YouSaid · the spoken record

Aravind Srinivas

lines on the record
225
first
2024-06-19
most recent
2024-06-19
sittings or episodes
1
sources
podcast

Every line below is reproduced as it was said and linked to the record it came from. Nothing here is summarised or generated. Directory · Search · Corrections

  1. “Core gigawatt. I mean, the point I clearly made in the poll was equivalent. So it doesn't have to be clearly millionaires wonders, but... It could be fewer GPUs of the next generation that match the capabilities of the Million H100s. At lower power consumption, great. Whether it be one gigawatt or 10 gigawatt, I don't know, right? So it's a lot of power, energy. Think the kind of things we talked about on the inference compute being very essential for future highly capable AI systems or even to explore all these research directions like models bootstrapping of their own reasoning, doing their own inference, you need a lot of GPUs.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  2. “I think it was like Google OpenAI Meta X. Obviously, OpenAI is not just OpenAI, it's Microsoft Doo. And Twitter doesn't let you do polls with. More than four options. So ideally, you should have added anthropic or Amazon two in the mix. Million is just a cool number”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  3. “The way I want to point out like a lot of people said it's not just open AI, it's Microsoft, and that's a fair counterpoint to that.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  4. “That's the kind of problems you need to solve whether you want to keep look, look, it's a whole reason it's called elastic. Some of these things can be scaled very gracefully, but other things so much not like GPUs or models, like you need to still make decisions on a discrete basis.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  5. “Now, I think he has this mentality that it all has to be in But it frees you from working on problems that you don't need to be working on when you're scaling up your startup. AWS infrastructure is amazing. It's not just amazing in terms of its quality. It also helps you to recruit engineers easily because if you're on AWS and all engineers are already trained using AWS. So the speed at which they can ramp up is amazing.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  6. “There was a clear competition between YouTube and, of course, prime video is also a competitor, but it's sort of a thing that, for example, Shopify is built on Google Cloud. Snapchat uses Google Cloud. Walmart uses Azure. So there are examples of great internet businesses that do not necessarily have their own data centers. Facebook have their own data center, which is okay. They decided to build it right from the beginning. Even before Elon took over Twitter, I think they used to use AWS and Google for their deployment.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  7. “It keeps changing the dynamics. By the way, everything's on cloud, even the mottos we serve are on some cloud provider. It's very inefficient to go build your own data center right now at the stage we are. I think it will matter more when we become bigger. But also companies like Netflix still run on AWS and have shown that you can still scale with somebody else's cloud solution.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  8. “Yeah, I mean, you got to make decisions like should I go spend like 10 million or 20 million more and buy more GPUs or should I go and pay another model providers like 5 to 10 million more and get more compute capacity from them?”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  9. “Like llama based models, we can work on it ourselves by optimizing at the kernel level, right? So there we work closely with NVIDIA, who's an investor in us, and we collaborate on this framework called Tensor RTLLM. And if needed, we write new kernels, optimize things at the level of making sure the throughput is pretty high without compromising on latency.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  10. “Could have certain queries that are at the tail failing more often without you even realizing it. And that could frustrate some users, especially at a time when you have a lot of queries, suddenly a spike. So it's very important for you to track the tail latency. And we track it at every single component of our system, be it the search layer or the LLM layer. In the LLM, the most important thing is the throughput and the time to first token. We usually is referred to as TTFT time to first token. And the throughput, which decides how fast you can stream things, both are really important. And of course, for models that we don't control in terms of serving, like OpenAI or Anthropic, we are reliant on them to build a good infrastructure. And they are incentivized to make it better for themselves and customers. So that keeps improving. And for Modos, we survive.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  11. “Inspiration from Google. There's this whole concept called Tay latency. It's a paper by Jeff Dean and one other person where it's not enough for you to just test a few queries, see if there's fast and conclude that your product is fast. It's very important for you to track the P90 and P99 latencies. Which is like the 90th and 99th percentile. Because if a system fails, 10% of the times you have a lot of servers.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  12. “We are not taking off the shelf models from anybody. We have customized it for the product. Whether we own the weights for it or not is something else. I think there's also power to Designing the product to work well with any model. If there are some idiosyncrasies of any model shouldn't affect the product.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  13. “Don't care That doesn't mean we're not going to work towards it, but this is where the model agnostic viewpoint is very helpful. The user care if perplexity Perplexity has the most dominant model in order to come and use the product. No. Does the user care about a good answer? Yes. So whatever model is providing us the best answer, whether we fine tuned it from somebody else's base model or a model we host ourselves, it's okay.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  14. “Going to be faster than clot models or 4.0 because we are pretty good at inferencing it ourselves. Like we hosted and we have a cutting edge API for it I think it still lags behind from GPD4 today in some finer queries that require more reasoning and things like that. But these are the sort of things you can address with more post training, RLHF training and things like that. And if you're working on it.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  15. “The model we train ourselves. We took llama three and we post trained it to be very good at fuel skills like summarization, referencing citations, keeping context, and longer context support. So that was called SOR.