Transcript
00:00:00How's everybody doing?
00:00:13Good?
00:00:14All right.
00:00:16Pretty crazy times we live in.
00:00:17Oh, my microphone.
00:00:19Okay, I am very excited to tell you all
00:00:21about Perfect Search, built for AI agents.
00:00:25And I'm going to say a lot of crazy things in this talk,
00:00:27so bear with me.
00:00:28They are all true.
00:00:30And I will tell you what is EXA,
00:00:32the story of EXA, so how we got here,
00:00:34and then where we're going.
00:00:36Okay, cool.
00:00:38If you take away anything from this talk,
00:00:40it is this slide.
00:00:42This is showing web searches per day
00:00:44over the past 30 years from humans and AIs.
00:00:47Obviously, it was all humans until around 2020s,
00:00:52and now we're in 2026.
00:00:53And actually, this year,
00:00:55we expect the number of searches from AI systems
00:00:57to exceed that of humans.
00:00:59Pretty crazy.
00:01:00And then in the next few years,
00:01:01it should be 1,000 times more.
00:01:03So AI systems, AI products,
00:01:05whatever AI system you use,
00:01:07some together will search 1,000 times more than humans.
00:01:10That's a pretty crazy world that we're getting into.
00:01:13And the entire ecosystem of search
00:01:15is changing because of it.
00:01:18So EXA is the search engine for AI agents.
00:01:21You know, we were the first ones to be like,
00:01:22we were the first search engine for AI,
00:01:24and now things are getting kind of wild.
00:01:27We now serve over 5,000 companies,
00:01:30over 400,000 developers.
00:01:31I see some customers in the audience.
00:01:33We serve a very diverse set of agents
00:01:36from coding agents like Cursors.
00:01:37If you use Cursor, at some point,
00:01:39the Cursor agent decides to search
00:01:40for the latest technical documentation or the news.
00:01:42It'll be using Exxon under the hood.
00:01:43We serve go-to-market agents like HubSpot,
00:01:45so we help their users get, you know,
00:01:48really high-quality lists of companies to sell to.
00:01:51We serve all sorts of financial agents.
00:01:52I was in New York a few weeks ago,
00:01:54and pretty much everyone there
00:01:55is now building financial agents,
00:01:56and they need the best financial data.
00:01:58Yeah, and really just a huge diversity of agents
00:02:00from labs to YC startups.
00:02:05Okay, but how'd we get here?
00:02:07So what's the why of EXA?
00:02:08To me, that's always been the most important.
00:02:10And there are a lot of ways to say the problem,
00:02:12but the short way is just misinformed anarchy.
00:02:15This is what we're trying to avoid.
00:02:16We want to create the opposite of this.
00:02:18So what does this mean?
00:02:19Well, this is the Internet,
00:02:21or it's a visual depiction of the Internet.
00:02:23You know, you've got a bunch of pages.
00:02:24You've got blog posts.
00:02:25You've got company websites.
00:02:26You've got images.
00:02:26You've got tweets.
00:02:27You've got all sorts of things.
00:02:29And the Web is really, really big, right?
00:02:33This is showing a few thousand pages.
00:02:35The Web is, you know,
00:02:36on the order of a trillion pages,
00:02:37so, you know, a million times bigger than this.
00:02:40And it contains a huge amount
00:02:42of the world's information.
00:02:43And that means that, you know,
00:02:45if this is just readily available,
00:02:46if you could go to any link
00:02:47and just get all the world's information,
00:02:48then I'm sure we all walk around
00:02:50with, like, deep understanding of everything, right?
00:02:52Obviously not.
00:02:54It's messy.
00:02:55And so it's crazy,
00:02:56and it can't fit in our heads.
00:02:57So we need information tools
00:02:59that could help synthesize this
00:03:00or filter it into the things you need to know.
00:03:03And, you know,
00:03:04we have a tool called Google,
00:03:06and it's a pretty solid tool.
00:03:07It could get you things like the Costco homepage
00:03:08or information about Taylor Swift
00:03:10or whatever you want to search.
00:03:11But it's not perfect, right?
