Latest / What's Your Problem? / Google's Journey to the Edge of Search

Transcript

Google's Journey to the Edge of Search · What's Your Problem?

• 0:15
  Speaker 1: Pushkin. It feels like searching the web is a problem
• 0:23
  Speaker 1: that's been solved. You know, it's ridiculously easy for me
• 0:26
  Speaker 1: to say, find out when Alexander Hamilton was shot eighteen
• 0:30
  Speaker 1: oh four, or whether they are making Sing three. Not yet,
• 0:34
  Speaker 1: but Matthew McConaughey has expressed interest and yet. And maybe
• 0:38
  Speaker 1: this is not surprising. The people who spend their lives
• 0:42
  Speaker 1: working on search do not think search is solved. This
• 0:46
  Speaker 1: is partly because the people at the frontier of search
• 0:50
  Speaker 1: don't just want to search the web. They want to
• 0:53
  Speaker 1: answer every question that might cross your mind, even questions
• 0:57
  Speaker 1: you can't put into words. I'm Jacob Goldstein, and this
• 1:06
  Speaker 1: is What's your problem? The show where on entrepreneurs and
• 1:09
  Speaker 1: engineers talk about how they're going to change the world
• 1:12
  Speaker 1: once they solve a few problems. My guest today is
• 1:15
  Speaker 1: Kathy Edwards, vice president and GM of Search at Google.
• 1:20
  Speaker 1: Cathy's problem is this, how do you teach computers to
• 1:23
  Speaker 1: tell people what they want to know, even if they
• 1:26
  Speaker 1: don't know how to ask. Later in the conversation we
• 1:30
  Speaker 1: get to the frontier of what Kathy and Google are
• 1:33
  Speaker 1: working on now, but we started with the problem they
• 1:36
  Speaker 1: have largely solved in the six years Kathy has been
• 1:39
  Speaker 1: at Google, the jump from search results based on keywords
• 1:43
  Speaker 1: to search results based on natural language, the way people
• 1:46
  Speaker 1: talk in everyday life. So one of the problems that
• 1:51
  Speaker 1: we were working on around six years ago is this
• 1:54
  Speaker 1: problem of natural language queries. So, if you're old enough
• 1:58
  Speaker 1: to remember the early days of search on the Internet,
• 2:02
  Speaker 1: there was this idea of keywordies, right, that you had
• 2:05
  Speaker 1: to sort of take this idea you had in your
• 2:09
  Speaker 1: mind of what you what you needed to know and
• 2:11
  Speaker 1: figure out what were the exact right keywords to enter
• 2:15
  Speaker 1: into the search engine to get your results back right.
• 2:17
  Speaker 1: I mean an example as an example, very early in
• 2:20
  Speaker 1: the you know, I remember being taught how to query
• 2:23
  Speaker 1: back in you know, nineteen ninety nine and being told
• 2:28
  Speaker 1: never used the word and never used the word because
• 2:32
  Speaker 1: the word and or the word there is in almost
• 2:35
  Speaker 1: every document on the Internet. And the way it worked
• 2:38
  Speaker 1: back then is you did this word matching, right, and
• 2:41
  Speaker 1: so if you had a word that was in your
• 2:44
  Speaker 1: query and there was that same word in the document,
• 2:47
  Speaker 1: then that document would be returned and potentially scored right.
• 2:51
  Speaker 1: And that was very helpful if it was a word
• 2:54
  Speaker 1: like genetics, right, which was highly specific and wasn't in
• 2:58
  Speaker 1: a heap of documents on the internet. But the word
• 3:00
  Speaker 1: and not very specific. And you know, in the very
• 3:04
  Speaker 1: early days the Internet, these words weren't even weighted particularly right,
• 3:08
  Speaker 1: The word and count for as much as the word genetics,
• 3:11
  Speaker 1: and so a document might have a ton of the
• 3:14
  Speaker 1: uses of the word and and one use of the
• 3:17
  Speaker 1: word genetics, and it would score really highly, even though
• 3:20
  Speaker 1: it wasn't particularly genetics. Folk. Now, by the time you
• 3:22
  Speaker 1: get to Google, that part is solved, right Google buy
• 3:25
  Speaker 1: that part is years ago? Is waiting genetics more more
• 3:29
  Speaker 1: heavily than it's waiting the But but what what part
• 3:33
  Speaker 1: six years ago was not solved? That's solved now or
• 3:36
  Speaker 1: solved ish now. But we were still seeing people do
• 3:40
  Speaker 1: these very keyword oriented queries. So they weren't saying things
• 3:44
  Speaker 1: like what wine pairs best with chicken? Or if they were,
• 3:48
  Speaker 1: they were doing those queries and getting not the best results,
• 3:53
  Speaker 1: because not only is there a question of word matching
• 3:56
  Speaker 1: and how much each word counts for, there's also the
• 3:59
  Speaker 1: question of like does the word what appear at all? Right?
