Latest / Agentic Conversations (formally mlops.community) / Agentic Marketplace

Transcript

Agentic Marketplace · Agentic Conversations (formally mlops.community)

0:14
I'm with Donae and Pedro. Pedro's working at OLX, Donae is
0:18
working at Process all doing AI stuff, and we're going to talk a
0:23
little bit about what their visions are and how they're
0:26
working with agents. So Pedro, you've got all kinds
0:32
of cool ideas. I want to hone in on the idea of
0:36
how you take the vision of what is possible and then go into
0:42
actually like executing on it. And then later on, let's talk a
0:46
little bit about where your vision has been thinking about
0:52
just like these millions of things that we could be doing
0:55
with the new paradigm of agents. Sounds exciting.
0:59
Thank you. Thank you for having me.
1:01
OK, So just to, to broaden up the scope a little bit.
1:04
So to, to be fairer to what our LX is, we actually operate let's
1:10
say in two main verticals and then the general classifieds.
1:14
So we have motors, so essentially cars, we have real
1:18
estate, so 2 very specific verticals and then we have
1:21
general classifieds. Maybe we can take a step back, I
1:25
think which is as part of these disrupt efforts, both of us
1:32
within, within OLX, the idea is really to come up with new
1:36
experiences for our users, whether sellers or buyers.
1:41
But I think the, the first thing that was really required and is
1:45
still required is to actually understand what's this new
1:47
experience? What, what's the disruptive
1:49
experience? How do you go from the current,
1:52
the status school, the current way of doing things into another
1:54
thing? And what's this thing?
1:55
And that's probably the most difficult step across the world
2:00
development. I'm not sure if you agree or
2:02
not, but that's, that's my perception.
2:04
And thankfully it's mostly a challenge for product people.
2:10
The, the world challenge of trying to come up with a new way
2:13
of doing old things is not, not something I envy because there
2:18
are always a lot of constraints, right?
2:20
Every time you want to innovate, you end up with the same, in the
2:23
same old places in the same chat interfaces, in the same UI
2:28
interfaces, in the same floating widgets or something.
2:32
And it's not easy. But I think that both within
2:36
motors and real estate, we are able, we were able to carve out
2:39
these new experiences within real estate for the buyers,
2:43
within the motor space for for the sellers.
2:47
And I think we were able to deliver something which is
2:50
disruptive within these specific areas for a few reasons, some
2:54
technical reasons, some design reasons, some product reasons.
2:57
And motor and real estate are where folks can just say I'm
3:01
selling my house or I'm selling my car, and then they find it's
3:05
like peer-to-peer selling in that regard.
3:08
Yes, although as major categories within OLX they are
3:12
super focused on professional sellers, so have most of our
3:16
customers are real estate agencies and most of the the the
3:19
automotive customers are stands. Car, car.
3:24
Car dealerships. Dealerships.
3:26
That's the word, yeah. That's the word I was looking
3:28
for. I couldn't find it either, so I
3:31
understand. So, yeah, so the challenge is
3:35
you as a buyer looking for a house, what does it look like to
3:38
have for you to have a different experience that you normally
3:41
have, which is basically selecting filters or going on a
3:44
map and just trying to understand whether you can find
3:47
something within your budget and your, I know, preferences
3:51
lifestyle. So how do we translate this
3:53
into, into a new experience And the way that we did it or the
3:57
way that our product team did it was OK.
3:59
Maybe the shift here is actually considering looking for a house
4:02
as a major lifestyle event or a major life event, but mainly
4:08
driven by lifestyle. Meaning when you're looking for
4:10
a house, you're not just looking for a house, you're looking for
4:13
the amenities around the house, the services, whether it has
4:16
parks, neighborhoods, exactly the neighborhoods, what type of
4:19
neighborhoods you're looking for.
4:22
So the, the, the challenge was really to try and translate this
4:25
into a flow where an agent tries to extract this from you without
4:31
being too invasive, without being too laconic, without being
4:35
too funny, with the, the proper tone and very quickly, hopefully
4:41
getting the persona out of you, trying to map you into an
4:44
intent, which of course is dynamic.
4:47
You can go into into all those details later.
4:50
But the goal really is to have an interaction with you and
4:54
being able to model you in a way to build your profile and then
4:58
provide the within our offering within our stock, provide the
5:03
best matches for you which are enriched in a way that usually
5:10
these platforms don't do it. Specifically.
5:13
Making sure that these new houses are within like an
5:16
acceptable commute time from your workplace, for example.
5:20
Making sure that like the neighborhood safety scores or
5:25
services scores are sufficiently high to match your preferences.
5:28
And basically using all of that and enriching all of that and
5:31
presenting to you as well. Recommendations And this is just
5:35
a chat that I say these are the things I care about.
5:38
It's a mix. So you have so we have not Gen.
5:42
not generative UI specifically, but basically components which
5:46
you pre populate with content which is generated by the agent.
5:49
So you can click and choose and not having it be 100% chat,
5:54
right? Because we have these intuitions
5:58
which are maybe wrong that maybe pure chat is not the way that we
6:03
want to go. But then again, all the examples
6:06
that we saw yesterday were pure chat.
6:07
So I don't know anymore. Yeah.
6:10
And so between this mix of chatting and clicking and and
6:15
nice navigation, then on the on the map page, on the listing
6:19
page, you'll have these new more enriched experience of searching
6:23
for a house. And now the whole idea of coming
6:27
from like idea to execution, that was something that I know
6:33
you were fascinated with because I think you've had a few war
6:37
stories. Yeah.
6:38
I mean, so yeah, we come from the same perspective, right?
6:41
Where like there's an opportunity to disrupt an
6:43
experience here and then what? What does disruption look like?
6:48
You know, if you can change everything, but if they're not,
6:51
if it's not a valuable experience, then you've changed
6:54
it for nothing. So figuring out the limits of
6:58
what people are willing to accept in terms of the change of
7:01
their flow in order to get a better experience.