00:03:13And I love this example
00:03:14where you type shirts without stripes,
00:03:16and if you notice,
00:03:17you get shirts with stripes.
00:03:18Why is it doing that?
00:03:19Well, it's not trying to be a database
00:03:20of the world's information
00:03:21that gives you exactly what you want.
00:03:22It's kind of like a recommendation engine.
00:03:24And you can see this with more examples.
00:03:27So find me everyone in Singapore
00:03:29who works on AI search
00:03:30and any blog post or research paper
00:03:31they've written.
00:03:32I bet you've never typed in anything like this
00:03:34to Google
00:03:34because you know it's not going to work.
00:03:36You're going to get like links
00:03:38and documents that contain some of those words
00:03:40but not like actually a database result
00:03:43of all, you know, 412 people who match.
00:03:46And it gets pretty serious, right?
00:03:47Like I'm a citizen.
00:03:49I'm trying to be informed
00:03:49about what's going on in the world.
00:03:51I want to find the most important
00:03:52U.S. news across all media
00:03:53or, you know, U.S. news articles,
00:03:55whatever it is.
00:03:55You just don't trust Google
00:03:57to give you that, right?
00:03:57It's just going to recommend some things.
00:03:59It's almost akin to like social media
00:04:01in the sense.
00:04:01It's kind of like a recommendation engine.
00:04:04Okay.
00:04:05So that means that no one really
00:04:08has a complete understanding of anything.
00:04:11I actually, when I walk around like SF
00:04:12or wherever I'm walking around
00:04:13and I see people,
00:04:13I often think like,
00:04:15no one knows what's going on in the world.
00:04:16Everyone's like this, including myself.
00:04:18You know, you can imagine this person
00:04:20on the right on their phone
00:04:21like trying to find a new job.
00:04:22They're looking for biotech companies
00:04:24to work for.
00:04:25Are they going to get like all the possible
00:04:26biotech companies that match?
00:04:28No.
00:04:28So there's always going to be like
00:04:30this unknown of what's out there.
00:04:32Or, you know,
00:04:33maybe this other person
00:04:33with the headphones.
00:04:35Maybe they're trying to, yeah,
00:04:35they want to understand
00:04:36what's going on
00:04:37in some region of the world.
00:04:40They're just not going to have
00:04:41a deep understanding.
00:04:42They can't,
00:04:42not only can they not find the information,
00:04:43they might not be able
00:04:44to trust the information.
00:04:45So we basically are in a world
00:04:46where we live without this,
00:04:48like, key information infrastructure
00:04:49that is so critical.
00:04:50And I think this is extremely important.
00:04:52So important that
00:04:53if we don't fix this problem,
00:04:54I believe we get a world
00:04:56that looks like this,
00:04:57a dystopia.
00:04:58I'm not kidding.
00:04:59This is AI-generated version
00:05:00of San Francisco in 2035.
00:05:03And it's basically a world
00:05:04where no one,
00:05:05no person really understands
00:05:06what's going on.
00:05:07If people don't understand
00:05:08what's going on,
00:05:08then as we have this crazy AI technology
00:05:10that we all are talking about today
00:05:12that's coming,
00:05:13and the world is getting
00:05:13way more powerful,
00:05:14there's going to be conflict,
00:05:15all these things.
00:05:15If people don't know
00:05:16what's going on in the world,
00:05:17this is very bad.
00:05:18Like, we will be manipulated,
00:05:19we will make really bad decisions
00:05:20as individuals and as a society.
00:05:22And I think that if we could fix this problem,
00:05:25it would be way better.
00:05:26And when I think about
00:05:27all the possible problems
00:05:28that are really important
00:05:29and neglected,
00:05:30to me,
00:05:31like, solving information
00:05:32is the most important
00:05:33and neglected problem.
00:05:34Okay.
00:05:34So that's the Y of Exa.
00:05:36Quick story of how we got here.
00:05:38So I've been thinking about this problem
00:05:40for a very long time,
00:05:41way before even 2021
00:05:42when we started,
00:05:43even since high school.
00:05:44But I think what was really cool
00:05:45is that in 2021,
00:05:48it suddenly became possible
00:05:49in our eyes
00:05:50to build a new type of search engine
00:05:51because Transformers
00:05:52had gotten really good.
