Pablo Mendes on How AI is Changing Search and the Next Trends

In the exciting new episode of Growth Talks, I had the honor of hosting Pablo Mendes, a true pioneer in the field of artificial intelligence. With an impressive work background at industry giants like IBM and Apple, Pablo has been working with AI since its early days, long before it became a "cool" trend.

During his time at IBM, Pablo Mendes was involved in the Watson project, and in his experience at Apple, he worked closely with the renowned virtual assistant Siri. Now, with his startup called "Objective," he has set out to revolutionize the way we search for information, harnessing artificial intelligence in innovative and extraordinary ways.

In this gripping episode, Pablo shares his insights on AI, exploring why the topic has exploded in recent months and how AI is fundamentally transforming the way we conduct online searches. Join us for an enlightening conversation about the future trends of AI.

Don't miss out on this episode! Tune in until the end to discover the exciting possibilities that AI holds. Listen now!

📚 Suggested book:
https://amzn.to/46TdCn3
https://amzn.to/40okvdI
https://amzn.to/3QIkC0t
https://amzn.to/3seEK0A
https://amzn.to/3u5Dopm

🛠 Suggested tool:
https://www.notion.so/

👍 Follow our guest here:
https://www.objective.inc/
https://www.linkedin.com/in/pablomendes

🧲 Watch my free lead generation course:
https://gaito.link/skillshare

🤖 Watch my free ChatGPT course:
https://skl.sh/3ZKHFcV

🙏 Subscribe to the channel:
https://gaito.link/subscribe

📚 Download the Reading List:
https://gaito.link/gtbooks

Episode Transcript

(00:00):
Show, everyone, and welcome backto Growth Talks. I'm rough your host,
and my guest today is Pablo Mendez. Hi Pablo, how are you?
Man Joao? How's everybody there?Great to be here, I'm great,
I'm great, super excited for this. What about you. Yeah,
it's been it's been a long timethat I've watched the show wanted to be

(00:23):
here, so thanks for having me. Yeah. I've tried, I think
for at least maybe three or fourmonths to do this episode with you.
But we're funny here. I can'tbelieve it's so super happy and and super
excited. Well, Pablo, weusually start, you know, with a
classic one, so very very simple. The first question is, uh,

(00:44):
just let's talk a little bit aboutyourself. So what's your story? You
know, who's Pablo. I don'tknow your back ground. You know,
what did you do in the past, what you are doing now? I
think that you can. You thinkit's kind of you know, interesting for
my audience and also to be thekind of a you know, the contest
for these episodes. O cat sure. Yeah, Well it's start from the

(01:06):
beginning. I am originally from Brazil. I came to the US to do
my PhD in computer science. I'vebeen working in AI and topics and artificial
intelligence since before it was cool.I've done a few different things there,
starting from like the end of mybachelor's and my master's and through my PhD

(01:30):
and then after that in my worklife. So the beginning of my research
career was in universities. I workedin the Free University of Berlin for a
while, and then I moved onto industry. So I worked for a
research lab at IBM Research. We'reworking on the IBM Watson project, which
is a project that did question answeringthat was famous for beating a human or

(01:55):
human competitors on a TV show calledJeopardy, which is you know, question
answering. And so I worked on, you know, taking the results of
that project to more customers and doingresearch to improve it over time. And
then I decided that I wanted thestartup experience, so I left this this
big research lab joined the company calledLattice Data. We were doing information extraction,

(02:23):
which is basically trying to understand what'sin text and then extract facts that
can be put in a database forpeople to do analytics and other things with.
That company was acquired by Apple.So I worked on Apple for a
few years where we were responsible forinformation seeking questions. So if you went

(02:46):
to Siri or Spotlight and asked likehow old is Obama? Or how many
COVID cases around me? What timedoes the oscars ceremony start? All these
questions about quote unquote anything, theywould come to this question answering system,
and we were responsible for several componentsthere to understand the question and then to

(03:07):
produce an answer. At that time, we were this was like way before
chat GPT. We were noticing howmachine learning was evolving and and there was
this big push around language models.Right the language model is becoming better with

(03:28):
neural networks, but also becoming larger, and we we we saw that,
like machine learning is going to change, is changing very rapidly. So that
inspired me to kind of question alittle bit the experiences that I was having
in the world that at the timewe had I had my first son.
My wife and I had our firstson, and I was looking for,

(03:53):
you know, those things that youput like around a baby that's learning to
crawl so the baby doesn't just likerun away somewhere. I didn't know the
name for that right as a firstparent. So I was trying to search
for in stores and it just couldn'tget anything useful. And I was like,
you know, thinking about language modelsand in machine learning and AI,

(04:14):
like really like search should understand isbetter? Right? And I found out
later I think it's called playpen.There are many like different things that serve
the same purpose, right, butplaypen was the one that I was trying
to find. And then I wasinspired by, hey, you know,
there's a shifting technology clearly is goingto be really important for people to find
things they want to learn about orthey want to buy. So I decided

(04:38):
to start a company with two cofounders, and we moved into a house
with our families in Hawaii for threemonths and it's kind of my co founders
and I locked ourselves in the houseand we hacked on our first prototype while