7:04
And for us that was like, it's an ongoing process.
7:07
I think as well. We're like learning it's rating
7:11
in smaller steps. So, so I think like from the
7:14
real estate perspective, because that's also like a high trust
7:18
experience that they have to try and you know, to try and engage
7:21
in because they have to really be like, this is a more reliable
7:25
experience than if I search by myself, I'm.
7:28
Not missing out on anything on anything.
7:29
Yeah, exactly. Yeah.
7:30
So yeah, it's just like, how do you how do you start that
7:34
process of earning their trust or do you go straight to like a
7:37
full on chat experience? I think there are a couple ways
7:41
of going about this. So we had some previous data
7:45
from similar initiatives like Magic, which allowed us to model
7:49
the user behavior a little bit. So try to understand how exactly
7:51
people on the users were interacting with these
7:53
platforms. That's one way.
7:56
The other way is through research, right?
7:58
Because you can roll out these prototypes, whether functional
8:00
or not, with a user base and just extract feedback out of
8:04
them, Out of them, whether it's explicit feedback as in asking
8:07
them what they actually think about things.
8:09
And the user interface and flow are implicit as in how they
8:13
navigate and in the proposed app, the prototype.
8:17
And then with control rollouts and.
8:19
And I guess having in place is sufficiently like elastic
8:25
structure to quickly iterate on whatever comes back from these
8:28
rollouts, right? So our calendar is it's very
8:32
detailed in the sense that we are doing initial rollouts for
8:35
internal teams, initial rollouts for internal teams, but slightly
8:39
more encompassing and bringing over real estate experts, then
8:44
going on to public beta but invite only.
8:47
And then eventually the full rollout tune to the old user
8:50
base that we have from him over 12.
8:52
And throughout all of these moments, you have to be ready to
8:54
collect feedback and iterate whatever you are doing.
8:58
So I guess there are a few ways of validating how we are
9:02
translating this vision, but there's also a lot of trust in
9:07
the process. Trusting what?
9:08
Trusting now our own vision for the product and that it will
9:11
eventually. Yeah, it's going to be a better
9:13
experience. Yeah, I could see it easily.
9:15
As you said, buying a house. That's probably one of the
9:18
bigger decisions that you're going to make in your life.
9:20
If you're the average person, you don't go and buy 10 or 20 of
9:26
them through your life, maybe 1-2, three Max.
9:30
And so you wanna be very sure when you're buying the house,
9:35
what you're getting yourself into and you potentially having
9:40
this. These questions, just asking
9:43
questions about are there playgrounds around 'cause you
9:45
have kids, how's the school system?
9:49
You don't get that on other platforms.
9:52
I mean there are some, some offerings, but not we believe as
9:58
sophisticated and as enriched as we are trying to offer that's.
10:03
Are you also doing that with the auto stuff?
10:07
Because that also is a big light.
10:09
You buy cars probably more than you buy houses, I would assume
10:13
for the average person. So, but it is still a lot of
10:17
money that you're putting down. Yep, not right now.
10:20
We haven't been working on that experience for the sale for the
10:23
buyer side on motors. We have been focusing focusing
10:27
mostly on the seller side, the dealers.
10:29
Experience, yeah, I mean, this is again like a choice you make
10:33
with, with development and drive to develop to a sort of more
10:37
like mature, sophisticated place and then go live or to sort of
10:41
do a very small, very beta pieces.
10:44
You know, that that you that more people see.
10:48
I think for that, like for us with the dealers, like it's a
10:50
little bit safer to do that because they're not it's not a
10:54
single action. They're gonna take over the
10:55
platform where you have one sort of flow with them.
10:58
And then once they've bought the house, probably for a few years,
11:01
you're not gonna need to interact with the dealers.
11:04
You know, if there's a small mistake, they will probably be
11:06
back. So you have a a higher risk
11:07
tolerance because they'll be back.
11:10
So you can you try again. It's obviously not ideal, but I
11:14
think in that case we we had a little bit more safety and
11:17
releasing very quickly so. So what are you doing with the
11:21
dealers in the agentic way now? Yeah.
11:25
So I think that's what trying to figure out what you know what,
11:29
what does it mean to use agent like an agentic experience for
11:32
dealers busy beyond the chat UI? I think this is the big, big
11:35
question. Like we obviously you start with
11:38
the the chat UI because that's how we can start to collect
11:40
data. We start to earn their trust and
11:42
we start to like figure out what they want.
11:45
But how do you move past the chat?
11:47
And this is kind of the big question because there is a like
11:52
an over like overflow of, of how many like chat interfaces there
11:56
are. And like for the dealers, what
11:59
we're seeing is that they are they have their routines and
12:02
they know how to navigate the platform to get what they want.
12:05
And so you have to interrupt that flow in order for them to
12:07
find value in the agent. And the initial test we did, we
12:12
found that actually for them to move into a raw chat is very
12:17
like a huge hurdle because they're not sure what questions
12:21
to ask. They don't know what the agent
12:23
can help with, right? Because it's, it can't do
12:25
everything at this point, especially when you start,
12:27
right? It's got limited functionality
12:29
and they kind of either ask nothing or they ask like massive
12:32
questions where you're just like, that's a lot for one, you
12:36
know, for V1. So there's there's definitely a
12:40
learning process there for for how to give them value without
12:45
alienating them, I think. I was thinking about this
12:48
yesterday with the Thiago agent, the ifood payment fintech agent,
12:57
and how there's this necessity for agents to really help you.
13:05
They kind of have to embed in everything.
13:08
And with the Fintech, if it's embedded in your bank and it's
13:12
embedded in your PNL, it's really able to do a lot more
13:16
because it knows, oh, here's all your invoices, here's what your
13:20
cash flow looks like. But for in order to do that,
13:24
like you got to ask for, can we be your bank?