00:05:53So this is a time
00:05:53when, like, GB3 had recently come out.
00:05:55GB3 was, like, magical.
00:05:56You type in a paragraph of text
00:05:58and it fully understands you.
00:05:59At the same time,
00:06:00you know,
00:06:00as we saw with Google,
00:06:01it's like,
00:06:02it doesn't fully understand you.
00:06:03And so what if you could combine
00:06:04the power of GB3
00:06:06with a search engine
00:06:07and maybe you could have
00:06:08perfect search
00:06:09over the world's information?
00:06:10And actually,
00:06:10the thought experiment
00:06:11that always drove me
00:06:12was, like,
00:06:12you know,
00:06:13if we take a query,
00:06:14a complex query,
00:06:14and a document,
00:06:15and we run GB3 over it
00:06:16and we say,
00:06:16does this match?
00:06:17It'll do a really good job
00:06:18of saying does it match.
00:06:19Now, do that
00:06:20over a trillion documents
00:06:21for every search
00:06:22and you get a perfect search engine
00:06:23or near perfect.
00:06:24The problem is
00:06:24that would cost, like,
00:06:25$10 million per search.
00:06:26So it really becomes
00:06:27an interesting optimization problem.
00:06:28How do you, like,
00:06:29billion X or trillion X
00:06:30reduce the cost of that?
00:06:31So that's kind of, like,
00:06:32the ideas that started Exa.
00:06:34It's actually the first day
00:06:35of Exa, 2021.
00:06:36It's actually, by the way,
00:06:36our five-year anniversary
00:06:37as of yesterday.
00:06:40So, yeah,
00:06:40it's been a crazy time.
00:06:43I wish I took a better selfie here,
00:06:44but this was the first day.
00:06:47And Exa was basically built
00:06:49on the idea that, look,
00:06:51like, traditional search engines,
00:06:53they use keywords.
00:06:53Keywords are very efficient
00:06:55and they can handle simple queries,
00:06:58but if you want to handle
00:06:59more complex queries,
00:06:59you just need to use neural networks.
00:07:02And particularly embeddings
00:07:03are a way of, like, encoding.
00:07:05You can't run, like I said,
00:07:07like a neural network
00:07:07over every document
00:07:08for every query,
00:07:09but you can pre-process
00:07:10every document
00:07:11into some sort of structure
00:07:12like an embedding,
00:07:13and then you could use,
00:07:15and then that captures
00:07:15a lot of the intelligence
00:07:16of a neural network,
00:07:17and then you could use
00:07:17those embeddings.
00:07:18Of course, embeddings
00:07:18have their own problems,
00:07:20and often you want to combine
00:07:21embeddings and keywords,
00:07:22but certainly embeddings
00:07:23are a big part of the picture,
00:07:25and that's how you can handle
00:07:26shirts without stripes.
00:07:26We can handle this kind of clear.
00:07:28Another way of saying it
00:07:29is just stack more layers.
00:07:30Very bitter lesson-pilled.
00:07:32We were very early on
00:07:33to being bitter lesson-pilled.
00:07:34I don't know if you know this meme.
00:07:35If you don't,
00:07:36it probably looks really weird.
00:07:37Okay.
00:07:39But anyway,
00:07:39so we were very bitter lesson-pilled,
00:07:40so we did some crazy things, right?
00:07:41We raised a couple million dollars.
00:07:43We spent half of it
00:07:43on a GPU cluster.
00:07:44That was crazy at the time.
00:07:46We did a huge amount
00:07:47of research for really years,
00:07:48just heads down.
00:07:50And we did a lot of...
00:07:51We were very new to the search,
00:07:52to be honest.
00:07:53Like, we were just
00:07:53really obsessed with the problem,
00:07:54and so we invented
00:07:55a lot of new stuff
00:07:55that I still haven't seen
00:07:56even today.
00:07:59And so, just, yeah,
00:08:01history of XSO.
00:08:01So, 2022,
00:08:02so this is like a year
00:08:03and a half after starting.
00:08:04We were called
00:08:05Metaphor at the time.
