built like Microsoft
33:57and apple but shows a parallel story at
33:59the same time because they competed with
34:01each other and I I I saw this movie when
34:04I was in campus here uh someone in the
34:07campus told me about it and uh and I
34:11read these two books uh these I mean
34:13people joke about these two books as
34:15propaganda books for Google it doesn't
34:17matter like it has a very positive
34:19effect on a lot of people uh how Google
34:22works and in the Plex and these books
34:24had a lot of impact on me
34:27even you know not it's not because of
34:29reading these books that we working on a
34:31search engine uh it's just that when
34:34you're are working when you're a
34:35researcher you often think like it's not
34:37your job to start a company uh companies
34:40are meant for these MBA type of people
34:42or like people are like constantly
34:44building products and like not doing
34:46research so Google is one of the
34:48companies that truly changed that um you
34:51know dynamic where the founders of
34:54Google are PhD students at Stanford and
34:56they they had an idea called page Rank
34:59and they converted it into like a
35:01trillion dollar company obviously many
35:03more things on top of that but that was
35:05the seed of the company and you know
35:08when I read the forward by Larry Page
35:10the Google founder and CEO at the time
35:14uh you can read the first paragraph he
35:17just said if he were to think about his
35:20future either he wanted to be a
35:22professor or start a
35:25company you know either option would
35:27give him a lot of autonomy the freedom
35:29to think from first principles real
35:31world physics than having to accept
35:32prevailing wisdom and and uh it's really
35:36true like both both these professions
35:38are truly you know entrepreneurial in
35:41spirit different ways of going about it
35:44uh but both of them you know you kind of
35:46have to take bets you kind of have to
35:48hire really talented people in a
35:50professor's case you hire really
35:51talented students in in in a company you
35:54want to hire really talented people uh
35:56employees and uh you know professors
36:00have to raise funding companies have to
36:02raise funding like and you you you want
36:03to be able to make great things happen
36:06right and that Spirit definitely is
36:10impacted me a lot when I was in PhD I
36:12really thought okay you you either had
36:14to start a company or be a professor and
36:15like I kind of didn't want to be a
36:17professor so I thought company was the
36:19only way forward and so I I was
36:23definitely you know when I was in deep
36:24mind doing my internship I used to ask
36:27people their you know fun lunch or
36:28dinner conversations were always like
36:30what is the page rank of 2019 and we
36:32were always talking about gpts
36:34Transformers there's always you know
36:36skepticism around whether it was really
36:38going to take this form that it exists
36:41today CU that was still the age where it
36:43was just gpt2 most people did not
36:45believe in all of this so but definitely
36:48there was you know clear that
36:50Transformers was a special thing and so
36:52I spent a lot of time working on that
36:54and uh so this is kind of how like my my
36:56trajectory was when I was here in itm I
36:59did a lot of AI research and deepl
37:02Professor ravindr here and this was all
37:05like you know not because it was cool at
37:08the time actually uh you know this is at
37:12that time probably he was the only
37:14Professor here in maybe in whole India
37:17working on AI or machine learning like
37:19or deep learning nobody else was even
37:21available so it just happened that it
37:24was very interesting and there were like
37:25a bunch of people very excited about it
37:27at the same time and so we worked on
37:29these ideas like much before most people
37:32now it's like oh it's a hot thing and
37:33everybody wants to work on these topics
37:35but that time was pretty unpopular and
37:39same thing when I went to
37:40Berkeley uh RL was became popular
37:43because of alphao and then U Alpha zero
37:47and all that deep mine so I was like
37:50what is actually like new because in PhD
37:52you're supposed to identify some things
37:54that other people are not really deeply
37:56looking into and identified this area
37:59that okay RL is going to be sample
38:01inefficient it's not going to work off
38:02its own you need to be able to take
38:05ideas from somewhere else and make it
38:07more efficient so I looked at generative