• 4:03
  Speaker 1: Like are the answers to that question actually just documents
• 4:07
  Speaker 1: that talk about the best wine to pair with chicken
• 4:10
  Speaker 1: is you know, chardonnay, right, and not so much talking.
• 4:16
  Speaker 1: You know, they didn't include the question, and so we
• 4:18
  Speaker 1: sort of saw these like SEO documents that would spring
• 4:23
  Speaker 1: up that would have the questions kind of baked in
• 4:26
  Speaker 1: and an attempt to match. But those documents weren't necessarily
• 4:29
  Speaker 1: the best answers. And so this is when we started
• 4:33
  Speaker 1: to go just that next level deeper in our language
• 4:36
  Speaker 1: understanding with these AI models, their language models that really
• 4:42
  Speaker 1: can start to map out in a concept space, things
• 4:46
  Speaker 1: like this sort of translation between how you might ask
• 4:49
  Speaker 1: a query and then what that might look like in
• 4:51
  Speaker 1: the document. So, to take the example you gave of
• 4:54
  Speaker 1: what wine pairs best with chicken, even as late as
• 4:57
  Speaker 1: six years ago when you got to Google, you're saying
• 5:00
  Speaker 1: Google wasn't great at delivering the best results to a
• 5:05
  Speaker 1: query like that because it was written as speech, not
• 5:09
  Speaker 1: written as a series of keywords. So six years ago,
• 5:11
  Speaker 1: I would have been better off typing chicken wine pairing.
• 5:17
  Speaker 1: I would have got better results if I did that,
• 5:19
  Speaker 1: you're saying, because that's kind of the way. That's the
• 5:21
  Speaker 1: way Google had mapped the web. It was like a
• 5:23
  Speaker 1: series of important words and what sites are reliable and
• 5:26
  Speaker 1: they it just the technology wasn't there to actually try
• 5:29
  Speaker 1: and understand the way people ask questions in real life. Absolutely,
• 5:35
  Speaker 1: And it was this idea of bringing AI into search
• 5:41
  Speaker 1: and having these like large scale language models. That first
• 5:44
  Speaker 1: one was called Bert. We now use one called mom,
• 5:48
  Speaker 1: which is get to mom. But let's talk about Let's
• 5:50
  Speaker 1: try talking about Bert for a second. So how do
• 5:53
  Speaker 1: you get from search results that are fundamentally keyword based
• 5:59
  Speaker 1: to search results that are fundamentally you know, answering questions
• 6:03
  Speaker 1: that are posed in a more natural way, like how
• 6:06
  Speaker 1: do you make that leap? So the fundamental insight is
• 6:11
  Speaker 1: you go from looking at these words as tokens that
• 6:14
  Speaker 1: get matched against each other to suddenly you look at
• 6:18
  Speaker 1: all the words in all the documents on the Internet
• 6:22
  Speaker 1: and you create what's code an embedding space, which is
• 6:25
  Speaker 1: essentially you can think of it as a map of
• 6:27
  Speaker 1: the concepts that these documents know about. And suddenly, by
• 6:32
  Speaker 1: being able to say, okay, you can take a query,
• 6:34
  Speaker 1: map that into this concept embedding space. You'd take these documents,
• 6:39
  Speaker 1: map that into the content embedding space. You can start
• 6:42
  Speaker 1: to actually match together not these words, but what people
• 6:47
  Speaker 1: actually mean what they actually mean when they ask these questions,
• 6:50
  Speaker 1: and what they actually mean when they write these web
• 6:53
  Speaker 1: pages on the Internet. That seems I mean, A, it
• 6:56
  Speaker 1: seems super hard, right, and B As I'm parsing that,
• 7:00
  Speaker 1: I'm tempted to use a lot of anthropomorphic language, right,
• 7:03
  Speaker 1: I'm tempted to say, like, you have to go from
• 7:06
  Speaker 1: the computer just sort of having a list of words
• 7:09
  Speaker 1: and kind of weights around those words to a computer
• 7:12
  Speaker 1: understanding what people mean. Like, am I right to say that?
• 7:16
  Speaker 1: Or is that just my like layperson intuition getting in
• 7:19
  Speaker 1: the way of what's going on? I Mean, the first
• 7:22
  Speaker 1: thing I'll say is I think we're very far away
• 7:25
  Speaker 1: from the computer having any sort of sentience and truly understanding.