00:08:06Some of you might know it.
00:08:07We launched our first
00:08:08search engine to the world,
00:08:10and it was pretty exciting.
00:08:12Like, it was a new way
00:08:13of doing search.
00:08:13A lot of people
00:08:14were really excited about it.
00:08:15The next big thing
00:08:16that happened two weeks later
00:08:17was ChetPT came out,
00:08:18and that really changed the world.
00:08:21And, like,
00:08:21this is what San Francisco
00:08:22looked like at the time,
00:08:23if you remember.
00:08:24And this is the XO team
00:08:25at the time.
00:08:26Thank you.
00:08:27All right.
00:08:30But everything was saved
00:08:31when we got this message
00:08:34on Twitter
00:08:35from this person,
00:08:36whatever,
00:08:37who wanted an API access
00:08:38to our search engine.
00:08:39And that was really weird
00:08:40because we were never thinking
00:08:41that, oh,
00:08:41this was going to be an API.
00:08:42We were just trying
00:08:43to build a better search engine
00:08:43on Google.
00:08:44We'll figure out
00:08:44how to make money later.
00:08:45Then someone asked us
00:08:46for an API.
00:08:46We were like,
00:08:47no, we don't have an API.
00:08:48Sorry.
00:08:48But then, like,
00:08:49we started getting
00:08:49more requests
00:08:51for API access,
00:08:52including, you know,
00:08:53my roommate
00:08:53who lived downstairs.
00:08:55And then we very quickly realized,
00:08:56wait a sec,
00:08:58like,
00:08:59there's a business model.
00:09:00Like, okay,
00:09:00what we realized
00:09:01was AIs need search
00:09:02because, like,
00:09:02the argument is basically,
00:09:04look,
00:09:04even GB5,
00:09:05gigantic model,
00:09:06it's tiny
00:09:07in comparison
00:09:08to the Internet, right?
00:09:08So, like,
00:09:09these systems
00:09:09always need to search.
00:09:11You're never going
00:09:11to have GB6,
00:09:12GB7.
00:09:13It's not going to be able
00:09:13to, like,
00:09:13just know everything
00:09:14about the world.
00:09:14It needs to be connected
00:09:15to a retrieval engine.
00:09:16And that was
00:09:17a really interesting insight
00:09:18because these things
00:09:18now need a search API, right?
00:09:20And so we pretty quickly realized,
00:09:22okay, wait,
00:09:23AIs are going to search the web.
00:09:24In fact,
00:09:25they're going to search the web
00:09:25way more than humans.
00:09:28And they're going to search
00:09:29in very different ways.
00:09:31So this is an example
00:09:33of what humans,
00:09:34this looks like
00:09:35when humans search, right?
00:09:35They search symbol queries.
00:09:36This is what Google was made for,
00:09:38like SpaceX News.
00:09:39It's good at that.
00:09:40But an AI system
00:09:41is very different, right?
00:09:42It kind of looks like
00:09:42this information guzzler creature
00:09:45that's, like, insane.
00:09:46And, like,
00:09:47it would be crazy
00:09:47if the same search engine
00:09:48that was optimal for humans
00:09:49was also optimal
00:09:51for these AI systems.
00:09:53So anyway,
00:09:53you know,
00:09:54we realized,
00:09:55okay,
00:09:55if we build a search API
00:09:57for these AI agents,
00:09:59or it wasn't called
00:10:00AI agents at the time,
00:10:00it was just AIs for LLMs,
00:10:02then we could make money
00:10:04from that.
00:10:04That's a nice business model,
00:10:06and we think
00:10:06it's going to grow really fast.
00:10:07And also,
00:10:08the beautiful thing is
00:10:09it matches our initial,
00:10:10our mission,
00:10:11which is perfect search, right?
00:10:12Like,
00:10:12AI systems really want
00:10:14perfect search.
00:10:14They don't want SEO.
00:10:16They don't want ads.
00:10:17They just want
00:10:18almost like a database
00:10:19of the world's information,
00:10:20which is what we were
00:10:20always trying to build.