13:26
Can we do this? Can we know your PNL, which can
13:29
be a big ask for a lot of things, But I'm thinking about
13:32
like with these sellers, the more access that they give you,
13:37
the more you're going to be able to help.
13:40
Yes. And that's where the the trust
13:42
is the important part because if, if they kind of start to see
13:47
value or they start to see impact, then you can
13:49
progressively ask for more trust from them.
13:52
But if you kind of start on your first release and and
13:55
immediately you're like, can I have access to your?
13:57
But that's probably not going to go very well.
14:00
You should try it, see if it works tell.
14:02
Me how it goes, I mean, we like it really is a process.
14:06
We you know, the way we released is we had the chat and then we
14:10
had like a couple of shortcut buttons that they could use and
14:13
one of those all it did was like on the page create like a filter
14:19
for them so they could see a specific view of their data.
14:23
But the way we labeled that they were concerned that it would
14:25
take an action on all of those ads.
14:28
And so they refused to press it because they weren't sure what
14:30
would happen. And that that's kind of like,
14:33
you know, the the trust process and I think also the education
14:36
process. But yeah, I mean, if if we
14:39
released a button that actually took an action on their behalf
14:42
right there and then you can see we would have immediately like
14:45
lost them. So, you know, it is a
14:47
progression of in this case, we'll just show you the things
14:50
you might want to like act on. And then maybe once they trust
14:54
that, you can sort of start to do you want us to take an action
14:57
for you? And then only do you start to
14:58
like automatically do things so that that progression of like
15:01
earning their trust. I think is there any way you
15:04
sort of start to eventually get insights into much deeper
15:07
things? That being said, like with
15:09
agents becoming more like readily accessible, I think
15:12
people are gaining a certain amount of inherent trust to just
15:15
have their stuff integrated because they like the
15:18
convenience. But I think it's been a process
15:20
to get there. I.
15:21
Just realized you don't really have a control Z and the undo
15:26
button for when an agent takes an action.
15:29
It would be very nice if you could just oh oh, undo.
15:32
Like the reverse? Like the reverse changes in
15:34
cursor. Yeah, exactly.
15:36
That's actually a really good point.
15:37
That's probably a design element.
15:39
We should start thinking about an agentic products as a general
15:42
rule. Well, especially because, yeah,
15:44
if you're, if someone is a little bit skeptical or hesitant
15:49
to take an action, if they know, Oh well, I can just roll back
15:53
really easily with a control Z type of fast movement, then I
15:59
would be more aggressive in the actions I take.
16:01
If I can always just like go back five steps.
16:03
It's. A really, it's a really
16:05
interesting idea We'll. Be here all week, don't worry,
16:10
happy to give you all kinds of great ideas.
16:13
What you're really talking about here is something that I've been
16:17
saying for a while since I think, like, I don't know, 2024
16:21
on the interface is the main friction point.
16:27
And people not understanding the interface and not understanding
16:32
what it's capable of, what it's not capable of, what they can
16:34
ask of it, what it will do, what it won't do because it is a new
16:39
interface, really feels like it gives you this wow moment, but
16:44
it also can make you give you pause.
16:47
Yeah, I, I think I, I kind of think of it as a learning curve,
16:50
right? Because you're giving someone a
16:51
new experience often with like not a lot of information and
16:56
then they kind of have to sort of trial and error their way
17:00
through it. And that's where the, the trust
17:01
element comes in again, right? Because you're not going to
17:03
trial and error if you're worried that a mistake will be
17:05
long term damaging. So like for us with the
17:09
automotive project, I think that the chat UI is the natural start
17:15
because that's what how we kind of think about agentic problem
17:19
solving is that it, it does the thinking and then gives you an
17:22
answer to like a normal question.
17:24
So our chat agent does that, at least initially is it's, I kind
17:29
of think of it as a like a data analyst for the dealers where
17:32
they can get more insights into the data because there's a lot
17:36
and there's across multiple like pages.
17:38
And like, you know, synthesizing that into a conclusion is a bit
17:42
of a process. So the project kind of started
17:45
with the idea that we could, we could give them those insights,
17:48
but I don't know quite how we landed up having an interface
17:53
where the chat was only a component of what they saw.
17:56
And then we gave them also what we call like shortcut buttons or
17:59
quick actions. And some of those are AI and not
18:02
all of them are some of them are workflows.
18:04
But the like the interesting thing there, and this is where
18:07
I'm kind of becoming more of a like zealot for just the UX is
18:14
everything. And when you're talking about
18:15
agents and functionality doesn't matter until your UX is solved.
18:20
Because we did see they asked questions that were either too
18:23
big or like very unsure. But if you gave them an action
18:27
that fed them into the chat with like a pre set thing that we
18:32
could do for them, then they were they had a better idea then
18:35
much clearer idea of what to ask.
18:37
And then they were more likely also to just follow up with
18:40
natural questions. And also, I think that's for me
18:43
at least, like changing how I'm thinking about building agent
18:46
products because they get really smart and there's still this
18:51
this resistance. And so you're not getting the
18:53
full like use out of them until you figure out how to give
18:56
people a a mechanism to use it that makes sense.
19:00
I think that we are stuck in this transition step between
19:04
pure chat interfaces and non chat interfaces.
19:06
And what non chat interfaces actually implicate meaning is
19:12
the does the user understand that if it's clicking a button,
19:15
it's also AI behind it? It's also doing something
19:19
intelligent behind it. So maybe generative UI is like a
19:23
step in between. Like you generate the elements
19:26
that you need to interact with the user.
19:28
Maybe, I don't know, I'm pretty dumb when it comes to UIC.
19:32
Why? But maybe there's something
19:33
there because then you don't need to like predefine all of
19:37
the visual aspects and you don't need to go with pure chat.
19:40
Maybe there's something in between.
19:41
Maybe you can have a selection of widgets that can be populated
19:46
or used by the agent in order to convey the information to the
19:49
user and then to extract information from the user as
19:52
well and extract actions from the user so.