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  16. “Traditional search cannot because they're all about physician and recall simultaneously. In Google, even though we call it 10 blue links, you get annoyed if you don't even have the right link in the first three or four. I so tuned to getting it right. LMs are fine. You get the right link maybe in the 10th or 9th, you feed it in the model. know that that was more relevant than the first so that that that that flexibility allows you to like rethink where to put your resources in in terms of whether you want to keep making the model better or whether you want to make the retrieval stage better it's a trade-off in computer science it's all about trade-offs right at the end”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  17. “LLMs really help there. So, what albums add? Even if your initial retrieval doesn't have an amazing Set of documents There's really good recall, but not as high a precision. LLMs can still find a needle in the haystack.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  18. “But not in a perquiry basis. You can do that too when you're small just to delight users, but it doesn't scale. You're obviously going to, at the scale of queries you handle as you keep going in the logarithmic dimension, you go from 10,000 queries a day to 100,000 to a million to 10 million. You're going to encounter more mistakes. So you want to identify fixes that address things at a bigger scale.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  19. “And this really depends on the query category. And that's why I search is a lot of domain knowledge involved problem. That's why we chose to work on it. Everybody talks about wrappers, competition models. There's this insane amount of domain knowledge you need to work on this. And it takes a lot of time to build up towards a highly Really good index with like really good ranking. All these signals.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  20. “You need a combination of all a hybrid. And you also need other ranking signals outside of the semantic or word-based. This is like page ranks like signals that score domain authority and recency, right?”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  21. “Like when OpenAI released their embeddings, there was some controversy around it because it wasn't even beating BM25 on many, many retrieval benchmarks, not because they didn't do a good job. BM25 is so good. So this is why just pure embeddings and vector spaces are not going to solve the search problem. You need the traditional term-based retrieval. You need some kind of N-gram-based retrieval.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  22. “Don't always have to be stored entirely in vector databases. There are other data structures you can use. And other forms of traditional retrieval that you can use. There is an algorithm called BM25 precisely for this, which is a more sophisticated version of TFIDF. TFIDF is term frequency times inverse document frequency, a very old school information retrieval system that just works actually really well even today. And BM25 is a more sophisticated version of that. you know beating most embeddings on ranking”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  23. “Ranking part that depending on the query you ask fishes the relevant documents from the index and some kind of score. And that's where you have billions of pages in your index. You only want the top K, you have to rely on approximate algorithms to get you to the top K.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  24. “It's not purely a vector space. It's not like once the content is fetched, there is some BERT model that runs on all of it and puts it into a big gigantic vector database. Retrieve from it's not like that because packing all the knowledge about a web page into one vector space representation is very, very difficult. There's like, first of all, vector embeddings are not magically working for text. It's very hard to understand what's a relevant document to a particular query. Should it be about the individual in the query or should it be about the specific event in the query? Or should it be at a deeper level about the meaning of the query such that the same meaning applying to a different individual should also be retrieved? You can keep arguing, right? Like what should a representation really capture? And it's very hard to make these vector embeddings have different dimensions be disentangled from each other and capturing different semantics. So what retrieve will typically, this is the ranking part, by the way. There's an indexing part assuming you have like a post-process version per URL. And then there's a”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  25. “And then you also need to decide the periodicity with which you recrawl. And you also need to decide what new pages to add to this queue based on hyperlinks. So that's the crawling. And then there's a part of building, fetching the content from each URL. And once you did that through the headless render, you have to actually build the index now. And you have to reprocess. You have to post process all the content you fetched, which is the raw dump into something that's ingestible for a ranking system. So that requires some machine learning, text extraction. Google has this whole system called now boost that extracts the relevant metadata and like relevant content from each raw URL content.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  26. “Yeah, it depends if some publishers allow that so that they think it'll benefit their ranking more. Some publishers don't allow that. You need to Keep track of all these things per domains and subdomains”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  27. “Like even deciding what to put in the queue, which web pages, which domains, and how frequently all the domains need to get crawled. And it's not just about knowing which URLs. It's just deciding what URLs to crawl, but how you crawl them. You basically have to render headless render. And then websites are more modern these days. It's not just the HTML. There's a lot of JavaScript rendering. You have to decide what's the real thing you want from a page. And obviously people have robots that text file. And that's like a politeness policy where you should respect the delay time so that you don't like overload their service by continually crawling them. And then there's stuff that they say is not supposed to be crawling stuff that they allow to be crawl. And you have to respect that.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  28. “So indexing is multiple parts. Obviously, you have to first build a Google has Googlebot, we have Perplexity bot, Bing bot. There's a bunch of bots that crawl the web.