00:10:22So we built the first
00:10:22search for LLMs,
00:10:24and, yeah,
00:10:24actually, like,
00:10:24in 2023,
00:10:25we said,
00:10:25soon AIs will search
00:10:26more than humans.
00:10:28You know,
00:10:28three years later,
00:10:29it's now happening.
00:10:31Okay.
00:10:32Yeah,
00:10:32and so then,
00:10:33the next couple years,
00:10:33we built a lot
00:10:34of really crazy stuff.
00:10:36It's way more complex
00:10:37than just embedding search.
00:10:38It combines all sorts
00:10:39of systems,
00:10:39some of which are included here.
00:10:41And now,
00:10:42you know,
00:10:42we're a much bigger team.
00:10:44And that's how we got here.
00:10:45Okay, cool.
00:10:46So just a quick,
00:10:47like,
00:10:48what can you do with Exa?
00:10:49And then I'll talk about
00:10:50where we're going,
00:10:50how we're going to get
00:10:51the perfect search.
00:10:52So the present.
00:10:53So right now,
00:10:54yeah,
00:10:54we're the highest quality
00:10:55information for AI agents,
00:10:57and you could do
00:10:58all sorts of things.
00:10:59So for example,
00:11:00a lot of people
00:11:01like really complex queries.
00:11:03So you want to find
00:11:03every startup funded
00:11:04by YC working on AI,
00:11:06give me their batch
00:11:06and status.
00:11:07You could do that
00:11:07with Exa now.
00:11:08And you could make this
00:11:09like arbitrarily complex.
00:11:10Like,
00:11:10a lot of people
00:11:12aren't aware of this,
00:11:12but you could just use Exa
00:11:13and just find really
00:11:14like any list
00:11:15of companies,
00:11:16people,
00:11:18like blog posts,
00:11:20news articles
00:11:20that you want.
00:11:21It will take some time,
00:11:23so it'll take,
00:11:23you know,
00:11:24maybe not seconds,
00:11:24it might take minutes,
00:11:25but you'll get
00:11:25the information you want.
00:11:27At the same time,
00:11:28we also have
00:11:28the fastest search API
00:11:29in the world.
00:11:30So we have a 200-millisecond
00:11:32search endpoint,
00:11:33and that's what it feels like.
00:11:34So it's super fast.
00:11:35It's way too fast for humans,
00:11:36right?
00:11:36But we're not serving humans.
00:11:37We're serving AI systems.
00:11:38And so, like, for example,
00:11:39we serve some voice agents,
00:11:41and if you're a voice agent
00:11:42and you talk to the voice agent
00:11:43and it wants to do a search
00:11:44underneath the hood,
00:11:46you know,
00:11:46every millisecond counts.
00:11:47You want it to do a search
00:11:48really fast
00:11:48so that it could go,
00:11:49you know,
00:11:50process it with an LLM
00:11:50and then output the best audio
00:11:52back to the customer.
00:11:55We also have cool things
00:11:56like super-efficient token extraction.
00:11:59So everyone's talking
00:12:00about the compute crunch
00:12:01and how everyone's spending
00:12:02way too much on LLMs.
00:12:04Well, actually,
00:12:04Exa could help there
00:12:05because, you know,
00:12:07when the LLM makes a query,
00:12:08it wants to get
00:12:09just the information it needs,
00:12:11just the tokens it needs.
00:12:12And so we take the documents,
00:12:14we give you 10 documents,
00:12:15and then we'll give you
00:12:15only the most important,
00:12:17like, 100 tokens
00:12:17from those documents,
00:12:18and that will save you
00:12:19a lot of downstream LLM costs.
00:12:22We also, you know,
00:12:23some people,
00:12:23they don't necessarily want,
00:12:24you know,
00:12:25snippets from each document.
00:12:26They actually want
00:12:27structured output.
00:12:27So, hey,
00:12:28you know,
00:12:28let's say you're building
00:12:29a recruiting AI agent
00:12:30and you want to find,
00:12:31you know,
00:12:31all the engineers
00:12:32who recently left
00:12:33their big lab job,
00:12:35give me, like,
00:12:36you know,
00:12:36the most cited paper
00:12:40that they've written,
00:12:41give me the college
00:12:41they went to,
00:12:42and the year they graduated,
00:12:43and we'll just give you
00:12:43that as structured output.