(04:59):
our families enjoyed the beach, anduh then we went fundraising and in kind
of the company was born. Ican talk more about that if it like
later, but that's kind of that'sthe story. Uh yeah, that's the
origin story, right from from Pablouh in Brazil, through research and then
to the startup world. That's whereI am. Now. So you said

(05:21):
that, you know, you startedyour the intro saying that you were working
on AI and machine learning before itwas cool. What do you mean by
dad? And also, why doyou think it's cool now? What happens?
What was the like the key momentsyou know in the whole AI ecosystem

(05:42):
that made it cool suddenly. Yeah, that's a great question. I mean
I should I should preface that statementthat I make in jest with it's been
cool and uncool and then cool again. So I like waves during the years.
Okay, yeah, they call likethe AI winters when it's like uncool,
right, and then it comes backup again. But yeah, so

(06:04):
you know, way back in whatsixties or like way way back in the
nineteen hundreds, you know, peoplewere talking about AI, and you know
this guy at MIT, Marvin Minski, famously predicted that AI is going to
be solved very quickly. And it'stwenty twenty three and I don't think it's

(06:24):
been solved yet, right, theartificial general intelligence. And I worked at
IBM Research. I worked with DanielGroll, who was a student of Marvin
Minsky. But back then that thekind of that first wave of AI it
became really cool because everybody was like, oh, my god, machines,

(06:46):
right, they're going to do allthese things for us. And then I
think it kind of hit a wallin terms of computing power and data availability
and a few things, and thenpeople are like, oh, actually,
it's not going to do what wethink it's going to do yet, right,
And then it started losing that kindof excitement around it. But then
researchers moved on, or most researchers, I should say, moved on to

(07:08):
like more practical things for the time. So new algorithms came up for how
to train machines to be really goodat specific tasks. Right. So that's
kind of when the machine learning theAI term started losing excitement. But then
the machine learning term started gaining excitementand it focused away from like, oh,

(07:31):
let's do everything that humans can do, and focus too, let's let's
get the machines to do useful thingsfor us. Right, So you train
the machines a very specific task.Typically would require somebody to sit down and
label example, so show the kindsof inputs and outputs that you need,
and then there is an algorithm wherethe machine learns to repeat the output from

(07:57):
given an input, and also howto as much as possible to generalize the
things that hasn't seen before based onpatterns that it recognizes. Right, So
that's like there's a there's this supervisedmachine learning area where it's it's it's the
showing examples of learning patterns that neededengineers to then kind of show for each

(08:18):
example, show the machine what arethe important aspects for each example. So
it's called feature engineering, right,So you're trying to like, Okay,
so if I'm trying to predict theprice of a house, it's important to
know if it has a pool,and it's important to know how many square
meters. So you start describing allthe characteristics, all the features that an

(08:39):
example has and the input, andthen the machine learns how to predict an
output based on everything that has seenin terms of patterns across these features.
Then, so that it was youknow, there was a lot of interest
in the technical community around that,but the general community was just like,
as you probably like, if you'renot a technical person and you heard my

(09:00):
explanation, now your eyes probably glazedover a little bit and I'm like,
where is this guy talking about allthis nerdy stuff right, and so it
was like inaccessible to many people oruninteresting to many people. And then there's
this new wave that came after that, which people called deep learning, which
is like this resurgence of neural networkswhere you don't have to do this feature

(09:22):
engineering anymore. This process of liketelling the machine what to look for and
you let the neural network learn thefeatures from the examples. Right. So
when that was around twenty twelve orso when it started getting really hot again
and then people are like, oh, there's so much more that we can

(09:43):
do now. And then ten yearslater, i'd say, right, or
a little less than ten years laterfrom the deep learning era, then it's
the large language networks era, whereyou don't even have to be a software
engineer to tell a neural network todo something for you, because now the
large language models understand English right orquote unquote understand English right. So you

(10:05):
can't get really useful things out ofthem. They're not super reliable. Sometimes
that will come up with things thatdon't exist and it will do the wrong
thing. But this idea of likeit became cool again. I think it's
very much related to who can nowderive value from from AI. Right.

(10:28):
It was at some point it wasvery very technical people only, right,
and then it became like, ohmore people and now even more people.
The barrier of entry in a senseto like program these ais right, or
to like tell the AI to todo something for you, that barrier of
entry has lowered. So that's mymy hypothesis for like why it's blowing up

(10:50):
now right this. There are sometechnical things that enable that, like transformers
in GPUs and TPUs, et cetera. I could talk about that if you
want to get into more technical things, but I'll stop here for now.
No. I like, I likethe uh, you know, the overview
that you you gave us, andyou said you made the example you know
of the what's the name of thething you were okay, yeah, So

(11:13):
you said that that was maybe theI don't know, d D d D
the light bulb moment for you toto to see that maybe there was an
opportunity in the market. Uh,what was the opportunity and what you're doing
on that? I mean, what'syour company about? Yeah, well,
our company is called Objective Inc.We used to be called Kailu eLabs because

(11:37):
we created the company, uh andwent to spend some time in in Hawaii
in the city called the Kailua overthere. Uh, and we launched the
company now with our forever name calledthe Objective, Inc. Uh And UH
what we realized was the previous journgeneration of machine of search systems just wasn't

(12:05):
focused around understanding what people are saying. It was focused around matching keywords in
the query with keywords. And let'ssay the product description. If you're talking
about buying things, right, sothen if you say, if you use
a slightly different wording, or ifyou search like like a in that example,