38:08models at the time now it's called
38:10generative AI at that time it was not
38:12called generative I was much less
38:14popular and there were plenty of startup
38:17ideas at that time like working on
38:20compression I mean I'll admit it it was
38:22very much like based on watching this TV
38:24show Silicon Valley oh like you can you
38:27know use generative models for
38:28compression generative models for search
38:30all these ideas were there but none of
38:33them materialized into a company that's
38:35again something I want to highlight is
38:37you can have a lot of startup ideas but
38:40it takes a ton of energy to like
38:42actually start a company and get things
38:44rolling like that energy in the
38:46beginning that is the founder energy and
38:49I definitely did not have it during
38:51PhD uh then I worked started the startup
38:54in 2022 with the simple idea that you
38:58just productize llms there was no like
39:01concrete plan stumbled upon this idea so
39:04everything was kind of like
39:05unfashionable but somehow like the dots
39:07connect like you know my reading of
39:09these books or like watching all these
39:11movies or like all these ideas that I
39:13had before some of it all comes together
39:15so uh if there's anything you take away
39:18is like you know don't always look for
39:20success in every outcome you're doing in
39:22your life right there are many many many
39:24research projects that fail for me here
39:26in fact the first project I did here in
39:28I Madras and research fail first project
39:31at Berkeley fail it's a lot of like
39:34research projects fail and the first
39:37startup idea we started with was
39:39searching over databases that fail like
39:42but somehow we ended up stumbling on a
39:44much better idea so what matters more is
39:47like you willing to just keep going
39:48through the
39:49Journey and um not really just trying to
39:53knock it out of the park at every single
39:55thing you do
39:57and the SQL thing also ended up helping
40:00us by the way you know we put it out
40:02just in case people were excited about
40:04it by saying okay we scrape all of
40:06Twitter and like power search over that
40:09and that's what made like the world
40:11notice
40:13perplexity uh because Jack dorsy the the
40:16actual founder and CEO of Twitter at the
40:18time before Elon took over um tweeted up
40:21tweeted about it and that got us like
40:24100,000 people on our site at once our
40:26site crashed whole world came to know
40:29about us and then we build the
40:31infrastructure to scale to so many users
40:33at at once right so some somehow things
40:37will work out in hindsight just just
40:38believe in the Steve Jobs thing and
40:42um try to aim for doing things you're
40:44like likely to be much better at than
40:46other people uh this is something I
40:48realize very
40:50late uh and good way and obviously you
40:52can ask like how do I know what they are
40:55uh and one heris is like whatever comes
40:57much easier to you but not so much to
41:00people around you it's it's very it's
41:02it's uh it's not going to be too many
41:04things obviously if it's too many things
41:05you're super human so it's going to be a
41:08few things and you pick that and like
41:10just strengthen your strengths even more
41:13it could be that you're much better at
41:14programming you're much better at
41:16presenting things you're much better at
41:18storytelling anything like you know
41:20whatever things you you think like
41:21you're so much more natural that than
41:24others you're likely to you want you
41:26probably want to be strengthening that
41:27even more and as for like companies like
41:31you know work with folks you can respect
41:33and trust it's very
41:36important um and you obviously know your
41:39weaknesses and strengths don't try to
41:40like impose your opinions on things that
41:43you're not an expert at uh even though
41:45like you might be the founder
41:48and invest a ton of your energy like
41:51let's say a week a great founder told me
41:54this recently that he used to spend 30%
41:57of the time 30% of his work hours in a
42:00week just recruiting people because it
42:03mattered so much and so that's something
42:06that you got to really focus on
42:07recruiting great
42:09people and I guess with that I'll I'll
42:13end it my I'll end my presentation thank
42:21you Q yeah so we we can do some
42:24questions if you guys have questions
42:40yeah I I'll repeat it so the question
42:42was what was the experience raising
42:45capital and like how many times have you