00:12:44It makes it really easy.
00:12:46And, yeah,
00:12:47I think one takeaway here
00:12:48is we're not building
00:12:50one search engine.
00:12:51Like, perfect search
00:12:51is not one thing.
00:12:52It's actually,
00:12:53we have 5,000 search engines
00:12:55for each of our 5,000 customers,
00:12:56right?
00:12:57We want to build our system
00:12:58so it's super flexible
00:12:59because we want every business
00:13:00to be super optimized.
00:13:01And that's, I think,
00:13:02a beautiful thing
00:13:02because, like,
00:13:03we don't want to declare
00:13:04what is the perfect search.
00:13:05We want you to almost,
00:13:06you know,
00:13:06tell us what exactly
00:13:07do you want.
00:13:08Do you want super fast?
00:13:09Do you want, you know,
00:13:10the highest possible quality
00:13:11even though it takes minutes?
00:13:12Do you want to search
00:13:13only over these thousand domains?
00:13:14Do you want to never
00:13:15search over those thousand domains?
00:13:16Do you want to search
00:13:16within this time window?
00:13:18Do you want to never
00:13:19get product pages?
00:13:20Some people ask us for that.
00:13:21So there's all sorts of things
00:13:22that you could do with Exa.
00:13:23It's very flexible,
00:13:24customizable.
00:13:26And, yeah,
00:13:27I mean,
00:13:27our search quality
00:13:28is really good
00:13:29for these AI agents
00:13:30because we've spent years,
00:13:31you know,
00:13:31doing research
00:13:32into how do you build
00:13:33a new type of search engine
00:13:34for agents.
00:13:35It's even better than Google,
00:13:36which was built for humans,
00:13:37which makes sense.
00:13:40Okay,
00:13:41a cool thing
00:13:41that we released recently
00:13:42was Exa Connect.
00:13:44So, you know,
00:13:45agents,
00:13:45they don't really care
00:13:46whether the information
00:13:47is from the public web
00:13:48or from private data sources.
00:13:49They just want the truth, right?
00:13:51And so it's always been obvious
00:13:52to us that, you know,
00:13:53we want to assemble
00:13:54all the world's information.
00:13:55It's perfect search
00:13:56over all the world's information,
00:13:57not just the web.
00:13:57And so now we have a system
00:13:58where data providers
00:14:00can actually partner with Exa
00:14:02so that developers
00:14:03can then get the data
00:14:04from those providers.
00:14:05So we're basically creating
00:14:06like a new market,
00:14:08a new economy for agents
00:14:10where, you know,
00:14:11if you have high valuable data,
00:14:13you could get paid
00:14:13from all the developers
00:14:15who want their agents
00:14:16to access that data.
00:14:17So we're creating
00:14:18this beautiful marketplace.
00:14:19I think it's really cool.
00:14:19It's like a free market.
00:14:20Like the data providers
00:14:22can decide how much
00:14:23they think their content is worth
00:14:25and then developers
00:14:25can decide what data they want.
00:14:29And then you could do
00:14:30cool things like this.
00:14:32Research AI in for companies,
00:14:33monthly website visitors,
00:14:34similar web.
00:14:35Oh, I guess it went too fast,
00:14:36but you could get like
00:14:36basically this is combining
00:14:37information from the public web
00:14:39and also information
00:14:40from similar web
00:14:41which is not publicly available.
00:14:44So you could do queries
00:14:44like that right now.
00:14:46Okay.
00:14:47So that's where Exa is right now,
00:14:48but the future
00:14:49has always been super exciting to me
00:14:51and the goal has always been
00:14:52perfect information
00:14:53and now we know
00:14:54it's for AI agents.
00:14:57So instead of a world like this,
00:14:59no, bad,
00:15:00we want search to kind of feel like this.
00:15:03It's actually really hard
00:15:04to describe what perfect information
00:15:05feels like.
00:15:06Best way you could say it
00:15:08is like literally
00:15:08any information query you have,
00:15:10it just works.
00:15:11No matter how complex that is.