• 7:30
  Speaker 1: But I think it is true. It is fair to
• 7:32
  Speaker 1: say that there is a level of deeper understanding that
• 7:37
  Speaker 1: you're not just looking at these words as as you know,
• 7:41
  Speaker 1: bits in a computer, but you're actually starting to model
• 7:44
  Speaker 1: in a way that a human might, a brain might
• 7:46
  Speaker 1: model what the concepts are. And I do think that's
• 7:50
  Speaker 1: a first step of getting closer to this sort of
• 7:53
  Speaker 1: natural human understanding. So is there a way to talk
• 7:59
  Speaker 1: about how that works? It's it's pattern matching effectively right,
• 8:04
  Speaker 1: And it just so happens that if you magnify pattern
• 8:07
  Speaker 1: matching on a very large scale, that can be a
• 8:10
  Speaker 1: pretty compelling understanding. And so that's the sort of big idea,
• 8:15
  Speaker 1: the theory of how it works. I'm sure in actually
• 8:19
  Speaker 1: building the thing in building Burt, which was this big
• 8:23
  Speaker 1: model that did work, it wasn't that easy, right, I mean,
• 8:28
  Speaker 1: is there a is there a story version of how
• 8:33
  Speaker 1: you built it? So I think there were two hard
• 8:36
  Speaker 1: points along the journey. The first hard point was just
• 8:40
  Speaker 1: these models were being built at a scale that was
• 8:44
  Speaker 1: unprecedented the amount of information. You know, traditional neural networks
• 8:48
  Speaker 1: would run on thousands, maybe millions of training examples. Suddenly
• 8:53
  Speaker 1: you're trying to model all the words on the Internet
• 8:58
  Speaker 1: and this scale. Firstly, this scale is what gets you
• 9:02
  Speaker 1: the amount of training to actually get the concepts model
• 9:04
  Speaker 1: to be compelling. But frankly, the computers just couldn't process.
• 9:09
  Speaker 1: So you're you're building this model and saying, okay, now
• 9:12
  Speaker 1: to learn what you need to learn, read literally every
• 9:15
  Speaker 1: word on the Internet, is that right, yes, and not
• 9:18
  Speaker 1: read at once, because every layer of the neuronet needs
• 9:22
  Speaker 1: to read it and reprocess it. Right. So you're reading
• 9:25
  Speaker 1: every word, you know, a massive number of times. And
• 9:29
  Speaker 1: at the time we didn't really have the compute power.
• 9:32
  Speaker 1: You just needed more more computers, essentially more and more chips,
• 9:37
  Speaker 1: more more engines to just process and process and process.
• 9:41
  Speaker 1: So our research team had developed these these chips that
• 9:47
  Speaker 1: these processes that were really optimized for doing a sort
• 9:51
  Speaker 1: of deep learning work. And it was that these chips
• 9:55
  Speaker 1: and the way that we could sort of put all
• 9:56
  Speaker 1: the chips together at a work in concert to solve
• 9:59
  Speaker 1: this problem that really unlocked the amount of processing power
• 10:03
  Speaker 1: and needed to even build these models in the festival.
• 10:05
  Speaker 1: So the binding constraint wasn't like the theory or the
• 10:09
  Speaker 1: ideas of it, like you knew how to do it,
• 10:10
  Speaker 1: you just didn't have enough enough horsepower to actually make
• 10:14
  Speaker 1: it happen. Well, we knew that we could do it,
• 10:18
  Speaker 1: we didn't know offer to be any good, right, it
• 10:20
  Speaker 1: wasn't it, and you couldn't even try, right, yeah? Right?
• 10:25
  Speaker 1: And so then we tried it and we found out, actually,
• 10:28
  Speaker 1: this thing is pretty compelling. It can understand things that
• 10:32
  Speaker 1: our models previously have never understood. You know. But I
• 10:36
  Speaker 1: will say the second and this gets to the second
• 10:39
  Speaker 1: hard part. We once we had these large scale language models,
• 10:44
  Speaker 1: we didn't quite know how to put them into search ranking.
• 10:49
  Speaker 1: This was not something that had been done before. So
• 10:52
  Speaker 1: we have in search this incredibly rigorous methodology for testing
• 10:57
  Speaker 1: any given change to our algorithm, and it's it's based
• 11:01
  Speaker 1: in statistics, and it's statistically samples queries, and we look
• 11:05
  Speaker 1: at the before and they after, and there's a scoring
• 11:08
  Speaker 1: system to say is it better or not? And I
• 11:11
  Speaker 1: remember looking at the early experiments from this burst integration
• 11:17
  Speaker 1: into our search engine, and the queries that it was
• 11:22
  Speaker 1: impacting were just queries that, honestly before we would have said,
• 11:28
  Speaker 1: we don't know how we can solve this query. And
• 11:30
  Speaker 1: suddenly the model was just able to figure out these
• 11:35
  Speaker 1: sort of unspoken concepts that just our previous technology just
• 11:41
  Speaker 1: would not have even been able to come close to.