42:47been said no uh to be very honest I was
42:50not said no ever but that was also
42:55because of my uh profile at the time
42:58that I had PhD and like worked at open
43:00aai Google and all these companies so
43:04the investment was not a no but the
43:06amount definitely was lower in the
43:08beginning right like I I only was able
43:10to raise one or two million to start
43:12with uh not like tens of millions like
43:14how most other people started many
43:16companies with so that was definitely
43:19it's all relative right like okay you
43:20might say one or two million is great
43:22compared to no money but at at that time
43:24I thought what we raised was pretty
43:29minimal yeah yeah
43:38yeah so as you said like while building
43:41initially you should just do bare
43:43minimum to put the product out right so
43:46like do weren't you being asked
43:48questions like what is the technological
43:49mode that you have why can't others come
43:51and build the same thing so how do you
43:53answer that yeah so most MO is something
43:56that everybody keeps asking but if Mo is
43:59something that you can build in like 2
44:01months then that shouldn't be a mo right
44:03if then somebody else can also do it in
44:05two months modes take like years to
44:07build or even decade uh so I usually
44:11just don't worry about that I I think
44:12it's you know not very important
44:16uh it it's also like the only Moe you
44:19have is an early stage startup is the
44:22raw energy you have to just keep going
44:24week after week after week can keep
44:26making improvements and and the speed
44:28and the velocity like sorry the speed
44:31and like
44:33the risk-seeking attitude you have that
44:36is your mode because that is what other
44:37people cannot replicate like the bigger
44:40companies don't have that they have to
44:42move slowly they have to have a room of
44:4310 people and try to achieve Consensus
44:46These are the kind of things you can
44:48skip so that's your mode and hopefully
44:51you can use all that to get to a point
44:53where you do have a technological mode
44:55in like maybe maybe two two to 3 4 years
44:58time frame but you shouldn't be aiming
45:00for that right away makes sense thank
45:02you um hi uh thank you for the talk it
45:05was great to hear you talk about
45:07perplexity uh my first question was on
45:10uh how does perplexity take care of like
45:13net neutrality so like if I search for
45:16uh five places to go and there's five
45:18lists by my friend and by me uh what
45:21does perplexity choose and why does it
45:23prefer one or the other like what does
45:24perplexity page rank
45:27yeah so that's a that's a good question
45:28so the way we think about this is there
45:31are obviously like some uh whatever the
45:34llm thinks is worth citing we want to
45:38respect that uh so we want to respect
45:41the Judgment of the LM and we also say
45:45okay that's part of the prompt like make
45:46sure you use authoritative sources
45:49relevant content and then you can ask
45:51like what if somebody SEO Bombs all this
45:54that's where we can use the authority of
45:55the do domain itself and filter a lot of
45:57things this is never going to be a
45:59perfect solution right uh because it's
46:01going to be a millions of queries and we
46:03could always be wrong on few things and
46:04great on others but the idea is that
46:07over at a at a aggregate level we can be
46:11pretty unbiased because we don't have an
46:14incentive to rank anything above or
46:17below since we're not getting paid by
46:19any of these domains for doing
46:22that that answers it thank you hi um
46:25what is the meaning of this logo first
46:27one oh the logo second thing is um what
46:30is your business model considering the
46:32fact that you're not going to get paid
46:33by advertisers okay okay great so the
46:35logo is uh the the center curse like
46:39cursor sort of like the you know the
46:42cursor when you're typing in a query and
46:44the
46:45Rolodex is sort of inspired by The
46:47Rolling index we feel like as as as
46:50people keep quering the product more our
46:53index keeps getting better so it's it's
46:55a rolling index and it's a center cursor
46:58so that's sort of uh how the logo came
47:01together and it's very distinct you know
47:04from other other
47:06companies as for what is the business
47:08model today we are monetizing through
47:11subscriptions so you have a Pro Plan and
47:13like you can upgrade to the Pro Plan if
47:15you want like a better quality AI model
47:18uh more features like uploading images