19:55
What you're trying to do is have the user take a certain set of
19:58
actions or gain insight. In this case with the auto
20:01
they're they have all this data, how can you see it differently?
20:05
Yeah. And how do you act on it?
20:07
I think that's that's our next hurdle.
20:09
Is that right? We have a way to show them
20:12
information and how do we give them a way to do something with
20:16
information? And that's where there's like AI
20:17
like this idea of like and interactive components, right?
20:21
So if you were to say OK to the agent or whatever, like here's a
20:27
predefined things of you can that you can probably do, but
20:30
only show the mechanism to do it to them when it matters.
20:34
And then the interactive UI components are great because
20:37
then it's not just, hey, this is possible.
20:40
Do you want me to do it? And then they kind of just say,
20:43
yes, but if there's like capacity to press a button or
20:46
like some like where they're taking the action.
20:50
But obviously the AI is, is. Yeah, exactly.
20:53
I think that's a really interesting idea.
20:55
And maybe, yeah, more of what we need to see from agents than
20:59
either full autonomy or none. Well, yeah, it does feel like
21:04
there's an easy win. You show a dashboard of data
21:08
that they didn't necessarily get to visualize in the past.
21:13
And you also say, you know what would help this number go up?
21:18
If you buy more from us. Yeah.
21:21
And then you give them a few options of how you can, hey, I
21:25
think you could change this number if you do one of these
21:27
things. Do you want to test this?
21:29
Yeah. Or all test them all that's the
21:32
ABC. Or D&D is all of the above.
21:35
Things like that are starting to happen, like in cursor when you
21:39
put it in planning mode, then it will ask you a bunch of follow
21:42
up questions and then it's just like a multiple choice form,
21:45
which is, yeah, I mean, it's a really nice way to keep
21:50
gathering new information. But again, moving away from just
21:52
raw chat because it would be a bit of it is a bit of
21:55
frustrating when you're in these flows and have to type the whole
21:57
time. This is what I meant and
21:59
everything. It's inefficient, yeah.
22:02
Yeah. And then on the other side of
22:03
it, it is probably more frustrating when it it the agent
22:08
goes off and thinks for a while or does for a while and then it
22:11
comes back and it's not what you wanted at all.
22:14
Yeah, the assumption was wrong and it's gone down.
22:16
Yeah. So I would rather have the
22:18
upfront questions and front load that and then I can go and get
22:22
some coffee while it's working. Yeah, exactly.
22:25
All right, donate. You've got to leave us for an
22:27
imaginary meeting that you're going to.
22:32
And Pedro and I, we get to talk a little bit about the future
22:35
and what you've been thinking about as potential realities in
22:38
the marketplace. Sci-fi meets marketplaces.
22:41
Nice, nice. All.
22:43
Right. Cool.
22:43
Thank you. You're trying to rethink the way
22:48
that you bring agents into the equation?
22:51
Yep, that's correct. What are some ways that you've
22:55
been thinking about? Within these general
22:57
classifieds, it's where we find, I think these not just within
23:01
general classifieds, but maybe it's more obvious that these
23:04
opportunities are there because that's where you find people
23:07
which are offloading their stuff, which are tired of their
23:12
chair and they want to refurbish their their room.
23:14
They are looking for for new mics, but they don't really want
23:16
a new one. So maybe in there somewhere
23:21
between the professional sellers will actually make a living out
23:23
of it. And there's a decent chunk of
23:25
people actually living off of, or at least we think they make
23:28
enough money to leave off of just spinning things on, on OLX
23:34
and the other types of sellers, which are just, let's say,
23:37
occasional sellers. Maybe there's an angle there, or
23:40
maybe there's a couple of angles there.
23:43
Yeah. What?
23:43
And what have you been thinking about?
23:44
Because some of the things that I can imagine are just easier
23:49
uploads. And so instantly I throw a photo
23:52
of my stuff and you have a description and you have and you
23:55
kind of have that in place already so that I don't have to
23:58
think as much when I want to upload something.
24:00
You suggest prices. In a way.
24:04
You kind of like tell me that's a little high for.
24:08
There are some price prediction levels.
24:11
And what else are you thinking though?
24:13
Like to cause the whole idea is to ease the friction, right?
24:17
Yep, exactly. And as a classifieds business,
24:21
the main point of friction is actually the transaction.
24:24
So you as a seller, me as a buyer, we kind of connect
24:28
through the ad page, but then it's a plus, right?
24:31
And so I have to send you a message and say, hey, is this
24:35
most common question? Is this still available?
24:39
Is it negotiable? Like Are you sure can't, can't
24:41
you like skim something off at the top?
24:43
Can you give me a discount here? Just have a kid.
24:46
Can you help me out? So this is the actual iteration
24:50
and probably where agents could start helping like because
24:54
imagine that not just from the perspective of a seller, but
24:57
also from the perspective of the buyer where you really want that
25:00
new toy for your kid, but you have some budget constraint.
25:03
You sort of have a time urgency as well, right?
25:06
To have a time window where it makes sense to buy that.
25:08
So you either spent all day searching on LX, refreshing your
25:12
searches, your saved searches and making sure that the buyer
25:15
still have these and that and repeat this all day.
25:17
Or you offload this to an agent and you just instruct your agent
25:21
to go up there, search according to our parameters.
25:24
And eventually if they find something actually negotiate for
25:26
you, 'cause you know your budget, right?
25:28
Maybe you can budget on your budget.
25:30
Maybe you can go a little bit on top, but you can formalize this.
25:32
You can tell your agent, hey, this is my budget, but I'm open
25:36
for negotiation. And maybe the agent, if they get
25:38
in touch with the with the seller and they can't really
25:41
comply with their constraints, maybe they can come back to you
25:43
and. Say, hey, found this, but it's
25:46
not in the area you wanted or it's not in the budget.