• 11:44
  Speaker 1: Like give me an example, like what kind of thing?
• 11:47
  Speaker 1: So here's a really great example. This is directly from
• 11:51
  Speaker 1: the one of the very first bit evaluations that we
• 11:56
  Speaker 1: did internally, and the query is can you get medicine
• 12:01
  Speaker 1: for someone? Pharmacy? Right? And so what's interesting about this
• 12:05
  Speaker 1: question is the users looking for something very specific. They're
• 12:09
  Speaker 1: looking like maybe my partner is sick. Can I go
• 12:13
  Speaker 1: and pick up their prescription at the pharmacy for them?
• 12:16
  Speaker 1: Or do they have to go and get it? Right?
• 12:18
  Speaker 1: It's also a goodly jankie where it's half in natural
• 12:22
  Speaker 1: language can you get medicine for someone? And half in
• 12:25
  Speaker 1: like keyword ease, they're just typing pharmacy at the end, right,
• 12:28
  Speaker 1: it's a weird exactly yeah. And so previously we didn't
• 12:35
  Speaker 1: know how to pause out this intent right, this idea.
• 12:38
  Speaker 1: You know, we could tell that it was about getting
• 12:42
  Speaker 1: a prescription from a pharmacy, but this notion of force
• 12:46
  Speaker 1: someone was a concept that was just slightly too complex. Oh,
• 12:53
  Speaker 1: I didn't even understand it until now. What they mean
• 12:55
  Speaker 1: is can I pick up someone else's prescription? That's what
• 12:58
  Speaker 1: they're actually asking, But it's very it's poorly worded, frankly,
• 13:02
  Speaker 1: and they'refore hard to figure out exactly right. And so previously,
• 13:06
  Speaker 1: before Bert, we would turn these wonderful web pages saying
• 13:11
  Speaker 1: this is how you get a prescription filled, which you
• 13:15
  Speaker 1: can imagine if you're this user doing this query, you're like, yeah,
• 13:19
  Speaker 1: I already know how to get a prescription that filled.
• 13:21
  Speaker 1: Thanks for me. What I need is filled for somebody
• 13:24
  Speaker 1: else exactly, and with Bert we were able to understand
• 13:29
  Speaker 1: pick up this idea of the force someone and put
• 13:32
  Speaker 1: the appropriate weight on it, that that was the sort
• 13:35
  Speaker 1: of you know, discriminating thing in the query, that that
• 13:39
  Speaker 1: was the key thing that the query turned on. And
• 13:42
  Speaker 1: then we were able to show this this web page
• 13:46
  Speaker 1: that talked about can I have a friend or family
• 13:48
  Speaker 1: member pick up a prescription for me? And that was
• 13:51
  Speaker 1: the sort of like aha moment where we could all
• 13:54
  Speaker 1: just sit around and be like, Wow, this is a
• 13:56
  Speaker 1: new level of understanding that we haven't got to previously.
• 14:02
  Speaker 1: So with birth, Google got to the point where it
• 14:05
  Speaker 1: was very very good at dealing with words in a deep,
• 14:09
  Speaker 1: complex way. But words make up less and less of
• 14:12
  Speaker 1: the Internet. Pictures and videos are a whole other story
• 14:16
  Speaker 1: that's coming up in a minute. Now back to the show.
• 14:26
  Speaker 1: So you have got to this place now where you
• 14:28
  Speaker 1: have you Google have made the leap from keyword based
• 14:32
  Speaker 1: searches to intention based search is what do people mean? Right?
• 14:35
  Speaker 1: Which is this big interesting leap? And so I'm interested
• 14:39
  Speaker 1: in kind of the next leap, like what's the next
• 14:42
  Speaker 1: big hard problem you're trying to solve? What's really interesting
• 14:47
  Speaker 1: to me is this idea of how many questions you
• 14:50
  Speaker 1: don't ask because you don't even know the words, right,
• 14:54
  Speaker 1: Like this is a bit of a sad story. But
• 14:58
  Speaker 1: I have at my house this oak tree, and the
• 15:01
  Speaker 1: oak tree I think is dead, and it's very sad
• 15:04
  Speaker 1: for me because a very beautiful oak tree. And what's
• 15:08
  Speaker 1: interesting is, you know, I looked at the oak tree,
• 15:11
  Speaker 1: I'm like, wow, those leaves are kind of brown, Like,
• 15:14
  Speaker 1: that's not it doesn't seem right to me. I wonder
• 15:17
  Speaker 1: if there's something's wrong with the oak tree, right, But
• 15:20
  Speaker 1: I can't necessarily right now really articulate that to a
• 15:24
  Speaker 1: computer this fundamental question of is this oak tree dead?
• 15:27
  Speaker 1: And if not, what can I do to save it? Right?