47:21uploading files unlimited uses of the
47:23co-pilot which is the interactive Storch
47:25experience
47:26and also like um more more uh support
47:30from us in terms of customer support and
47:32things like
47:33that uh and over time we plan to have a
47:36lot more value add through the Pro
47:41Plan so I wanted to ask like okay we saw
47:45that you did a lot of uh your Masters in
47:47your pH PhD I think right yeah PhD and
47:50your I think your thing in in AI how did
47:53you get to saying that okay fine and I
47:56want to research in AI like how what was
47:58the journey what was the road map there
48:00what led you to that because I'm
48:02assuming like at that time it wasn't
48:04that popular right it wasn't as hot as
48:06it as it is right now so how did you get
48:08to that point I mean I had a good
48:10professor he was pretty exciting and
48:15uh
48:19so so yeah we we took this machine
48:22learning class it used to be very hard
48:24to get into the class and I
48:27think Professor ravindran you
48:31know Professor ra I
48:37mean yeah so everybody tried to get into
48:39the class and he his exams were super
48:42hard so obviously when when when you
48:45find a challenge you want to take it
48:47right so it was just for the challenge I
48:49mean like why would you why do you write
48:51J it's not you don't actually know why
48:52what what you even want to do in your
48:54life right it's just very challenging
48:57Ching so you just do it for the
48:59challenge and similar to that and
49:04yeah yeah uh so my question is how will
49:07you deal with copyright issues like for
49:10example you definitely will take data
49:12from the internet from some random
49:14website and the website the traffic will
49:18be reduced from that website if if a
49:20person who need to visit from Google
49:22search so yeah so the question is how do
49:25you copyright and drive traffic to the
49:27content
49:28Publishers uh the way we are thinking
49:30about it is uh we are at least
49:33attributing every part of the answer
49:36where are we getting it from uh in terms
49:38of inline citations as well as the
49:40source panel at the top we are
49:42definitely not going to have as much
49:44traffic to the outgoing links as Google
49:48but the traffic is a lot more directed
49:50like we only drive traffic to the source
49:52that really matters to the to the actual
49:54answer whatever the user really wants to
49:56click on so that way the value of the
49:59traffic becomes higher but the volume of
50:01the traffic becomes lower so in terms of
50:03the product value times volume we hope
50:06that it at least Remains the
50:10Same yeah so uh first of all great
50:13presentation um so my first question is
50:16how did your time in the department of
50:17electrical engineering affect your
50:19current pursuit in AI like do you feel
50:21it helped you yeah I think it helped me
50:24because
50:26uh one thing I do remember
50:29distinctly
50:30uh is there was a professor who head of
50:34Department the time Professor Hari he
50:37thought as a Python Programming class
50:39this was my second year and like most of
50:42my computer science friends were just
50:43writing code in C++ and because he
50:46thought as a python in fact it was not
50:48just python it was numi and Mt plot lib
50:50and scipi all of those things I picked
50:54up all those libraries is very fast you
50:56know like broadcasting slicing all these
50:59basic stuff that ends up being so
51:01valuable in deep learning right nobody
51:03knew at the time so when when we
51:05actually did the machine learning course
51:07after that I could just do all these
51:09assignments like 10x faster than the Cs
51:12people because they would be writing for
51:13Loops for
51:14everything because they would write code
51:16in Python like how they write code in
51:18C++ except you don't need to and like
51:20that's a skill right and the other thing
51:23is uh in E like
51:25you're you're basically like learning
51:28all these like uh sigul and systems
51:30linear algebra
51:32probability like lot faster like earlier
51:35on in the curriculum that uh you just
51:39get so you just internalize it a lot
51:41more and when you're reading ml like it
51:44comes more naturally to you same thing
51:46like we had a course on Control Systems
51:48which is basically a flip of RL instead
51:51of reward it's like cost and uh all that
51:55helps you much much more you just you
51:57just are able to move much faster than
51:58the Cs people great so apart from your
52:01curriculum was there something that you