00:15:14Another way you could think about it
00:15:15is like it's as if you did
00:15:16a year of research in a second.
00:15:19So imagine no matter
00:15:20what you're looking for,
00:15:20whether it's people
00:15:21or companies or news,
00:15:22imagine you spent a whole year,
00:15:23you spent all of 2026
00:15:24just doing research for it,
00:15:26you get that in a second.
00:15:27That should give you a sense
00:15:28of what perfect information
00:15:29feels like
00:15:30and you do that for every search,
00:15:31all the crazy number of searches
00:15:33that AI agents are going to make.
00:15:36And so yeah,
00:15:36we want to move really fast
00:15:37at Excel.
00:15:39We're moving extremely,
00:15:40basically every,
00:15:41basically a quarter
00:15:42or two quarters now
00:15:43we have as much progress
00:15:44as we did the past five years
00:15:45and it keeps being like that,
00:15:46it's exponential growth.
00:15:47And so yeah,
00:15:47our ambitions for 2027
00:15:48are pretty crazy.
00:15:50We want the world
00:15:51to be like this
00:15:51where everyone walks around
00:15:52with deep understanding
00:15:53of what's going on.
00:15:54I think it's particularly important
00:15:55because things like
00:15:56the 2028 presidential election
00:15:57are coming soon
00:15:58and I would love
00:15:59for the entire world
00:16:01to have access
00:16:02to near perfect information
00:16:04so that everyone
00:16:04is very informed
00:16:05going into that election.
00:16:06You kind of can feel
00:16:07the gravity
00:16:07of what we're doing here
00:16:10of perfect information
00:16:11and I encourage others
00:16:12to try to do it too.
00:16:13It's very important
00:16:14for the world.
00:16:14It's like key information,
00:16:15it's key infrastructure.
00:16:17And yeah,
00:16:17you have to like,
00:16:18that slide I showed
00:16:19at the beginning,
00:16:20it's not the whole picture,
00:16:21right?
00:16:21If you play it out,
00:16:22we're talking a thousand times
00:16:23more searches
00:16:23from AI systems
00:16:24than humans.
00:16:25You can't even capture
00:16:26that on a graph.
00:16:26That's like 20 times.
00:16:28A thousand times
00:16:29it'll be all the way up there
00:16:30on top of a building
00:16:31or something, right?
00:16:31So it's crazy
00:16:32what's coming
00:16:33and it's really happening.
00:16:35Basically,
00:16:36humans on average
00:16:37search a few times
00:16:37a day on Google
00:16:38but when everyone
00:16:39has AI systems
00:16:40and every software product
00:16:41you use has AIs in it,
00:16:43every interaction
00:16:43you're doing
00:16:44is going to be grounding
00:16:45itself in search.
00:16:47So it's going to be
00:16:47a huge number of searches
00:16:48and if each of those searches
00:16:49are as true as possible,
00:16:51as near perfect,
00:16:52then the world
00:16:54looks like this
00:16:54in 2035.
00:16:57And yeah,
00:16:58I do think that if,
00:16:59I do think we're basically
00:17:00our future is limited
00:17:02by ourselves.
00:17:03Like we're basically
00:17:03getting into a world
00:17:04where our technology
00:17:05is so good,
00:17:05it's really just a matter
00:17:06of like can we coordinate
00:17:07and just decide together
00:17:09that like on sensible things,
00:17:11right?
00:17:11Like if you look
00:17:11at San Francisco,
00:17:12there's amazing things
00:17:13happening here
00:17:13and there's really stupid
00:17:14things happening here
00:17:14at the same time.
00:17:16Like that's just
00:17:16coordination problems
00:17:17and coordination comes
00:17:18from the information
00:17:18we consume.
00:17:20So that's what
00:17:21we're working on.
00:17:21You all have a role
00:17:22to play
00:17:22in also getting
00:17:23to this world.
00:17:24So thank you all
00:17:25for working on
00:17:27whatever passion
00:17:27you're working on
00:17:28and thanks for listening
00:17:28to me.
00:17:29All right,
00:17:29thank you.
00:17:29Thank you.
Community Posts
No posts yet. Be the first to write about this video!
Write about this video