25:50
Actually, I'll tell you a true story of when I was in Portugal
25:53
and I used OLX to buy my daughter a roller like one of
25:57
those little scooters. And I do remember that we spent
26:01
a considerable amount of time with the back and forth of the
26:04
sellers deciding when we are going to meet, where we were
26:08
going to meet. Why did you do that?
26:10
Yeah, exactly. It's it's.
26:12
That's scheduling. That's scheduling.
26:13
It's a solved problem. Yeah, yeah, exactly.
26:15
It's like, oh, this is a solved problem now.
26:18
And even like the next step is something that you were just
26:22
saying. And I was like, we got to turn
26:23
on the cameras for this with the lockers where the agents can
26:27
then play with the locker idea and then you don't have to even
26:32
worry about when to meet someone.
26:34
Yeah, then this was full improvisation on my part, based
26:36
on the hiring of humans that you mentioned before.
26:39
Yeah, for context on that one, I saw a website where the molts
26:43
can now hire humans to do things in real life.
26:46
And so you can hire a human to sign a contract for you or to do
26:51
whatever it is that they can't do inside of the computers,
26:54
which I found fascinating, and within.
26:58
That within that context, I, I realized that maybe, well, maybe
27:01
smart lockers are another like a, a, a physical arm of these
27:05
agents, right? Because because they could deal
27:07
with the logistic chain and just make sure that whatever smart
27:11
locker supplier that's next to you or closer to you now has
27:15
your roller for kid. Instantly my mind thinks like Oh
27:19
my agent has a budget for the roller but it finds a roller
27:25
that isn't in the budget and so then it goes and it buys like 5
27:29
different items and flips them, makes enough money to be able to
27:32
buy the roller and then or loses all my money and then says hey
27:37
yeah about that sorry. Or get you or get you banned off
27:40
of. But that doesn't make sense.
27:43
I mean, the whole logic of arbitrage is something that we
27:47
have explored like flipping objects with a with a certain
27:51
almost guaranteed margin, like identifying good deals, making
27:54
sure that they are in sufficient condition for us to like a price
27:57
a little bit. Just like tweaking the
28:00
description or taking the pictures or something like that
28:03
and then reselling for a profit. That's actually a project that
28:07
that we did. And again, there's a chunk of
28:09
users for for for whom this would make sense.
28:13
Yes, within, within all. I love watching YouTube videos
28:16
on people that go garage sales and like yard sales and then
28:21
they find, you know, the golden Pokémon card and it's like, this
28:24
is actually 2,000,000 bucks. Wow.
28:26
And we found it in the garage sale.
28:28
And that is something that it feels like, oh, cool, those are
28:34
the people that are the professional sellers those or
28:37
maybe they're not professional, but if they know enough about
28:40
what to look for in the in real life, then they can sell it
28:44
online, They can make some money.
28:47
The only problem I have with those videos is that they kind
28:50
of like say, oh, look, this is on eBay.
28:52
It's selling for 60 bucks right now.
28:54
So they kind of in the videos, they'll basically say this is 60
28:59
bucks, but they don't sell it yet.
29:02
And so it's not really 60 bucks until you actually sell it
29:05
because I see people like they'll buy a whole lot of mugs,
29:09
coffee mugs. I don't know who's buying coffee
29:11
mugs on eBay, but they say like, oh, there's 5 bucks, this is an
29:15
extra 10 bucks or whatever. I'm like.
29:18
And maybe they have they have their own mental bottles and
29:20
they are really confident of. Yeah, like this sells like
29:24
hotcakes. Exactly.
29:26
But we're still getting humans in the loop here, right?
29:28
Because OK, we are in the wake of this mega hype from and then
29:37
multiple, right? Multiple now has close to 2
29:39
million agents operating on it, whatever that means, full of the
29:44
degeneracy, full of grifters, full of scams, whatever makes
29:48
sense. But at the same time, again, 2
29:50
million independent agents, maybe not fully independent, but
29:53
the million agents actually ended up there, right?
29:56
Autonomously. So what if we are the first
30:00
stepping into that place? But from a marketplace
30:02
perspective, why? What if we are the first ones to
30:07
using niches words, Nero's words, to actually build the
30:11
harnesses for an agent's only marketplace?
30:15
Because then we skip all these attrition involving humans for
30:20
the better or the worse. But at least we would be able to
30:24
see what would happen if we have autonomous sellers, autonomous
30:28
buyers with access to some kind of stock.
30:32
So let's think about classified steel.
30:33
So still thinking about second hand stuff.
30:36
But what, what would this look like?
30:38
Right. Because technically we can do
30:39
it. It's it's fast.
30:41
I mean, notebook was built, I don't know, in a couple of days,
30:43
right. Yeah, yeah.
30:45
So there there's really no like no, no, no actual barrier for us
30:48
to occupy that space. No, no actual barrier for us to
30:50
build that structure for independent agents to act in a
30:55
in a marketplace like an agent to agent marketplace.
30:57
Yeah, we have the payment integration.
30:59
Stripe will be chatting about it today.
31:03
Google will be chatting about it today here at Brazos.
31:06
So like even the payment infrastructure is there at OLX,
31:08
we don't have that payment layer.
31:12
So it's all. Yeah, we kept it in cash.
31:15
Exactly. Exactly.
31:16
But what if we act as escrow? What if we take buyers take
31:21
rate, sellers take rate, whatever from these autonomous
31:24
interactions? I mean, it's incredibly
31:28
attractive. And I think what I mean apart
31:34
from all the security issues, apart from all the hype, that
31:37
will just be hype. I think what changed in this
31:40
past couple of weeks was that maybe the general population is
31:44
now actually considering having agents working for them and with
31:49
them on a day-to-day basis. Yeah, doing more stuff.
31:52
Doing more stuff exactly outsourcing actions to them
31:55
because the same way that everyone started getting used to
31:58
using a chat assistant from ChatGPT like a clear qualitative
32:02
change in the way people interact with apps.