52:02did that was that you felt significantly
52:04pushed you into this field in into AI
52:07yeah just the M yeah actually
52:11mostly I I do remember doing a contest
52:14an ml contest that was also conducted by
52:16some friend at CS Department that was
52:19actually before even I took the ML Class
52:22they said there's this contest you know
52:23there's this thing there's this a
52:25library called uh psychic learn just
52:29take it and like try a bunch of random
52:30algorithms and try to win the contest it
52:33was perfect trial and error I had no
52:36idea you know just importing different
52:38libr uh Al like you know decision trees
52:41and random Forest tweaking hyper
52:44parameters until I got the number one on
52:45the leaderboard but turn out that's
52:48exactly how ml works right so so that
52:51that that really helped me too that that
52:53experience was surreal when I
52:56first thanks uh hello so hey he so when
53:03you decided to step into the world of
53:05startups you took care of the tech you
53:08took care of AI who took care of the
53:10business side of it we didn't have a
53:13business right like when we started we
53:16we just had to launch a
53:17product uh in fact I wouldn't say I I
53:21alone took care of the AI like we have a
53:23CTO and like he took care of it more
53:25than me I took care more of the business
53:28actually in the beginning if if by
53:29business you mean like company
53:32operational uh fundraising hiring I I
53:35took care of it myself uh but that's
53:38actually not even business that's just
53:40running the company business means
53:42actually like figuring out revenue
53:43streams and
53:45Partnerships for which we recently have
53:47one more person now so is it because at
53:50that time you were the best person who
53:53could take care of that side because you
53:55had CTO and everybody else you have such
53:58a good profile a background in Tech
54:01that's right that's right so we we
54:04decided the split based
54:06on uh who was best position for that
54:09role at that
54:10point like you know also it's a good
54:12skill to be able to Market what exists
54:15well and also at the same time make the
54:18right technical decisions you know like
54:20a sweet spot of that's why I emphasize a
54:23lot on like soft skills too it's pretty
54:27important and and uh yeah I I for
54:30example my CTO is way better at me at
54:33programming I'm not bad at programming
54:35at all but he's just way
54:37better okay that answers my question
54:39thank you AR uh Google was found here
54:43here in the front row oh yeah Google was
54:45founded in 98 and uh 25 years later it
54:50is still a recognizable brand so in the
54:53year 204 eight do you expect perplexity
54:57to still be a recognized brand and to
55:00get there what are things you need to do
55:03today I mean how do you even care about
55:05that idea I I hope so right obviously
55:09why why would I say no to this but
55:13uh what do I need to do today for that I
55:16think we just need to First survive for
55:18five to 10 years if we get there if we
55:21reach the 5 to 10 year mark I think we
55:24can definitely start thinking more long
55:25term there's like this uh thing I read
55:28recently that for a company it's like
55:31halflife halflife concept if you survive
55:34for one year you can plan for the next
55:36one year if you survive for two years
55:37you can plan for the next two years so
55:39uh if you survive for 5 years or 10
55:41years then we can plan for the next
55:42decade
55:47itself um uh so uh over the so and so
55:52okay and so over the last the last the
55:56last year AI has moved forward by Leaps
55:58and Bounds uh AI tools like Char gbt and
56:01purpose have become insanely good and of
56:04course all this has led to talk of how
56:06uh AI is going to steal our jobs uh
56:08people bring up terminator codes and all
56:10that stuff and um so uh and my question
56:13is as um as college students um what can
56:17we do to ensure that uh years down the
56:19line when we uh when we
56:22graduate what can we do now to ensure
56:24that we can actually add value to the
56:26world
56:28that that we can that we cannot be other
56:31place by by $20 chat GPT and uh that's
56:34what I wanted to ask okay yeah um I
56:37would
56:38say you just uh you I mean it's a good
56:43question number one obvious answer is
56:46use these tools yourself make yourself
56:48more
56:49productive uh be able to do a lot more
56:52things per day than people who are not
56:54using these tools like for example