(12:26):
if you search by the description ofthe thing you're looking for, if
you just use slightly different words,you just don't find what you want,
right. Uh. So the insightwas, Hey, we need to really
focus around understanding what the products offeror what the content contains, and we
need to focus around understanding what peoplemean. And with large language models,

(12:50):
there's there's a huge opportunity now thatthere's actually that kind of understanding available not
only for text, but also forimages and video and audio and et cetera.
Right, so that that's going tochange things. So the idea of
the company is, let's offer anAPI that's very simple to use. You
put your data in, you putyour text in, your images, your
video, your audio. You putall of that into an index that we

(13:15):
host as a service, and wegive you an API that you can search
or you can put that API inyour app. You can put that API
in your website. Right. Sowhen people go to your app, they
type what they want, you callour API. We understand what they need,
We fetch the objects that match thatand we return them to you and
then you can show them to people. Right. So we have people some

(13:37):
of our customers. For example,dribble dot com is a designer platform that
it's a community, it's a placeto find work, is a place to
find inspiration. So people go thereto search for designs and they can express
like, you know, I wantlike a coffee shop logo in the style
of the Pacific Northwest, right,and just like pour your dreams into the

(14:01):
machine and kind of let the machinework for you. And the idea is
that, you know, our APIcomes smart out of the box with knowledge
about the world, and it continuesto learn as people interact with them.
Right, So we shift the focusfrom like instead of programming them how to
match words or how to match youknow, what people are talking with,

(14:22):
what's in the documents. We shiftedfrom that, so like, how can
we help people achieve their goals throughthe use of an AI. Right,
That's that's the mindset shift that ourAPI is bringing to the market. And
I think what that's going to dois the marketized access. So any full
stack engineer can just save a bunchof time instead of having to learn go

(14:43):
there and like figure out how toconnect a bunch of components to get this
to work, which is, bythe way, super hard to do.
We just save your time. Hey, you need to solve a search problem
in your app or in your website, so you just come and use our
API and focus on achieving your goals. So so that's kind of that's the
the insight that we're bringing. Ifyou are enjoying this episode, please check

(15:07):
out my lead generation course. Youcan watch it for free on Guido,
dot link, slash skill Share gA. I T. As an entrepreneur,
marketer or business owner, you knowhow crucial lead generation is. In
this course, I'll be sharing withyou twenty proven tactics for lead generation in
both B two B and B twoC markets. You can watch it for

(15:28):
free on Guido, dot Link Slashskill Share g A. I t oh,
you'll find the link in the description. And I suppose the idea behind
is that people don't need to uselike keywords or you know, a nerdy
approach to search, but they justneed to write it down in plain English.
So I think about it. Iwrite it down, and the system

(15:52):
kind of understand what I mean bythat, and it gives me the right
information, the right answer right,and enables that behavior, but also even
for the keywords right. So fornow, if you went to search for
photos and you set recession. Let'ssay I have a slide deck for my
board and I want to have likea nice picture that talks about the recession.

(16:14):
If you just type recession and yousearch for a picture in a stock
photo website, if you try tomatch the tag, you're going to find
very few examples or none at all. But if you were able to understand
the concept of recession and know thatit has something to do with finance,
has something to do with financial ruin, maybe then you can find all these
photos. And you can try thaton our website in one of our demos

(16:37):
for image search and you'll see thedifference of like keyword search with this kind
of understanding. Even if you're usingkeywords, the deeper understanding of what each
keyword means also leads to better results, right, and then beyond that,
it allows you to search with morenatural language, just speaking what's on your
mind into the search box. There'ssomething coolious ere because I brought up blog

(17:00):
post maybe a month ago or somethinglike that. Sorry guys, it's it's
in town. But basically the blogpost was about that. I realized that
in the last months I've not beenusing Google Search anymore. And the reason
is that chgipit became part of mydaily routine, so it's always there on

(17:22):
one screen, you know, ona site every time I have like a
question, every time I have adoubt, every time I need to search
something or you know, to dosomething like this. My previous behavior was
to go on Google, you know, search for it and then you know,
go through the in the sert inorder to find the right link.
But what happened in the last yearsis that in my at least in my

(17:48):
experience, in my opinion, isthat the quality of Google was you know,
going down and down in town becauseof you know, too many adds,
too many widgets, also too manyany I don't know, you know,
blogs and useless you know stuff inresults, people trying to sell you
stuff with affiliated link. So itwas really really hard to find the right

(18:11):
information. And also it was kindof time expensive, you know, I
had to go through ten different linksand you know, and search for it
and find the answer. Now Idon't tag, I write it down in
playing English on playing Italian, andthe system understands what I mean and gives
me the just that information, youknow, without all the noise, the

(18:33):
background noise around it. Are thosetwo things somehow related? So I'm kind
of thinking is it a bottom upprocess or a top down process? So
is it the technology that's changing andso we are adapting our behavior or it's
our behavior that is changing, andso the technology is kind of you know,

(18:57):
allowing us to do that. Yeah, I think those are self reinforcing
cycles. Right. So it's like, so I feel luck at like Google's
trajectory when it was like back thenit was ten blue links. It's very
much keyword search that was already betterthan what was before, which was just
navigation. Right, if you thinkof like alta vista. Yeahoo back then,