• 15:30
  Speaker 1: So what I do is I go and type in
• 15:32
  Speaker 1: some queries, I say, you know, oak tree dead? How
• 15:36
  Speaker 1: do I know if my oak tree is dead? You know?
• 15:39
  Speaker 1: And I get back results. But those results aren't necessarily
• 15:42
  Speaker 1: taking what they're not taking in any context of this
• 15:46
  Speaker 1: particular tree and what do the leaves look like? And
• 15:49
  Speaker 1: so this idea of how can you start to ask
• 15:53
  Speaker 1: these questions using all of the information around you, using
• 15:59
  Speaker 1: your camera to actually capture this particular oak tree, using
• 16:03
  Speaker 1: your location to know, you know, what are the native
• 16:07
  Speaker 1: oaks in this area? And what's the current incidence of
• 16:12
  Speaker 1: sudden oak death syndrome, which is a thing that I
• 16:14
  Speaker 1: have recently learned exists. Okay, so I get why this
• 16:17
  Speaker 1: is a hard thing to search in a text box, right,
• 16:22
  Speaker 1: And so the thing that's interesting to me is how
• 16:24
  Speaker 1: can we facilitate asking those types of questions where it's
• 16:29
  Speaker 1: a mix of here's something that you're looking at, Here's
• 16:33
  Speaker 1: something that you're saying with your words that adds to
• 16:37
  Speaker 1: the picture. You know, here's a lemon tree that's got
• 16:40
  Speaker 1: some black spots on it? What's wrong with it? Like?
• 16:43
  Speaker 1: How can you help me understand what I should do
• 16:46
  Speaker 1: about this? You know, these sorts of questions I think
• 16:50
  Speaker 1: are right now. We have to do a tremendous amount
• 16:54
  Speaker 1: of work to try and translate these questions into text
• 16:57
  Speaker 1: that we would issue to a search engine. And yeah,
• 17:01
  Speaker 1: we use that. Yeah, yeah, normal people. Yes, we're all
• 17:05
  Speaker 1: doing it. And when you think it's you know, sometimes
• 17:08
  Speaker 1: it's very easy, right, but sometimes you're like really having
• 17:12
  Speaker 1: to work hard to come up with a query that
• 17:14
  Speaker 1: will actually get you the answers that you need. And
• 17:16
  Speaker 1: I think that's really the next frontier for us is
• 17:19
  Speaker 1: how do we on the query side help users just
• 17:23
  Speaker 1: naturally intuitively express whatever information need they have. And then
• 17:30
  Speaker 1: how do we understand the whole universe of information, not
• 17:34
  Speaker 1: just the web pages, that all the images and video
• 17:38
  Speaker 1: and audio out there, and take that next level of
• 17:43
  Speaker 1: like concept understanding to match those together so that we
• 17:47
  Speaker 1: can get users even more precise answers that really help them. Great.
• 17:53
  Speaker 1: So that's the like vast dream slash big problem. Can
• 18:00
  Speaker 1: we reduce it a little bit so we can talk
• 18:02
  Speaker 1: in sort of practical terms about what you're working on.
• 18:05
  Speaker 1: I mean, I know there's this new AI model that
• 18:09
  Speaker 1: integrates images, like you can, you know, whatever, take a
• 18:12
  Speaker 1: picture with the camera on your phone and put in text.
• 18:14
  Speaker 1: So like, well, you have this new model, and it,
• 18:17
  Speaker 1: like the old one, has this worm fuzzy acronym. Right,
• 18:20
  Speaker 1: it's called MUM, which stands for hold on, I gotta
• 18:23
  Speaker 1: look at my notes, the multitask unified model. So like,
• 18:28
  Speaker 1: tell me about MUM. So MOM is our next level
• 18:33
  Speaker 1: model that you know Bert was about language. MOM is
• 18:36
  Speaker 1: about all these different modalities of information coming together, particular
• 18:41
  Speaker 1: images and language. I mean, is that if really images
• 18:44
  Speaker 1: and language and we've got some limited applications of it
• 18:49
  Speaker 1: in search today. So for example, you can take the
• 18:54
  Speaker 1: take the photo of somebody's handbag and say you want
• 18:58
  Speaker 1: to shop it, and that will work today. And that
• 19:01
  Speaker 1: is like we were not able to do this previously,
• 19:04
  Speaker 1: and that in and of itself is a big breakthrough.