32:04
Maybe we are upon that step now with with a new change where we
32:09
can actually trust an agent. Well, maybe not trust, but we
32:12
can actually have an agent doing doing stuff and I think we
32:15
should ride that wave. Talk to me more about this idea
32:18
of harness and building the harnesses.
32:20
Everyone or mostly everyone is focusing about building the
32:24
agent itself, but there's a lot of incumbents there, right?
32:27
We have I mean, all the all the big tech is building their their
32:31
own assistance. I'm using Gemini.
32:34
I'm not sure what to use. Every someone here uses
32:36
Entropic, Tachipt, whatever. Probably the capabilities from a
32:41
user perspective are already being built or are already exist
32:46
within these agents or are easily like extended with MCP or
32:50
tools or whatever. What we really need is the
32:54
infrastructure with which they interact, like the actual
32:56
platform for, for these agents to in this case, negotiate their
33:00
items to make their postings, to upload their pictures, their
33:04
descriptions of their items. This is what the harness is.
33:07
This is what the like the, the, the exoskeleton for this
33:10
structure is, and probably what we should be focusing on right
33:15
now before it gets irrelevant. And then you have agents talking
33:20
to agents directly. That was what I was going to
33:22
say. Like why would there need to be
33:23
a platform in between? Why would it not just be agents
33:27
talking to agents directly and sending photos or or putting out
33:31
a bat signal saying I've got this to sell if anybody wants
33:35
it? Technically speaking, currently
33:37
for that agent to agent communication to exist, you
33:40
still need some sort of configuration in between, right?
33:43
So we either start by exposing the agent 1st and sort of a
33:46
server and then have the other ones explicitly connecting to
33:50
it. As of now we don't.
33:53
You don't really have that sort of mapping capability or radar,
33:57
the ability of checking out what's out there.
33:59
And in order to check out what's out there, especially when we
34:02
have multiple in multiple technical solutions, multiple
34:06
existing assistance, I don't see how you can go around them
34:09
having some sort of intermediate layer that at least tells you
34:12
what other agents are out there. And if it tells you what other
34:15
agents are out there, which protocols they are using, which
34:18
preferences they have or whatever, maybe there's room
34:21
then to still carve out something specifically for
34:24
marketplaces. Because then you will have this
34:26
layer where, OK, maybe you don't exactly provide all of the
34:30
capabilities of a market of a traditional marketplace, but you
34:34
provide the discovery of other agents.
34:36
And as of now, I don't think there's a good solution for it.
34:39
I might, I mean, everything moves so fast.
34:41
Maybe I'm I'm wrong but I haven't seen it so.
34:44
And when you're talking about marketplace, you're talking
34:46
primarily about selling things, right?
34:49
Because I've heard. And our services, our services.
34:52
Yeah, I've seen a lot of folks try this idea of, Oh, well, in
34:57
the future agents are going to need things done from other
35:01
agents, which I'm not so sure I fully grok or believe is going
35:07
to happen, but that is trying. That is like in flight right
35:14
now. And I haven't seen any of them
35:17
take off as of yet. Maybe it's because we needed
35:19
this moment in time or maybe not, but you're really talking
35:24
about how do we enable stuff to be sold, like physical things to
35:32
be sold? I guess the services piece, I
35:34
was thinking like, oh, services from agents, like an agent needs
35:38
a job done and so it will go. And yeah, pay for that.
35:43
One of the categories that we have within Alexis services and
35:45
by services I mean plumbing, electricians, painters,
35:48
whatever. So you can actually buy those
35:51
services, right? You can actually find the people
35:54
who are willing to do the service that you need for, for a
35:57
fee obviously. Yeah, but it's not like a
35:59
newsletter agent. I, my agent goes and contracts A
36:03
newsletter agent because I, I see these agents becoming so
36:09
generalized that they don't need to contract specific agents to
36:13
do specific things. You have the agent that can do
36:17
whatever it needs to do on the computer.
36:19
Now, as soon as we get out of the digital world, then you need
36:23
things done. But right now in the digital
36:26
world, an agent isn't going to go to another agent and say, oh
36:29
cool, you know how to write newsletters.
36:32
Come, I'm going to contract your service.
36:33
Yeah, I'm with you. I'm super skeptical about that
36:36
model. I don't think that we can make a
36:39
business out of it because not just because these agents are,
36:43
are are becoming very generalist as you say, but also because if
36:46
they are not good enough, and we could see that from Tencent talk
36:51
yesterday, they can build their own tools, they can build their
36:54
own sub agents. So why would you sub hire
36:56
something which is 150% commoditized right, which is
37:00
building your own agent, unless it's something ultra specific,
37:03
which if you are within the real, within the digital realm,
37:07
probably you can replicate. I've seen something come up that
37:11
maybe is along these lines where it was a harness marketplace and
37:16
so you could buy harness of different agents.
37:19
And so maybe that's the closest thing.
37:21
But again, this is it's a really hard sell.
37:25
Yeah, to be like, I'm gonna actually pay money.
37:27
Yeah. For a harness.
37:28
Yeah, that's physically saying that your agency is not is not
37:30
good enough. It doesn't have the tools to
37:32
build its own tools, for example, which is more or less a
37:34
given. Or you're not willing to put in
37:36
the like, it's not that hard to make it work.
37:39
You know, like even a genius, maybe the most incredible
37:42
harness ever is. But then then when we jump to
37:44
the real world as you were saying, then I guess we get
37:48
either into the sci-fi realm. Well, not just not really
37:51
sci-fi, because again, Tencent investing heavily in robotics,
37:55
the whole Chinese ecosystem investing heavily in robots as
37:58
that last mile of interaction with the world or you get humans
38:03
doing it. Yeah, which humans apparently
38:05
are happy to do these days, judging by how many people have
38:10
signed up for this website that you can hire a human for.
38:14
Yep. Yep, because someone will need
38:16
to come come in here and actually paint.
38:19
Yeah, that's smudge off the wall, right?