(19:18):
right, they were like the directorywith categories. You need to go
into the category, then look throughthe website. This is the one I
want, right, so you wereyou were the search engine right manually,
but then a note to like okay, now they can match keywords to the
content. That was brilliant, andthen you know, over time they started
incertaining these features where it was likeyou can also ask and then there's like

(19:38):
fully like full questions with the answersunderneath, right, or I started doing
the auto suggest so when you typeof keyword, it shows you like longer
completions, right, so that likeyou can click on that, and then
you start asking more natural questions.And that's good because when like what they're
trying to do there is like bringthe behavior to the market so that you

(20:00):
can give them more information about whatyou actually want so they can understand you
better if you just give keywords.It's really hard to know what people mean,
right. You say, like you'resaying for app, like are you
trying to build an app? Buyan app? Or you know what's going
on? But if you say Iwould like to build an app, what
is the best technology now? Theyknow what you want, right, and
they can try to serve that.I think that you know, where where

(20:22):
I got interesting for Google is likesome of the circles that I that I
made. People joke that Google's nota search company, They're an ADS company
with a good search engine, right, and I think there's there. They
may have crossed the line there.But what you're saying is like, so,
I think, you know, thechanging technology enables new things, and
then people discover those new things,they start changing their behavior, and that

(20:45):
changing behavior then pushes the tech toadapt to that, and then that generates
this like virtuous Hopefully virtue is notvicious circle, right, where things get
better and better, And that's verymuch kind of the part of the the
thing that we're really excited about.You talked about using chat ept instead of
Google Search. I think it doesn'tend there. I think the conversational interface

(21:08):
where you like type this really longquestion and then it answers you. At
some point, it may start askingyou clarification questions as well, so you
can refine what you're saying, andjust like back and forth is a very
intuitive interface for people to communicate.That's how I'm communicating with you. We're
talking to each other, asking questions, asking clarification questions, telling things,

(21:30):
right, But it's also not theonly way that we communicate, and sometimes
it's not the preferred way. Sowhen I'm buying something, I walk into
a store, I don't go directlyto the clerk and explain everything that I
want and then you know, theygive me the thing. Right, I
will browse a little bit. SometimesI will search on my phone to see
which I all need to go tosave some time. Sometimes I will want

(21:52):
to talk to somebody for advice.Right. So I think similarly the interaction,
like the search interface that we're goingto have with the way we're going
to look for things through computers isgoing to have a mix of browsing conversational
filtering like this thing of like takeme directly to the thing that I want

(22:14):
to look at. And the wayI see this happening is all the AI
will disappear into the background. Rightnow, you go into chat, GPT
and then you do something, thenyou go into Google. Right, So
it's like the the AI is verylike localized, right, And I think
that you need to know where togo to do the thing that you need
to do. But in the future, like if you just think a few

(22:37):
years ahead, I think everything isgonna be powered by AI. Everything is
going to be smart. Every websitethat you're going to go to is going
to have some sort of AI behindit, right, And that's where we're
also positioning ourselves with these APIs thatobjective that we're going to make it really
easy for people just use AI intheir site in many different places, not

(22:57):
just the search box. Search boxesare start, but over time, more
and more things. Right, So, I imagine a world in which you
don't even know you're using AI anymore, but your life is just so much
easier. It does appeared into thebackground. I can talk to things,
I can ask questions, I cango directly to where I need, I
can smartly navigate through the content.Right. So I just imagine this world

(23:18):
where like it's just so much productivefor anything, sorry, for anyone to
get anything they need to get donedone. Right, it just becomes easier.
Maybe the experiments that Google is doingwith bart and Microsoft with being are
you know, going in that direction, so where you have you know,
kind of the AI power integrated inthe search engine somehow, I don't know

(23:45):
if it's the right way to putit. But I feel like now AI
is kind of a product. It'snot the right word, but it's a
product. In a few years isgoing to be a feature. So or
maybe it's already becoming a feature.I mean, if I open notion,
there is a AI notion. IfI open Canva, there is in Kanba.

(24:07):
I mean it's everywhere. Now thereis a button. I click on
the button, I can write stuffand they generate images or or you know,
or copy or or you know,videos or anything that I need using
API and you know, third partytechnologies. So that's the direct direction that
you see on the market. Yeah, I think it's it's like a product,
then it's a feature, and thenit's just table stakes. It's like

(24:30):
you can't claim anything like I haveAI. It's just like oh I have
I use a programming language, right, or like you know, it just
becomes like something that exists and itdisappears into the background. There's no point
in calling it out anymore because everything, if it's not parted by AI,
is just not good enough for humansto use anymore. Right. Yeah,

(24:51):
Yeah, it's there and it's partof the experience, and you don't even
know anymore that there is a behindOkay, that's Is there any apart from
these one any trends that you seein the whole, I would say AI
and especially you know search together somethingthat is opening that you see from you

(25:15):
know, your point of view.Yeah, I mean the two that I'm
that I'm really interested in are multimodality, which is a nerdy word for content
in different formats, right, Socontent and text that's one modality, right,
or content and image format or invideo format or an audio format.