56:56you're writing code like try to get the
56:58same task done in like 2 hours instead
57:01of like 8 hours don't debug manually
57:04like use whatever tool you like chat GPT
57:07perplexity B and uh you know if you're a
57:12programmer that is or GitHub co-pilot
57:15and
57:16uh the other thing I would say is human
57:19judgment is still super valuable like
57:21these tools are great at uh generating
57:24content but taste human taste is
57:27incredibly valuable you know there's
57:29this funny video of Rick ruin where they
57:31ask him do you know music do you know
57:33any instruments like do you have any
57:35technical knowledge and he just says no
57:37so then what what what are people like
57:38why are they why do you even have fans
57:40like why are people listening to your
57:42music and he says because I have
57:44incredibly good taste and I think
57:46that'll be very in fact way more
57:49valuable than what it is
57:53today so judgement and Cas matters I I
57:57have a small question I mean first I
57:59wish to congratulate you Arin for
58:00whatever you're doing thank you great
58:02work I've been sitting next to your
58:04professor and been hearing what work you
58:06did in the ins and you got nine papers
58:07through this ins before you left so it
58:09is phenomenal so congratulations but
58:12having said this I
58:14[Applause]
58:17okay having said this I see a little bit
58:19of a mixed message that you're giving to
58:21the young audience I never seen this
58:23audience so full like this so that is
58:25also a big sign that you know you are
58:27appealing to a large uh you know section
58:29of the you know cool guys the Next
58:32Generation guys of this institute but I
58:34think there's a little bit that you have
58:36to care carefully uh you know share see
58:40is 80% right and getting it into the
58:43market is okay for the digital
58:45Revolution but I probably think boing
58:47did the same and release the max and I
58:50think it's a disaster so we have to be a
58:53little careful Where We Are apply this
58:5580% and why we don't apply this 80%
58:57that's fair that's fair so I think this
58:59message has to be little clear to the
59:00younger generation who may be getting
59:02influenced early in life and you know
59:06Boeing was run by Engineers if some of
59:08you read the stories Yeah by brilliant
59:11Engineers of the of the world but today
59:13it's not run by Engineers it's a totally
59:15transformed organization run by
59:17marketing people who wanted to put into
59:19the market the product before it was
59:21ready totally fair so I guess the C
59:25to everything I said in this slides is
59:27that it applies to if you're building
59:30software products right definitely like
59:33you when you're building something in
59:35the physical
59:36world uh you cannot be 80% right because
59:40the mistakes are too expensive right
59:42even if you're working on S driving cars
59:44you cannot be 80% accurate right the
59:47longtail matters way more there uh so
59:51yeah also another another that's another
59:53heris stic if you if you want to working
59:55software use cases big use cases that
59:58are where like you know the 80% accuracy
01:00:01is tolerable like it's not risky to any
01:00:03human life if you're 20% inaccurate to
01:00:06start with and and this this product is
01:00:08one such example in general like most
01:00:11software is pretty harmless in terms of
01:00:13going
01:00:15wrong hey uh I have a question uh so
01:00:18what do you think is the bottleneck for
01:00:21these kinds of AI agents uh when it
01:00:23comes to Hardware because I'm starting
01:00:25up right now and I'm I want to build the
01:00:27Boston Dynamics of India or something
01:00:29like figure robotics you mean robots
01:00:31yeah uh so now now that I've seen that
01:00:33perplexity is tying up with rabbit
01:00:34Hardware device and uh there are such
01:00:36more devices like tab so do you think
01:00:39it's a bottleneck that without Hardware
01:00:40it will not reach the master properly or
01:00:42smartphone is
01:00:43enough I mean smartphone will definitely
01:00:46take us far or even laptops for that
01:00:49matter uh but AI native consumer devices
01:00:53will be like creating new experiences
01:00:57that were not possible before right uh
01:01:00leaving aside like rabbit or Humane or
01:01:02tab um the the stuff you want to work on
01:01:05robots it's actual physical labor right
01:01:08it's not possible with a rabbit or an
01:01:10iPhone um so the next big thing is going
01:01:13to be human art robots like robots and