• 19:06
  Speaker 1: But there's still so much headroom, right like, this, still
• 19:09
  Speaker 1: so much ability to say, you know, you can add
• 19:13
  Speaker 1: sort of I would I would classify our current ability
• 19:18
  Speaker 1: to process words in this multimodal context as you know,
• 19:22
  Speaker 1: kind of like back in the early days days of
• 19:24
  Speaker 1: the internet, you can say near me to find where
• 19:26
  Speaker 1: you can buy it. Near me, you can say buy,
• 19:29
  Speaker 1: but you can't necessarily like ask an incredibly complicated question
• 19:34
  Speaker 1: about a picture, right like, so we're kind of back
• 19:38
  Speaker 1: to keywords in this new pictures plus words universe. Let
• 19:42
  Speaker 1: me ask a dumb question, why why can't you just
• 19:45
  Speaker 1: take all of your brilliant intent AI and copy and
• 19:51
  Speaker 1: paste it to fit with the image AI. So a
• 19:56
  Speaker 1: couple of things. The first is that anytime we develop
• 20:02
  Speaker 1: sort of this new technology, we also need to see
• 20:04
  Speaker 1: how users start using it, right And so I think
• 20:07
  Speaker 1: it's also fed say that we don't have. You know,
• 20:12
  Speaker 1: we have a ton of people using this, but we
• 20:15
  Speaker 1: haven't yet. There hasn't been time for that new technology
• 20:18
  Speaker 1: to really be accepted by the world. And then we
• 20:21
  Speaker 1: have this vast set of queries that we're doing poorly on, right.
• 20:25
  Speaker 1: So that's the other thing you should know about Google.
• 20:27
  Speaker 1: We spend a lot of time looking at the queries
• 20:29
  Speaker 1: where we're failing. That's one of the other reasons we
• 20:31
  Speaker 1: have a deep appreciation of how search is an unsolved
• 20:33
  Speaker 1: problem because we're just constantly looking at queries whether the
• 20:37
  Speaker 1: users clearly not getting what they're looking for. And I'll
• 20:41
  Speaker 1: use that as a as a siege to figure out
• 20:45
  Speaker 1: how to make things better. So do I understand you
• 20:47
  Speaker 1: that the fundamental thing you need now is just lots
• 20:50
  Speaker 1: of people to use this thing so you can see
• 20:53
  Speaker 1: the weird ways people search and the things they sort
• 20:56
  Speaker 1: of do that are hard to understand. That's certainly one
• 21:00
  Speaker 1: of the things we need. I mean, it is fundamentally
• 21:04
  Speaker 1: search works in service of our users, right, and so
• 21:07
  Speaker 1: understanding the the failures is critical to how we get better.
• 21:13
  Speaker 1: I think there are also just things that we know
• 21:15
  Speaker 1: that we need to do on the AI and the
• 21:17
  Speaker 1: model side that we'll continue working through, right, the ability
• 21:22
  Speaker 1: to really bring together more of the two step process
• 21:28
  Speaker 1: of how do you conceptually understand the words, conceptually understand
• 21:31
  Speaker 1: the image, and then bring those two things together and
• 21:35
  Speaker 1: have that be a bit deeper on both sides rather
• 21:37
  Speaker 1: than just the combination together and those sorts of things.
• 21:41
  Speaker 1: But yeah, I mean people coming in and using it
• 21:44
  Speaker 1: and then having a bad time, we'll then make it better.
• 21:48
  Speaker 1: It seems like there have been two main threads of
• 21:53
  Speaker 1: AI research. One is basically language and the other is
• 21:57
  Speaker 1: basically vision and images. I mean, it is it right
• 22:02
  Speaker 1: to think of what you're trying to do as the
• 22:05
  Speaker 1: synthesis of those two sort of main AI traditions. Yeah,
• 22:10
  Speaker 1: I think so. I think it is clearly the case
• 22:16
  Speaker 1: that just like uh, you know, with Bert, we took
• 22:20
  Speaker 1: all these words and we got down to concepts. Right,
• 22:23
  Speaker 1: It is clearly the case that human beings understand the
• 22:26
  Speaker 1: world through concepts, and they do that visually, and they
• 22:28
  Speaker 1: do that with language, and ultimately the concepts are the same. Right,
• 22:32
  Speaker 1: So being able being able to say, okay, here's here's
• 22:36
  Speaker 1: a concept, and we can attach to that what that
• 22:39
  Speaker 1: concept looks like or that visual representation of that concept
• 22:43
  Speaker 1: as much as it has one and the words surrounding
• 22:45
  Speaker 1: that concept. That's when we can really unlock this true
• 22:51
  Speaker 1: natural way of understanding the world that we think is
• 22:56
  Speaker 1: going to enable people to ask all those questions that
• 22:58
  Speaker 1: they have that they're not asking right now. Are there
• 23:03
  Speaker 1: applications that go beyond search that come to mind if
• 23:08
  Speaker 1: you figure this out? Yeah, I mean I think that
• 23:16
  Speaker 1: search has this connotation of kind of find what's out there.