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  29. “The model is smart enough, it'll know that I specifically said these are ways a model can go wrong and it'll use that and say,”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  30. “But in such a case, if a model is skillful enough, it should just say, I don't have enough information. So there are like multiple dimensions where you can improve a product like this to reduce hallucinations, where you can improve the retrieval, you can improve the quality of the index, the freshness of the pages in the index, and you can include the level of detail in the snippets. You can include the model's ability to handle all these documents really well. And if you do all these things well, you can keep making the product better.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  31. “So you retrieve the right documents, but the information in them was not up to date, was stale or not detailed enough. And then the model had insufficient information or conflicting information from multiple sources and ended up getting confused. And the third way it can happen is you added too much detail to the model. Like an index is so detailed, the snippets are so, you use the full version of the page. And you threw all of it at the model and asked it to arrive at the answer. And it's not able to discern clearly what is needed and throws a lot of irrelevant stuff to it. And that irrelevant stuff ended up confusing it. And made it like a bad answer. So all these three, or the fourth way is you end up retrieving completely irrelevant documents”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  32. “Yeah, there are multiple ways it can happen. One is you have all the information you need for the query. The model is just not smart enough to understand the query at a deeply semantic level and the paragraphs at a deeply semantic level and only pick the relevant information and give you an answer. So that is the model skill issue. That can be addressed as models get better, and they have been getting better. Now, the other place where hallucinations can happen is you have poor snippets, like your index is not good enough.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  33. “Otherwise, you can still end up saying nonsense or use the information in the documents and add some stuff of your own, right? Despite these things still happen, I'm not saying it's foolproof.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  34. “Yeah. So Rag is retrieval augmented generation, simple framework. Given a query, always retrieve relevant documents and pick relevant paragraphs from each document and use those documents and paragraphs to write your answer for that query. Principle and perplexity is you're not supposed to say anything that you don't retrieve, which is even more powerful than rag because rag just says, okay, use this additional context and write an answer, but we say don't use anything more than that do. That way we ensure factual grounding. And if you don't have enough... Information from documents to retrieve, just say we don't have enough search results to give you a good answer.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  35. “Except, like, you know, it's not easy for them to serve that mission anymore. And nothing stops other people from adding on to that mission. Rethink that mission too, right? Wikipedia also, in some sense, does that. Does organize the information around the world and makes it accessible and useful in a different way. He does it in a different way, and I'm sure there'll be another company after us that does it even better than us. And that's good for the world.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  36. “We felt like that is a mission that's bigger than competing with Google. You never make your mission or your purpose about someone else because you're probably aiming low, by the way, if you do that. Want to make your mission or your purpose about something that's bigger than you and the people you're working with. And that way you're working, you're thinking. Like completely outside the box too. And Sony made it their mission to put Japan on the map, not Sony on the map.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  37. “And most importantly, a lot of people clicking on the related questions too. So we came up with this vision. Everybody was asking me, okay, what is the vision for the company? What's a mission? I had nothing, right? Like it was just explore cool search products. But then I came up with this mission along with the help of my co-founders that, hey, it's not just about search or answering questions, it's about knowledge, helping people discover new things and guiding them towards it, not necessarily giving them the right answer, but guiding them towards it. And so we said, we want to be the world's most knowledge-centric company. Was actually inspired by Amazon saying they wanted to be the most customer centric company on the planet. We want to obsess about knowledge and curiosity”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  38. “Because there's no need for people when they hang out, their family and chilling and vacation to come use a product by a completely unknown startup with an obscure name, right? So I thought there was some signal there. And okay, we initially didn't have it conversational. It was just giving you only one single query. You type in, you get an answer with summary, with the citation. Had to go type a new query if you wanted to start another query. There was no conversational or suggested questions, none of that. So we launched the conversational version with the suggested questions a week after New Year. And then the usage started growing exponentially.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  39. “Honestly, the way we thought about it was let's release this. There's nothing to lose. It's a very new experience. People are going to like it. And maybe some enterprises will talk to us and ask for something of this nature for their internal data. And maybe we could use that to build a business. That was the extent of our ambition. That's why most companies never set out to do what they actually end up doing. It's almost like accidental. So, for us, the way it worked was we put this out, and a lot of people started using it. I thought, okay, it's just a fad, and the usage will die. But people were using it in the time, we put it out on December 7, 2022. People were using it even in the Christmas vacation. I thought that was a very powerful signal.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  40. “Like at least some amount of relevance. But we knew that's not like that's like a one time thing. It's not like every way is repetitive query, but at least that gave us the confidence that there is something to pulling up links and summarizing it. And we decided to focus on that. And obviously we knew that this Twitter search thing was not scalable or doable for us because Elon was taking over and he was very particular that he's going to shut down API access a lot. So it made sense for us to focus more on regular search.