(25:38):
So the multimodality is the ability tounderstand content no matter how it's recorded,
if it's in video or text orimage, right, which which with their
networks, we're just getting better andbetter during that. And couple that with
the fact that everybody has a cellphone now, everybody is making videos.
Everybody's consuming content like the yours forexample, right as they're driving the listening

(26:02):
to some audio or a they're sittingin front of the computer, they watch
the video with the images as well. They're reading your blog posts, which
are you know, in italent textright. So I imagine just with everybody
carrying a cell phone producing more contentover and over again and sharing that content
through TikTok or and stuff or whateveris the product of choice. Maybe uh,

(26:25):
stores are going to start collecting uh, you know, photos and videos
of the products that they sell whilepeople are using and doing them. Right
right now, when I go toI was looking for something for the pool
and I just found some some randomguy in Florida fixing his pool. I
wanted to see, like, howexactly do I put the thing in?
Right? And so I found someuser generated content for that. So as

(26:48):
people generate more content, it alsobecomes harder to just browse through it.
You want to like, you know, find it, find the one that's
relevant for you and find it quickly. Right. So this ability to understand
and these richer formats like images andvideos and the amount of these, uh,
the amount of content in these formatsincreasing, the problem just compounds and

(27:11):
the opportunity grows. Right. Sothat's the first trend that like, I'm
I'm super excited about and I haveanother one. But do you want to
do me to talk about both ofthem or one at a time? No,
one at a time. Yeah,let's dive into this one. Yeah.
So yeah, So that's the multimodality that's basically more more data being
created, more opportunities showing up,and it becomes really hard and and to

(27:36):
to deal with that as humans tonavigate that I a'mount of content and with
large language models becoming smarter and smarter, how to deal with these content.
Then now you can ask questions aboutit, you can navigate through it,
you can get directly to what youwant, right, which is again part
of part of the problem that we'retrying to solve. And what about the
second one, the second one washand the fact that a lot of as

(28:03):
technologies start getting created and evolving,they tend to be at a lower level
in the sense that you need likemore technical ability to if you think of
the machine as being the lower leveland the human being the higher level.
We start at the lower level andwe start building up right, So,
the first generations of search engines youneeded to program how to match the keywords

(28:26):
to the content. And as wego over time, you get to a
point where you allow the programmers,the developers that are using the search engines,
to focus on how to tell thesearch engine to achieve their goals.
So you know, if you have, for example, an e commerce website

(28:48):
and you're trying to get people tothe product that they need as quickly as
possible so that they get to buyit and they don't waste time and get
frustrated and leave and go buy somewhereelse. Right. So you want to
look at metrics like conversion rates uh, and you want the search to increase
the conversion rates. So conversion ratehere being somebody searches for something, then

(29:11):
they click on something and then buyit, add to cart and then buy
it, right. So you wantpeople to be that's a that's a statement
of success. The goal of theuser was to find something to buy.
They found something, they bought it, right, So that's that's the goal.
Can we then let the AI learnwithout me having to program every single

(29:33):
thing that the search system should do. Could we just let the AI learn
how to achieve that goal? Right? Uh? And this is one type
of goal. There are other typeslike sometimes you want to give people a
diversity of things for them to browsethrough and and and get inspired by what
they want to buy. Or sometimesyou may want to keep finding more educational

(29:56):
content to keep a student engaged inlearning. Right, So you can express
your goals and and have the engineoptimize for those goals. For you,
which is, you know, muchmuch more productive than having to then well,
okay, let me then program howthis keyword matches this keyword and if
it does match, then how doI, you know, figure out how

(30:18):
to put this thing on the topso that people can click on it.
It's it's just a different way ofthinking about it. Right. So with
the AI capabilities increasing, more andmore applications will become goal oriented, just
that you will start programming the goalsthat you have instead of programming the details
of how things are are matched.Right, and then you can do that

(30:41):
over any kind of content because youknow, the the the AI is becoming
better at understanding images and taxing andall of that. Right. So those
are two trends that I'm really excitedabout on the on the more I guess
technical side of it. A quickquestion on the second one because it sounds
I don't know, it sounds likesuper cool, and it also sounds like

(31:03):
something that you know, every companyshould do as a matter you know,
in which industry you work, soeveryone should do that. But you know,
there's a lot of discussion about thecost of AI that right now it's
kind of expensive to use, youknow, part of the technology you have
to pay someone else for APIs.APIs are kind of expensive and so on

(31:23):
and so forth. What's the solution. So, are we going to go
to a point where, you know, it gets really cheap or the alternative
is that everyone is going to buildkind of you know, their own internal
AI. Let's let's call it thatway. What do you see that in

(31:45):
the future building is even more expensivethan using I think very few companies will
want to build this. It's probablyeasily ten x for more expensive because you
need to hire like expensive engineers,and then you have to have the machines
to train and the machines to torun inference right, to run run time.

(32:08):
And then when you have these machines, now you need an infer team,
and then you need a data team, and then you know, and
it's like it gets really really expensive. So that's why we think it's a
good idea to have any to offerthis API, because I think we actually
lower costs, or I know thatwe actually lower costs by a lot.