01:01:15human F factor with human dexterity and
01:01:19like physical understanding Common Sense
01:01:21visual common sense and all the is like
01:01:24still challenge so what is the
01:01:26bottleneck actually I think the
01:01:28bottleneck is still an Al at an
01:01:30algorithmic level like we don't have the
01:01:33CH GPT of Robotics yet right y so until
01:01:37that
01:01:38happens um we're not going to see them
01:01:40in Mass market yet and is the bottle
01:01:43like in Hardware I don't think so yeah
01:01:45as Tesla Optimus as you've seen even
01:01:48though it's thly operated the hardware
01:01:50looks phenomenal and like pretty close
01:01:51to done right y yep thanks
01:01:55another question here here
01:02:12hello yeah so CH so chat GPT is a
01:02:15product it's I guess you meant gbt
01:02:173.5 uh yeah we are moving away from gbd
01:02:203.5 uh it's a pretty old model by now uh
01:02:24and and uh you we have already seen
01:02:27through evaluations that a llama fine
01:02:29tune version is better than that and uh
01:02:32and then if we go even further and take
01:02:34one of these models like mistra and F
01:02:36tune that it's going to be even better
01:02:38than llama so in the interest of serving
01:02:42users with a better model not in the
01:02:44interest of moving away from open but
01:02:46more in the interest of improving the
01:02:47product we're moving away from
01:02:493.5 we're not moving away from four
01:02:52though uh because four is still a better
01:02:55model than any llama or M in tun today
01:02:58so it'll still be there on the
01:03:01product yeah on the PO plan yes we not
01:03:05on the free plan hello last three
01:03:08questions hello yeah um can I
01:03:13go
01:03:15yeah can I go yeah so congratulations um
01:03:19thank you we all know that Google search
01:03:21engine was the 17th organization
01:03:23globally new kind of made them dance
01:03:25today uh what I really want to
01:03:27understand from you and whil you said
01:03:29next 5 years is going to be very crucial
01:03:31to make your mark could you tell us what
01:03:33are those top two things that you would
01:03:35like to do in the next 5 years to make
01:03:37your mark actually it's very simple uh
01:03:41you know we just need to be focusing on
01:03:43accuracy and
01:03:44speed and this is exactly what Google
01:03:47did ignored most of the other things
01:03:50that you could be doing and just focus a
01:03:52lot on making the product faster and
01:03:54making the product more accurate and if
01:03:57you wanted to add one more thing here
01:03:58it's distribution that's getting in the
01:04:01hands of as many people as possible and
01:04:03making them use it and if you can nail
01:04:06these three things continuously over a
01:04:08sustained period of time then you'll be
01:04:10very far ahead in after 5
01:04:14years hello sir
01:04:16llm llm hallucinate a lot right uh and
01:04:20hallucination yeah the the current
01:04:23architectures they hallucinate and they
01:04:25can't continually learn anything do you
01:04:27think there are there can be new
01:04:28architectures that can you know resolve
01:04:31these issues or yeah definitely I I
01:04:33think so uh definitely like models that
01:04:37can decouple memorization and
01:04:40reasoning will be way better at than the
01:04:43current system for hallucinations or
01:04:47memor like you know just saying
01:04:50arbitrary
01:04:52things definitely for sure but that's
01:04:54more like research progress that needs
01:04:55to be made right
01:04:57yeah hello sir first of all I have two
01:05:00questions to ask I've been using
01:05:02perplexity from the last one year and
01:05:05the beginning I saw that you guys also
01:05:07provide the sources where you guys get
01:05:08the data now we have B also doing the
01:05:11same thing and we have Google Gemini in
01:05:12the beta phase and we have gp4 also so
01:05:15what do you think the competive edge
01:05:16perity a has over these all llms I I
01:05:20believe we our Edge today over those
01:05:23other tools is the speed and accuracy
01:05:26like our default mode is like the
01:05:27fastest version in terms of the rag plus
01:05:30llms and our co-pilot is the most
01:05:34accurate or comprehensive version and so
01:05:36we offer these the the best
01:05:37interpolation of speed and accuracy but
01:05:40that said like we will be adding a lot
01:05:42more into our product that makes it
01:05:44sufficiently differentiated from Bard
01:05:46and Chach PT that uh you know we
01:05:49basically don't want to be the same like