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  41. “People laughed at, they would like, or like they either are spooked by it saying, Oh, this AI knows so much about me, or they were like, oh, look at this AI saying all sorts of shit about me. And they would just share the screenshots of that query alone. And that would be like, what is this AI? Oh, is this thing called perplexity? And you go, what do you do is you go and type your handle at it and it'll give you this thing. And then people started sharing screenshots of that and Discord forums and stuff. And that's what led to this initial growth when you're completely irrelevant.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  42. “So, our first, like the reason, the first release of perplexity went really viral because people would just enter their social media handle on the perplexity search bar. It's really fun. We release both the Twitter search and the regular Perplexity A week apart, and we couldn't index the whole of Twitter, obviously, because we scraped it in a very hacky way. And so we implemented a backlink where if your Twitter handle was not on our Twitter index, it would use our regular search that would pull up a few of your tweets. And give you a summary of your social media profile I will come about hilarious things because back then it would hallucinate a little bit too.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  43. “Think there is something powerful about showing something that was not possible before. There is some element of magic to it. And especially when it's very practical too, you are curious about what's going on in the world, what's the social interesting relationships, social graphs. I think everyone's curious about themselves. I spoke to Mike Kreiger, the founder of Instagram, and he told me that Even though you can go to your own profile by clicking on your profile icon on Instagram, the most common search is people searching for themselves on Instagram.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  44. “Concentrated. And then we built this demo where you can ask all these sort of questions top like tweets about AI. If I wanted to get connected to someone, I'm identifying a mutual follower. And we demoed it to a bunch of people like Jan Lekhan, Jeff Dean, Andre. They all liked it because people like searching about what's going around about them, about people they are interested in. Fundamental human curiosity, right? That ended up helping us to recruit good people because nobody took me or my co founders that seriously, but because we were backed by interesting individuals, at least they were willing to listen to a recruiting pitch.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  45. “So we built all this into a good search experience over Twitter, which we scraped with academic accounts before Elon took over Twitter. So we, you know, back then, Twitter would allow you to create academic API accounts. And we would create lots of them with generating phone numbers, like writing research proposals with GPT. Would call my projects as like brin rank and all these kind of things. And then create all these fake academic accounts, collect a lot of tweets, and like. Basically, Twitter is a gigantic social graph. We decided to focus it on interesting individuals because the value of the graph is still pretty sparse.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  46. “So it's almost like you should consider it was like programming with computers that had very little RAM. A lot of hard coding. Like my co-founders and I would just write a lot of templates ourselves for like this query, this is SQL, this query, this is a SQL. We would learn SQL ourselves. This is also why we built this generic question answering bot because we didn't know SQL that built ourselves. So, and then we would do rag, given the query, we would pull up templates that were similar looking template queries. And the system would see that build a dynamic few shot prompt and write a new query for the query you asked and execute it against the database. And many things would still go wrong. Sometimes the SQL would be erroneous. You have to catch errors. You have to do retries.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  47. “The reason we picked SQL was because we felt like the output entropy is lower. It's templatized Only a few set of select statements, count, all these things. And that way you don't have as much entropy as in generic Python code. But that insight turned out to be wrong, by the way.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  48. “Yeah, most recent tweets that were liked by both Elon Musk and Jeff Bezos. Couldn't ask these questions before because you needed an AI to understand this at a semantic level, convert that into a structured query language, execute it against a database, pull up the records and render it, right? Was suddenly possible with advances like GitHub Copilot. Had code language models that were good. And so we decided we would identify this inside and go again search over scrape a lot of data, put it into tables, and ask questions.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  49. “When we decided, okay, how would it look like if we disrupted or created search experiences over things you couldn't search before? Said, okay, tables, relational databases. You couldn't search over them before. But now you can because you can have a model that looks at your question, translated it to some SQL query, runs it against the database, you keep scraping it so that the database is up to date, and you execute the query, pull up the records and give you the answer.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source

  50. “And I always wanted my start also to be of this nature where. But it wasn't designed to work on consumer search itself. We started off with searching over the first idea I pitched to the first investor who decided to fund this Elod Gill. Hey, would love to disrupt Google, but I don't know how. But one thing I've been thinking is. People stop typing into the search bar and instead just ask about whatever they see visually. A glass I always like the Google Glass version. It was pretty cool He just said, Hey, look, focus, you're not going to be able to do this without a lot of money, a lot of people. Identify a wedge right now and create something. And then you can work towards the grander vision, which is very good advice.”

    2024-06-19 · Lex Fridman Podcast · #434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · IDENTIFIED FROM THE TRANSCRIPT · source