38:20
Yeah, yeah, exactly. Or fix the toilet when it is
38:25
broken. Well, what do you think about
38:27
that the because everything that we have discussed in these past
38:30
few days has been about selling stuff.
38:32
Yeah, you probably noticed that. On WhatsApp, on.
38:35
WhatsApp selling stuff on WhatsApp?
38:37
So is there room for any wine, social or ethical discussion
38:46
these days? It's a pretty weird feeling,
38:49
right? Cuz it seems like everyone is
38:51
developing at a really really really high pace.
38:54
But to push stuff for it and to buy.
38:56
Yeah, it is. It's amazing.
38:57
I thought about that at one point.
38:59
It's very much capitalism at its best.
39:02
I think it was when the $0.10 talk was happening yesterday.
39:04
Yep, Yep, Yep, Yep. Absolutely.
39:06
The world ecosystem is based off of recommendations.
39:09
Yeah, yeah. Like the the question about
39:11
someone in the audience ask but but can we trust
39:13
recommendations? How do we know if
39:14
recommendations are actually good for me or they are just ad
39:18
based? And I think the answer was
39:20
really fair is like no recommendations is that it's at
39:23
our core, it's embedded within the society.
39:26
Yeah. So every purchase that you do,
39:28
you're actually incentivized to leave a review.
39:30
So it's like it's next step capitalism or you wind up like
39:34
you as a human are embedded in the recommendation system.
39:39
You're just living in a recommended.
39:40
System exactly, and that will push the next bubble tea.
39:43
Yeah, exactly. It's not a simulation, it is a
39:45
recommendation. Exactly, exactly.
39:47
And you are and you are part of it.
39:48
Yeah, yeah, I do feel like it happens already.
39:54
And so why not figure out ways to reduce the friction?
39:59
Because like I was saying, that's right.
40:01
I have from being a human and living in the capitalistic
40:05
society, a lot of stuff that I've accumulated, especially
40:10
since I have two daughters. And as they grow, I just
40:13
accumulate more and more toys. And so this brings us to that
40:17
next idea that you were talking about, kind of adjacent to what
40:21
we were saying in the beginning. If I could just take a photo of
40:25
my living room right now and you could tell me.
40:28
I've got buyers for those Legos over there.
40:30
I've got buyers. They're willing to pay €20 for
40:34
those horses that you've got. You've got fifty horses.
40:38
Do you want to sell one or two of them?
40:40
I'm happy. I don't think my daughters would
40:42
ever realize they're gone and I can make an extra €50.
40:46
And free up the living room. No, that's that's impossible.
40:50
I would love that. Like you can choose how extreme
40:53
you wanna go on that much. Volume you have to yeah agent to
40:56
estimate how much stuff you need to also.
40:58
Like what is it the Yolo mode with cursor?
41:03
It's like I I want to go Yolo mode.
41:04
But, but, but that idea, what does it require?
41:07
It requires this again, this platform, this harness for your
41:10
agent to be able to expose your stock, in this case, let's say
41:14
to multiple other agents, yeah, which have maybe previously
41:18
registered with some sort of interest.
41:21
And then they would negotiate between themselves.
41:23
You would know nothing about this.
41:25
And at the end of the day you would get you some notification
41:27
to confirm the cell or something.
41:29
That is different than how I was thinking about it too.
41:33
I like that cuz they, the agents register interest and they're,
41:37
they're just saying like, hey, if you got anything.
41:39
As buyers do nowadays, right? Because when you go in a
41:43
platform, what you are doing when you are clicking and
41:46
navigating. You're registering interest?
41:48
Yeah. You're building an intent, and
41:50
we in the background are trying to extract that intent, right?
41:53
But agents probably are most likely needed to be explicit,
41:57
right? So I'm looking for these or
41:58
these and that within this budget, as we were saying
42:01
before, from the perspective of the buyer.
42:03
But then you have a million of these guys and then your seller
42:06
agent can just either broadcast or I don't know the details, but
42:12
are directly go through some previously vetted buyers.
42:18
I mean, the possibilities are endless here.
42:20
If agents are acting, how can fraud be exposed on that level?
42:28
You have to be that layer of trust.
42:30
That's probably one of the best reasons to justify why we need
42:34
this middle layer. This middle layer.
42:37
It's trust, right? It's an escrow.
42:40
Yeah, it's trust. I don't I don't know other ways
42:42
of. You would in this case.
42:44
Right now with OLX you're not taking any payments, but in this
42:47
case. It would.
42:48
There's a subset of categories and countries which have
42:52
explored paying cheap, but it's not a business model that it's
42:56
going forward. Yeah.
42:57
Yeah, and but in this case it would be more hold your money
43:01
until you get your thing. Maybe, maybe or we could use the
43:05
same model that we have right now, which is vases valley vases
43:08
services where you as a seller and the paying for extra
43:12
services within the platform, so more ads to promote the ads and
43:17
so on, which could clearly work with agents, obviously, right?
43:20
Your seller agent is limited in our platform.
43:22
You can only, your agent can only post one or two things.
43:25
So you can only sell one or two horses.
43:27
But then we can copycat the model that we have and say, OK,
43:30
you want them to sell off your 25 horses just it's a.
43:34
But when I go back to the friction, 1 friction point that
43:37
I had when I was buying the roller on OLX is how to get cash
43:42
and never operate with cash these days, which is for better
43:46
or worse. And so then I had to drive to
43:51
the person's house when they had their window of openness, right?
43:55
We had to figure out when that was and I had to come with cash.
43:59
So at first, before driving to their house, I had to go to an
44:02
ATM, pick up cash. All of that can be erased in.
44:07
All of that could be erased. I don't see why not.
44:10
If your agent has access to your Apple Wallet or or Google Pay or
44:14
whatever the transaction and then the transaction can can
44:17
happen immediately. And again, we have at least two
44:21
protocols now, one from Google and another one from open Eye,
44:25
right for for commerce protocols to try and unlock this and these
44:30
verified transactions between agents.