(32:29):
Now, even the cost of theAPIs will also go down over time because
I think that they will have moremachines, more and more GPU machines that
we're a little kids trained on theavailability of machines now, right, so
they will will have more of themand cheaper. Also that the models themselves

(32:50):
will evolve to be able to runmore cheaply, or there's a lot of
research and how to get the modelsto be smaller or even get the the
large size models to not need thefull computation to get to an answer.
So there's a lot of like reallytechnical and really exciting things here. So
I think that like all this isgonna just go down and and and costs

(33:13):
over time and become more accessible.Okay, okay, So there may be
several factors that now are impacting thecosts and everything is going to go down
in a few years maybe, soit's gonna be way more cheap and also
easy and accessible for companies to touse that. Yeah yeah, yeah,
okay, cool cool. And Iwas thinking, do you see any since

(33:37):
you said that you give us theexample of one of your clients that is
what was the platform jibbile, right, yes, yeah, so do you
see any I would say, whatwhat what? What are they looking for
at the moment? I mean companiesthat you know want to integrate this kind

(33:59):
of technolgy g inside their product.So it means that they have a you
know, the brand is a goodbrand. They have been there on the
market. You know, they havea community, they have an audience.
The audience trust them. They havea product, maybe it's a cool product.
What's the reason why they decide todo you know, the next step

(34:19):
and now it's time to you know, to have some of these technology in
our product? Why they do that? Yeah, I wouldn't want to speak
on behalf of any of our customers, but I think in general, what
a what a company that comes totalk to us. In general, what
they want is they want to speedup their time to market, right,

(34:40):
So they want to get to greattechnology really fast. And you know,
our team has built this kind oftechnology for year as well. People were
Apple, they were a Google video, YouTube and Twitch, Amazon, Right,
so there are a bunch of we'vemade a bunch of mistakes over time,

(35:00):
learn from them, and we're buildingthis technology. We'll carry a lot
of the weight of the infra andall that stuff. Right, So people
come to us so they get velocityto get to that so they can get
the users to have a better experience, so the users can engage more,
and then they can learn from thatengagement and continue to provide better experiences for
their users. Right. So it'slike trying to set up that cycle where

(35:22):
you know that where like things getbetter, the behavior changes and learn from
that and they become better over time. Right. So it's like trying to
set that up such that it generatesthis the marketplace happiness, right where people
are getting value from from the appor the website and they are then getting

(35:42):
more excited. Then they're using itmore, and then you know, and
it goes from there. Market ismarketplace happiness. I love it. Yeah,
I don't create that term. Ibelieve in borrowing it from Sarah Tavell.
I think maybe just pronouncing the nameas well. I only read it
and never had to say it,but I can look it up and do
the source. But I also reallyliked that when I heard it. I

(36:04):
love it. I love it.I love it. And you said that
you also made a bunch of mistakesduring the way. I always love to
talk about lesson learned. So isthere any I don't know something that was
kind of you know that now youcan say, oh, yeah, we
were completely wrong on that. Someimportant lesson that you guys learned during your

(36:30):
you know, the last couple ofyears when you were building your technology and
your startup. Yeah, I thinkone of them that everybody makes is all
the all this stuff is really exciting. Engineers love new things, right,
It's like, oh, I wantto try this out, and then you
try a model for a demo andthen oh my god, it works surprisingly,

(36:55):
especially if you worked with machine learningin the previous generation and then now
you're coming to to this, youknow, LM or stuff that comes pre
trained right with a lot of dataabout the world. I had no idea
that I, you know, Iwas going to be able to do this,
and they're like, great, let'sput this in PROD and then it's
wow, it's like completely a differentworld because it doesn't want to fast enough,

(37:15):
right. You can't. You can'trely on the outputs. It will
make mistakes that you didn't expect itwas going to make. Right, and
going from that like impressively good tolike actually really good in production for your
customers. That little it feels likeit's twenty percent of the work of the

(37:36):
you know, it's already eighty percentgood, so it's only twenty percent to
do. That's actually like more thaneighty percent of the actual work they need
to do, right, So ittakes a long time, and it takes
a lot of expertise and like andlessons learned. So so that that kind
of hey, we think we're almostdone and then they're like, no,

(37:57):
not at all. That's like that'sone of the missed that I think a
lot of people make, and wedefinitely made in the beginning. Uh.
The other one is, uh,just scale, Like INFRA is just really
hard, and it becomes harder withwhen you have machine learning and AI in
the middle of it. Then thenwhen I move by by scale, I
mean just being able to deal withhigh throughput, like with lots and lots

(38:23):
of volume, like lots of peopleare asking questions at the same time or
searching at the same time and doingthat fast. Every response needs to come
back really fast, so all ofthat stuff. I mean, you have
great people in the company, souh so that's a lesson we had to
learn before. But like to getthings to run really fast at scale.

(38:45):
Uh, what was really tough whenI when I started with us. The
other one I think is quality evaluationis way harder than anybody thinks. And
this goes for technical people and nontime technical people, like just defining what
good means. Right, So Iremember, like I was doing some research.