• 23:22
  Speaker 1: I think there's something, you know, we're thinking about what
• 23:26
  Speaker 1: this looks like in the generative space. So for example,
• 23:31
  Speaker 1: if you're looking for you know, I bake birthday cakes,
• 23:36
  Speaker 1: and sometimes I for my kids, and sometimes what my
• 23:40
  Speaker 1: kids want, and a birthday cake just actually doesn't exist
• 23:43
  Speaker 1: on the internet right or there's like only one or two,
• 23:47
  Speaker 1: So then I have to come up with it myself.
• 23:48
  Speaker 1: And like, what if AI could help us generate as
• 23:54
  Speaker 1: sample image just based on these concepts that I could
• 23:59
  Speaker 1: then use for inspiration. I think that's a pretty interesting concept.
• 24:03
  Speaker 1: There's obviously a lot of things that we need to
• 24:05
  Speaker 1: be very thoughtful about in this space as we do it,
• 24:09
  Speaker 1: But I think this idea of extending search past the
• 24:13
  Speaker 1: notion of connecting you with the information that's out there.
• 24:15
  Speaker 1: To actually synthesizing new information for you is pretty interesting
• 24:21
  Speaker 1: and something we're talking about a lot. You know. One
• 24:25
  Speaker 1: of the things that has become clear to me talking
• 24:27
  Speaker 1: with you is clearly, I think too narrowly about search. Right.
• 24:32
  Speaker 1: I have this very kind of twenty years ago idea
• 24:35
  Speaker 1: of like searching text on the web, and the web
• 24:41
  Speaker 1: has become much less text based in that time. Right,
• 24:46
  Speaker 1: the web includes Instagram, the web includes TikTok, and those
• 24:50
  Speaker 1: are places where, weirdly to me, lots of people go
• 24:54
  Speaker 1: to search like people go on TikTok to find whatever,
• 24:56
  Speaker 1: where to go out to eat, which would never occur
• 24:58
  Speaker 1: to me. So I mean it's that part of the
• 25:01
  Speaker 1: sort of motivation on some level for you to figure out, Oh, right, text,
• 25:06
  Speaker 1: that's not enough, clue, we got to figure out how
• 25:08
  Speaker 1: to search in video and what does that even mean.
• 25:12
  Speaker 1: I think we're really driven by what our users are
• 25:15
  Speaker 1: telling us, and we just have really robust mechanisms for
• 25:20
  Speaker 1: understanding what our users are doing. And it's pretty clear
• 25:23
  Speaker 1: that people around the world find image and video content
• 25:29
  Speaker 1: to be pretty compelling, right, I Mean that's sort of
• 25:34
  Speaker 1: a very obvious statement, but you know the Internet in
• 25:38
  Speaker 1: the early days, it was banned with constrained. It was
• 25:41
  Speaker 1: technology constrained. It had to be words because that's what
• 25:45
  Speaker 1: the technology enabled, not necessarily because that's what human beings
• 25:50
  Speaker 1: most enjoy in terms of an information consumption experience. And
• 25:54
  Speaker 1: so we really are driven by what we're seeing in
• 25:57
  Speaker 1: the user trends, and we're really driven by just this
• 26:00
  Speaker 1: mission of how do we keep helping people get the
• 26:03
  Speaker 1: best answers to their questions that we can give them.
• 26:11
  Speaker 1: In a minute, the Lightning round, including what you learn
• 26:14
  Speaker 1: about the Internet when you spend six years working at
• 26:16
  Speaker 1: Google Search. Now, let's get back to the show. I
• 26:27
  Speaker 1: want to do a lightning round. We close usually with
• 26:29
  Speaker 1: a lightning round on this show, just a bunch of
• 26:32
  Speaker 1: relatively quick questions. So in this instance, I googled best
• 26:37
  Speaker 1: Lightning Round questions, and right at the top of the
• 26:41
  Speaker 1: search results page I didn't even have to click through,
• 26:43
  Speaker 1: is this bulleted list. I'm just going to give you
• 26:45
  Speaker 1: a few from there. Sounds good. Favorite day of the week,
• 26:50
  Speaker 1: oh Monday, because I get to go to work and
• 26:57
  Speaker 1: not deal with my kids all day. Who I love
• 26:59
  Speaker 1: very daily. Good favorite city in US besides the one
• 27:06
  Speaker 1: you live in just reading here New York's City. Thank you.