44:32
So again, I think the tech is there, someone just needs to
44:36
occupy that. Space.
44:37
But going back to this locker idea, I'm not sure I fully
44:40
understood what the value of the locker would be or how the
44:45
agents could interact with the lockers.
44:48
I was, I was thinking about your idea of, or your problem, which
44:52
was you had to figure out like the, the physical connection
44:58
with the, with the seller. In this case, when what I think
45:03
these Amazon lockers is interlockers and the doing is
45:08
they basically do a very nice analytical work and try to
45:12
discover what, what are the best hubs, right?
45:14
And then they have their lockers there.
45:15
And then these lockers are connected to the careers career
45:19
companies. So you, you can send, you can
45:22
send stuff to a precise locker within, within a certain space,
45:26
which in principle is relatively close to, like it's in an
45:31
optimal location, right? So you as a buyer can probably
45:35
rely on one of these lockers to be sufficiently close to you at
45:38
any time and whatever you want would just be delivered to you
45:42
there instead of you having to negotiate and the location with
45:45
the with the actual seller, which is also a risk.
45:48
Like with this physical contact, as much as it's also a good
45:53
thing about the human like nature and interaction, it's
45:56
also a risk. That's why police stations
45:58
actually have locations for these transactions to happen.
46:01
That's why there are a lot of recommendations on how to do
46:05
this transaction. But still, maybe there's some
46:09
attrition there that could be reduced because ideally at any
46:12
time you would be sufficiently close to any of these smart
46:15
locker hubs to place your order, whether you're here or in Lisbon
46:19
or whatever. Yeah, well, not for you to place
46:22
your order, for your region to place your.
46:24
Order Yeah for me to just take a photo.
46:26
Yep. And say get rid of all these
46:28
damn toys. And then you have a
46:29
notification. Hey, your stuff is in locker C
46:32
Here's the pin code. Yeah.
46:34
Or in this case, if I'm selling it, I need to actually put all
46:38
of this stuff into a bag and then.
46:40
Or have the career pick it up. I mean, eventually it could be
46:42
like a drone like in San Francisco, like those automated
46:46
delivery boats that go on delivering food in San
46:49
Francisco. But.
46:50
Also a drone. Particularly close to the drone,
46:53
Yeah, I mean. In the.
46:54
Not my area of expertise, but yeah.
46:56
And I think I feel this playing with drones as well in Brazil.
46:59
Because there it does hit on the speed you want things faster.
47:04
Yep, Yep. And frictionless experiences.
47:07
So if you're able to get it to that locker as soon as the agent
47:13
says yeah, you want it, all right, cool, boom.
47:15
And you figure out a way to get. It the wheels are in place.
47:18
This does. It does feel like we're quite a
47:21
few steps away from that world, but it's not that many steps.
47:26
It is. It's not.
47:26
That many. Steps.
47:27
You could envision a world where agents are acting on your behalf
47:31
in the marketplace and it's as easy as you taking a photo, or
47:37
it's as easy as you saying. I want this.
47:40
Give me a plumber. Yeah, or give me a horsey if
47:44
you're on the other side of it. Yeah, on the equation.
47:48
And then a drone comes to your door and you just give it the
47:53
bag of horses and it goes through a locker.
47:56
And I don't know if the locker idea and the drone idea is the
48:04
most optimal, but it feels like it would be the easiest first
48:09
step, like it would be the path of least resistance.
48:13
Maybe do that, I mean, and and there's a marketplace, maybe
48:16
that's not our worry, right, Because that's last last mile
48:19
logistics. It's a whole different space.
48:22
Maybe we just connect to the existing solutions.
48:25
OK, so in this region the agent can rely on smart lockers, cool.
48:28
In this region, the agent can rely on ground autonomous
48:33
robots. OK, let's let's outsource it.
48:36
OK. In this region, they can
48:38
actually call for a call for a drone to come to come here.
48:43
But that I would say that that's not really.
48:46
I mean, it's part of the scenario, part of the the whole
48:49
equation, but not part of the problem that we should be
48:52
solving as Oleks. Yeah.
48:53
Or as browsers, I don't know. You're mainly focused on how can
48:57
we get the agents to interact in them in this scenario, in that
49:01
middle layer, how do we enable the supply and demand?
49:06
Yep, exactly to happen because right now the agents aren't able
49:11
to have that marketplace. Yep, exactly how do we unlock
49:15
this item circulation? And I had been thinking of it
49:20
before as I'm going to ask either Claude or Gemini directly
49:28
for something kind of like we do right now on Google.
49:32
I want XYZ. So I search on Google and then
49:35
boom, I look through things and great, I've got it.
49:39
I hadn't necessarily thought of it the way that this is where
49:44
and I guess you could still interact through cloud or
49:46
Gemini. Why not as long as it's
49:49
connected to but the best market was.
49:52
Yeah, exactly. You are making a bet that
49:57
Clutter Gemini is going to use you as the marketplace layer and
50:02
you would have to be that marketplace layer because you
50:05
would have the most supply I guess.
50:08
Or the most seamless experience. I mean, I guess there are
50:10
multiple angles here, either the now traditional steel, modern
50:15
Geo, right generative engine optimization, like making sure
50:19
that you pop up in these results, making sure that the
50:22
interaction between these agents and your platform is seamless.
50:26
And you could go the integration way and have like Shopify bib
50:30
integrate their their role ecosystem within ChatGPT.
50:36
You should become as you should become so good and so large that
50:42
they want to integrate themselves.
50:44
I mean, I guess there are multiple options here.
50:48
But what we, what I think we can see is that there's like the
50:53
demand for this is clearly here, right?
50:56
People want to buy, are starting to want to buy stuff in their,
50:59
in their or through their agent, through their personal
51:02
assistant. Yeah.
51:03
Through the chat. Which we were discussing
51:06
yesterday might not be like the best interaction layer but.
51:11
But it's here and it's what we're dealing with right now,
51:14
and so can we make it better?