(39:08):
I don't know over you mean qualityin the results, quality of what
precisely? Yeah? Like so sopeople say like, okay, I want
a really good system, and everytime I send a search result, I
want to get only good things back, right, very easy. Right,
Then you ask like, okay,so what does good mean? And then

(39:29):
putting that in words is tough,and sometimes it's even tough to like ask
people to point out which ones aregood or bad and then show another person
and ask them to do the same. And then if you do this with
like five seven people, they don'tagree on everything, right. It's like
there's lots of subjectivity that goes intothis. So the the the act of

(39:51):
evaluating quality or you know, ofa system or of a product is very
different to get right. And somethingthat we like, we've been working on
for a real long time and wehave good results, but it's you know,
it's difficult, and people just underestimate. And when I talk to people,

(40:12):
so I'm talking to prospective customers allthe time, and I'll ask them,
hey, you know, how's yoursearch? Are you pretty happy?
I'm like, yeah, it's prettygood. And then I'm like, okay,
let's try a few examples. Right, So I opened and we're trying,
and I don't say anything. Ijust show like they're like, oh,
actually, oh yeah, that's that'sactually not good. And it's because
like it's hard to track quality.So most people just don't do it.
They look at a few examples,they think it's like, this is pretty

(40:35):
nice. It works for the examplesI tried, but there isn't like a
rigorous let's collect, you know,a thousand examples and just make sure that
you know ninety nine percent of themare good, or like just try to
put a framework around it. Scientistsdo that all the time, but it
you know, it's it's hard tolike budget time to do that kind of

(40:57):
stuff for every single aspect of yourproduct because you're just building, right,
you have requests coming from users andall that. So this is one of
the things that we want to alsomake it a lot easier for for everybody.
It's just like, how do youtrack and improve quality over time?
Right? So that's that's a keything in our vision as well. And
I was thinking. You said thatif you ask, you know, five

(41:17):
different people what's good at what's notthey were not agreed. But I mean
that's maybe the best part of beinghumans, right, I mean we don't
agree because maybe you and I andand someone else we we you know,

(41:38):
take a look at those you knowresults, and for me, the first
one is good for you is notbecause you have a different I don't know
story and big ground you know,and then and then and stuff like that.
So I don't know, I'm I'mthinking maybe I don't know. My
question is is that you know,is there are like a unique solution for

(42:01):
dead or there is no point ofsearching for a you know, a unique
solution that because we have to beyou know, kind of I don't know,
to have different approaches and different ideasand in the way we evaluate quality.

(42:22):
Yeah, I mean on the technicalside, people call that personalization,
right, It's like, how doyou allow the system to adjust to the
definition of quality or the definition ofgood that Wrath has versus Pablo, Right,
So it personalizes to you. Butif we if we can get philosophical
for a minute, right, yeah, let's go. Let's go for it

(42:43):
all the way to the limit.And there's this problem of AI alignment,
which is essentially imagine if the AIcan do things like if it's it already
does things better than me. Right, there are certain topics that if Rath
asks chat BT, it will giveyou a much better answer than I public
could do as a human, justbecause I don't have that kind of knowledge,

(43:05):
right, And the AI brain canfit so much information and it can
read so much faster than I couldread, right that they it's it's easy
to imagine how the AD is goingto get smarter and smarter and eventually surpass
the amount of information that it canstore and the amount of questions that it
can answer correctly, and it willget to this superhuman ability right where we

(43:30):
need now, like the best sourceof information for anything that we want is
going to be an AI. Right, And imagine now the AI can not
only answer questions but perform actions,and they become the best you know,
solution for performing on action for us, and they start taking over more and
more of the things that we needto get done, which is super exciting

(43:52):
and in some ways scary a situationto be in, because you know what,
if they tell me the right thingto do is whatever, And I
don't know if that's the best thingor not because I don't understand that topic.
So I need to trust Do Itrust the AI or not trust the
AI? Right? So, ifwe continue to kind of play through that

(44:15):
vision, right, the I isdoing more and more things. It's more
capable than most individual humans. It'sprobably at some point capable then all the
humans put together. Right? Andnow, how do we know that the
decisions that the AI is making onour behalf are actually good for us?

(44:39):
Right? And how do we controlbeing that that's smarter and learns faster and
moves faster. Right? So theAI alignment problem is like how do we
get the AI behavior to match whatwe consider to be positive for us as
humans? Right? Yeah? Soso that, I think for me is

(45:05):
the ultimate challenge. I could seethe AI substituting lots of human labor,
and I think that the very lastprofession that's going to be left, it's
one where you your job is toensure that the AI's interests. I'm kind

(45:30):
of personifying all of that. Butthe i's decisions are aligned with what we
want humanity to be and become andevolved to be. Right. So Anyways,
it got really philosophical real fast.But something that like, you know,
it's really exciting and scary. Isthat like that when the AI is
doing everything for us? Right?Yeah, yeah, I see your point.

(45:52):
And I'm also reading and listening toa bunch of stuff on that problem.
And sometimes even are there too todefine the problem? You know,
to to to to ask the questions? Uh, because of course you know
when we talk about alignment and valuesand they are also different, I mean

(46:16):
for two different countries kind of youknow, different values and goals and and
ideas, so it's gets it getseven harder on that. But yeah,
as I said, it's kind ofmaybe super excited and also super scary at
the same time. And yeah,yeah, and your and your previous question
was like how do we know thatwe are? Well? I guess I

(46:39):
guess the way I interpreted the questionwas like, is there a one solution?
Like we're talking about search at thattime, right, like search?
And then and then quality is quality? Is it good? Is there something
that's good for everybody? Probably notright. We're probably different cultures, different
backgrounds, different people. We're allgoing to think different and what's right for

(47:01):
us is different from what it's rightfor another group or for another person,
and then think about I was talkingabout aligning the AI with humanity, but
humanity is not one thing humanity,So now how do you then align in
a way that you know, Iwas going to say pleases everybody, but

(47:22):
maybe pleasing is not the right thingto do. It's more like do the
right thing for everybody. Right,It's yeah, it's fascinating. So yeah,
it is, it is, itis. I love it. I
love it. Look, Pablo,we usually you know, before saying bye,
we close these conversations with two questionsthat are now kind of part of

(47:45):
the of the show. So toolsand books. So do you have any
books that you want to share withus with my audience, something that you
are reading right now, or somethingthat you have read in the last don't
know, a couple of month,I don't know, your average book,
whatever, something you think can theinterest for people that listen to this podcast?