• 27:12
  Speaker 1: Would you rather be able to speak every language in
• 27:14
  Speaker 1: the world or be able to talk to animals? Speak
• 27:19
  Speaker 1: every language in the world. I'm shocked, to be honest,
• 27:22
  Speaker 1: Although I get that like a Google might actually figure
• 27:24
  Speaker 1: that out. Does it not seem like you can already
• 27:27
  Speaker 1: get a translator for every language. Talking that animals would
• 27:30
  Speaker 1: be like a revolution and human understanding of the natural world,
• 27:35
  Speaker 1: I guess. But I do not speak any language as really,
• 27:38
  Speaker 1: and I constantly feel bad about it. So maybe that
• 27:42
  Speaker 1: was just a fair personal feeling of weakness. So okay,
• 27:48
  Speaker 1: so we're now we're pivoting out of the Google lightning
• 27:51
  Speaker 1: round questions into my own bespoke lightning round questions. What's
• 27:57
  Speaker 1: your favorite kind of cake to bake? Oh? Well, so
• 28:03
  Speaker 1: I really make these, like quite elaborate cakes for my
• 28:09
  Speaker 1: children because I want them to be able to grow
• 28:13
  Speaker 1: up and say, wow, I remember you making great cakes
• 28:15
  Speaker 1: for us. Mostly so I recently made in one of
• 28:20
  Speaker 1: my kids plays Minecraft a lot, and there is a
• 28:23
  Speaker 1: character called a slime, which is a sort of jelly
• 28:26
  Speaker 1: blob that kind of jumps on top of you and
• 28:28
  Speaker 1: kills you if you don't fight it off. And so
• 28:30
  Speaker 1: I made a slime cake with a cake embedded in
• 28:34
  Speaker 1: the jelly. So big idea one here? What do you
• 28:38
  Speaker 1: think you understand about the Internet that most people don't understand? Oh?
• 28:43
  Speaker 1: I like this one. I think most people don't understand
• 28:47
  Speaker 1: how much it changes every day. And you know, so
• 28:50
  Speaker 1: we have this astonishing stat that even I didn't believe
• 28:54
  Speaker 1: when I heard it, which is that fifteen percent of
• 28:56
  Speaker 1: the queries Google sees every day we have never seen before.
• 28:59
  Speaker 1: And that happens every day. That there's fifteen percent, I
• 29:02
  Speaker 1: just completely new. And the same happens on the internet side.
• 29:06
  Speaker 1: Every day we index a ton of new content we've
• 29:09
  Speaker 1: never seen before about ideas that are completely new to
• 29:13
  Speaker 1: humanity at that time, right, And you know, we have
• 29:18
  Speaker 1: to be able to continually understand that and keep up.
• 29:22
  Speaker 1: And I think that people sort of have this idea
• 29:25
  Speaker 1: that there's a fixed amount of information out there, but
• 29:27
  Speaker 1: actually human beings are astonishingly productive and are constantly coming
• 29:31
  Speaker 1: up with new ideas. If everything goes well, what problem
• 29:35
  Speaker 1: will you be trying to solve in five years, I
• 29:38
  Speaker 1: will still be working on making Google Search better for
• 29:41
  Speaker 1: all our users. I think we will I think we
• 29:44
  Speaker 1: will be working on this for the next hundred years.
• 29:48
  Speaker 1: Is there a narrower answer, like this particular problem you're
• 29:51
  Speaker 1: working on now of integrating image and words basically like
• 29:57
  Speaker 1: you think you'll obviously it won't be completely solved, but
• 30:00
  Speaker 1: you think that'll basically work. And if so, is there
• 30:02
  Speaker 1: a next thing? I think the problem of a video
• 30:10
  Speaker 1: I think will continue to be hard because there's just
• 30:13
  Speaker 1: such a large amount of information in a given video.
• 30:17
  Speaker 1: The other problem that I'm really interested in is helping
• 30:22
  Speaker 1: people pause information with helpful context. So, like, how you
• 30:27
  Speaker 1: know we've unleashed the all of the world's information on people,
• 30:32
  Speaker 1: how do you actually help help them sift through that
• 30:36
  Speaker 1: and make good decisions, whether it's choosing a reliable merchant
• 30:39
  Speaker 1: to buy from or finding reliable medical information? How do
• 30:43
  Speaker 1: you help people make those decisions for themselves and be
• 30:46
  Speaker 1: literate with their information choices. What's one piece of advice
• 30:49
  Speaker 1: you'd give to someone trying to solve a hard problem.
• 30:55
  Speaker 1: I would say, find a really great group of people
• 30:58
  Speaker 1: to help solve up with you, because generally trying to
• 31:02
  Speaker 1: solve hard things by yourself enser being an active fustriction.
• 31:09
  Speaker 1: Kathy Edwards is vice president and g M of Search
• 31:12
  Speaker 1: at Google. Today's show was edited by Robert Smith, produced
• 31:18
  Speaker 1: by Edith Russelo, and engineered by Amanda k Waugh. I'm
• 31:22
  Speaker 1: Jacob Goldstein, and we'll be back next week with another
• 31:25
  Speaker 1: episode of What's Your Problem.