01:05:52commoditized version yeah and one more
01:05:55question I have uh so during October de
01:05:58day GPT CH GPT announced about their GPT
01:06:01store and GPT Builder yeah like everyone
01:06:03who is using jpt 4 as subscription bis
01:06:06they can their own jpts so how do you
01:06:08think that how much sustainable that
01:06:10thing is and how is going to change the
01:06:12whole economy of gpts when everyone will
01:06:14be able to make their own gpts on their
01:06:16own yeah uh I'm I'm more skeptical of
01:06:19the GPT store uh I don't think any one
01:06:23particular GPT is so widely useful yet
01:06:27that you don't that you cannot get the
01:06:29value out of it just through the
01:06:30traditional CH GPT
01:06:33interface but uh I believe that like you
01:06:36know people if if you can do more
01:06:38complicated things than just a system
01:06:40prompt like actual say book tickets or
01:06:43something like that then I think that
01:06:46can become useful but that doesn't work
01:06:49yet today there so it's still not there
01:06:52yet yeah thank
01:06:56you
01:07:00last um m test hello sir can you hear me
01:07:04yeah I can hear you uh thank you for the
01:07:07presentation and the lecture it was
01:07:08thank you it was nice um I'm not
01:07:11entirely sure if this is a very
01:07:13well-framed question but there was a
01:07:15section on the presentation which talked
01:07:18about irrationality and how it helps us
01:07:22uh build conviction in the case there
01:07:24are negative self-doubts or questions on
01:07:27what you're
01:07:27doing uh I believe that every uh person
01:07:32does have a rational and irrational site
01:07:34and uh whichever is in majority
01:07:37classifies a person as a rational and
01:07:38irrational thinker so and it's hard for
01:07:42rational thinkers to build startups or
01:07:44companies uh as you said because it's
01:07:47hard to ignore the consequences of these
01:07:49questions
01:07:51so like if a rational person does want
01:07:54to be as successful or more
01:07:57successful and build an amazing
01:08:00startup how does he convince the
01:08:03irrational part of his mind to take over
01:08:06in these times where these self-doubts
01:08:08and convictions prevent us from actually
01:08:11doing what we are supposed to do yeah I
01:08:14mean I I agree with you there noan to
01:08:16saying someone is a rational person was
01:08:18a rational person was just being
01:08:20rational irrational about one particular
01:08:22thing right what I meant is if you if
01:08:25you're if you're just rational and
01:08:26wanted to make a lot of money in your
01:08:27life doing a startup is not the best way
01:08:30to go about it right um because it's
01:08:33very risky and like there's a lot of
01:08:36pain tolerance required so I would say
01:08:38if a rational person has sufficient pain
01:08:41tolerance then they can definitely like
01:08:43invoke the irrational part of their
01:08:45brain and go ahead and try to do
01:08:47it thank you
01:08:52sir
01:08:54all right I think that's all for the
01:08:56questions but thank you so much
01:08:58[Applause]
01:09:09everybody thank you sir for that
01:09:11insightful talk now I will invite Prof
01:09:14mahes Dean alni and corporate relations
01:09:17and Mr kavaj n and Professor rendan to
01:09:20please facilitate Mr Dr Irving
01:09:32[Applause]
01:09:52Nas
01:10:22back
01:10:27ladies and gentlemen as we B farewell to
01:10:30this enriching experience it's evident
01:10:33that our journey is gracefully con
01:10:35concluded a big thank you to Dr Irvin
01:10:38shinas for sharing insightful
01:10:39experiences today his wisdom has left a
01:10:42lasting impact thanks to our fantastic
01:10:45audience for making this event memorable
01:10:48cheers to New Perspectives and shared
01:10:49movements thank
01:10:52you
01:10:54uh I
01:11:22will
🎥 Related Videos
What vaccinating vampire bats can teach us about pandemics | Daniel Streicker
a16z Podcast | Things Come Together -- Truths about Tech in Africa
2024 TSCRS Applications of anterior segments diagnostic instruments in cataract surgery
a16z Podcast | The Infrastructure of Total Health
The Robot Lawyer Resistance with Joshua Browder of DoNotPay
NES Controllers Explained
🔥 Recently Summarized Examples
4 Steps to Master Any Complex Skill (quickly)
40 Years of Fitness Experience in Less Than 11 Minutes.
Gun Controlling Media Makes FATAL Mistake... They Have Tied Their Fate To Biden's & Gun Rights Win
GET READY! Palantir Is Officially The Next Nvidia.
Abundant Thinking: The Hidden Key to Get Everything You Want (Audiobook)