(48:08):
Yeah, a ton. I mean, if if you were interested in
any of the things that I saidhere, you're probably gonna like to read
some of these books that I'm goingto suggest. If you if you just
thought it wasn't interesting, maybe don'tgo there. But the one the book,
one of the books that I lovedwas The Alignment Problem. That book

(48:28):
is by Brian Christian. Incredible,talks a lot about it goes actually into
some of how the tech works,which is which is interesting but at a
at a level that's easy to digest, but it talks about alignment. Uh.
The other book, it's this oneis kind of like off the beaten
path. It's called Manna by MarshallBrain. It's actually available for free online.

(48:53):
It's very short and it talks aboutkind of a few different routes that
AI could take us. I don'twant to say too much so I don't
spoil, but yeah, and ifyou read it for free and you like
it, buy it so that theauthor gets gets get some money off of
it. But uh, it's Ienjoyed it. I read it like ten

(49:15):
years ago and I was talking withmy co workers about whether we programmers were
ever going to be replaced by AI. And I remember ten years ago I
said, no way, it's likewe're going to be the last ones.
And now there's like there's Cursor,there's like all these tools that actually like
do pretty well at writing code.So I was wrong. I also like

(49:39):
The Coming Wave by Mustafa sally Man, and I just pronounce peoples names all
the time. I apologize, buthe's a former deep mind. He also
talks about challenges with AI becoming moreand more powerful, So that one was
also very eye open and and reallyinteresting and for entrepreneurs. Uh. So

(50:06):
that's on AI side, but forpeople that like want to create their business
or are creating their business, orjust are interested in the topic. I
like Built by Tony Fadell, thethe one of the core guys on the
product side for the iPhone and theiPod first and then and then more and
the nest after that. Uh.And I also like this book about marketing

(50:30):
called They Ask You Answer. Uh. It's says a lot of simple things
that like everybody should just like understandand be doing, but somehow we don't.
You know, customers are asking questionslike your website probably should answer those
questions, right. So there's nota lot of like comments there that I

(50:51):
really enjoyed. It reminded me liketo to do a better job with all
that stuff. So cool. That'sthat's the books. Uh. I leave
all the links in the description.Guy, So everythink it's going to be
here in the descriptions and on thetools side, they have any cool tools
that you use with your team andyour date day, uh, you know,

(51:12):
jump shameless plug for our search ap I if you're if you're trying
to search through any kind of content. But let me think one thing that
was surprising for me when we started, uh was I, you know,
I used the Google Suite like GoogleDocs and sheets because you know, we
we wanted Gmail for the company,so we we bought all that. And
then one of my co founders like, we need to be using Notion and

(51:35):
it and it was weird in thebeginning, and I tried, and all
of a sudden, I found outwe're using Notion for everything, like we
track like CRM and candidates and there'sthere's so much going on there. I
don't think the full team is soldon that yet, but like it's like
I'm impressed by how much uh howmuch usage, like all the different types
of ways in which we used it, And so I think that's the that's

(51:59):
the that comes to mind. Imean obviously the you know, the the
coding if you're a coder, likeall the coding tools that help you to
write code faster, definitely give aproductivity boost. So if you're not doing
that yet, I highly recommend youstart playing with them. And I think
that's Yeah, those are the onesthat come to mind now. Cool,

(52:22):
nice, and I totally agree onnotion. I mean, my whole life,
my whole life is organized with motion. I love it. I love
it. Okay, So Pablo,where can they find you? Read you,
follow you? Do you have anyuseful links you want to share with
us? Yeah? So the companyis at objective dot Inc. We use

(52:46):
this what I think is very excitingnew top level domain that not many people
use. Objective dot Inc. Isthe company, and then there are demos
and all that. You can signup for the wait list and and and
chat with our team there. AndI'm on LinkedIn, Pablo Mendez, feel
free to connect right, write somethingabout how you you know, how you

(53:09):
heard, so I know where you'recoming from. You're not just a random
person, so I'm always happy tochat about anything. I'm on Twitter as
well, Pablo Menda's nice. Cool, and all the links are going to
be in the description below. Pablo, thank you so much for your time.
It was great. Thank you.Love that go ahead work raft.

(53:29):
Yeah, thanks for listening to thisepisode of the podcast. I hope you
enjoyed it and you learned something new. Make sure to subscribe to us on
YouTube and Spotify to stay updated onnew episodes. With your support, I
can continue to bring great content andgreat guest to this podcast. So hit
the subscribe button now and I'll seeyou in the next episode.