VCs surveyed across the industry ranked their most exciting enterprise tech companies and the #1 early stage pick was a name almost nobody had heard of. Eric sits down with Han Wang, CEO of Mintlify, the knowledge infrastructure platform that quietly powers the docs for Anthropic, Lovable, and thousands of other companies and found out their servers crashed overnight because of Open Claw before Han even knew what it was.
Then in the second half, Eric talks to Jesse Zhang, CEO of Decagon, ranked #4 on the late stage list in a category that includes some of the most well funded names in enterprise AI, on how agents are replacing call centers, why voice AI is closer than you think, and where the customer experience space is headed in the next three years.
Two of the most exciting under the radar bets in enterprise AI right now, in one episode.
00:00:00
Do you think you you ranked higher than Sierra on this on
00:00:02
this list which was interesting? Yeah, yeah.
00:00:05
We continue destroying the competitors, you know.
00:00:07
Massive, massive spike, especially during that one
00:00:10
weekend. I, I'll never forget.
00:00:11
Like our servers were having issues.
00:00:13
Everyone was questioning like where all these numbers were
00:00:15
coming from and it was just from this one deployment.
00:00:18
It wasn't even called open claw at a time.
00:00:20
I didn't even realize what the heck it was.
00:00:22
In the newsletter today we are publishing with Wing, the
00:00:26
venture capital firm, a list of breakout and prize technology
00:00:31
companies, the ET30. With Wing, we surveyed a bunch
00:00:35
of top venture capitalists to figure out what early stage, mid
00:00:38
stage, late stage, and giga stage companies they are most
00:00:42
excited about. And the number one on the early
00:00:45
stage list I'd never heard of, which is always a good sign that
00:00:49
you're, you're learning new things.
00:00:50
So that is Mintilify. And so we are going to have Han
00:00:54
Wang, the CEO of Mintilify, explain what the hell he's
00:00:57
doing. It's got venture capitalists so
00:00:59
excited. Then in the second-half, we will
00:01:01
have the CEO Jesse Zhang of Decagon, which is #4 on the late
00:01:06
stage list. Now, to Jesse's credit, the
00:01:09
light stage list is, is a heavy category.
00:01:12
It's the breakout one. I mean, it's, it's, you know,
00:01:15
it's kudos to you. But obviously number one is 11
00:01:18
Labs. Number 2 is versal #3 is open
00:01:21
evidence #4 is decagon #5 is glean #6 is Sierra Decagon's
00:01:25
rival. So kudos to Decagon for being
00:01:28
ahead of your rival. The Giga stage lists are the
00:01:30
names you'll recognize Anthropic data bricks, SpaceX, Open AI and
00:01:34
roll. So you can go to
00:01:35
thenewsletternewcomer.co. We'll publish all the lists.
00:01:38
We're going to throw some up here on YouTube.
00:01:40
Thanks to Wing for conducting the survey and working with us
00:01:43
on it. First up, we have Han from
00:01:45
Mintlify. I really like knew that I wanted
00:01:55
to start a company with my now Co founder.
00:01:57
Hanby started a company wanting to tackle a problem that we can
00:02:01
both relate to and so we picked a space that was so crucial,
00:02:04
crucial to us. It was about enabling and
00:02:07
empowering developers, as simple as that, right?
00:02:10
That was the anchor in which we knew we must had because we were
00:02:13
like, look, it's gonna take a decade plus to go build anything
00:02:17
significant, right? Like building a company is not a
00:02:20
thing you do overnight. Right, you need to be pretty
00:02:22
committed to the space. You have to be.
00:02:25
And so we were like, OK, what's that space look like?
00:02:27
And to us, it was about enabling other developers a problem that
00:02:29
we could deeply relate to. It's like.
00:02:31
So if we even did, let's worst case scenario.
00:02:33
Let's say we pivoted 8 times and you know like and spent a.
00:02:37
Were they, they were like big pivots or yeah.
00:02:39
Did you leave categories or? I didn't that's.
00:02:41
The thing it was always develop serving developers.
00:02:44
Exactly. And so the first application
00:02:47
was. What year was that?
00:02:48
When did it start? So this the first initial
00:02:51
iteration that eventually let us on this path started at the end
00:02:54
of 2020, 21. And then we didn't really land
00:02:58
on to what we're doing now with Milefi until I would say like
00:03:01
end of 22, like start at 23, depending on how you look at it.
00:03:05
All right, so explain what the company does today.
00:03:07
Yes. So we help companies build their
00:03:10
like knowledge base, developer platform, source of truth.
00:03:14
We call ourselves the intelligent knowledge platform
00:03:16
and the knowledge infrastructure.
00:03:18
So if you've been, for instance, to cloud code stocks.
00:03:22
Right, this is the documentation.
00:03:24
Space. This is the documentation space,
00:03:27
though it increasingly so. Millified does a lot more than
00:03:30
just docs. So if you've been, for instance,
00:03:32
reading up on how Lovable works, they're Lovable guides.
00:03:36
Those are all powered by Millify if you've been to the open claw
00:03:39
docs or health centers, right? They're also on there.
00:03:41
Now obviously we're going to get into are these docs for humans
00:03:44
or are they for agents? So, but we won't hit that
00:03:47
immediately, but it's sort of if you're building, if I'm a
00:03:50
company like Stripe or something building a big API, I wanna sort
00:03:53
of explainer out there why we made the decisions we made and
00:03:57
how to interact with it and how to get the most out of it.
00:03:59
Is that yes the right way to explain it?
00:04:01
Exactly the analogy I always like to say when people ask me
00:04:05
what Milefios is. When's the last time you
00:04:08
assembled IKEA furniture? I like swear it off.
00:04:12
I, I remember I was like, you were supposed to go to a party.
00:04:16
It was literally like Kara Swisher was having, I think like
00:04:19
a book party or something. And I started assembling it with
00:04:22
my roommate who is like kind enough to get wrapped into it.
00:04:24
And I, you know, I thought I was going to go to some party at,
00:04:27
like, 10:00 PM, It was like 3:00 AM before we'd ever, you know,
00:04:31
like, finish the furniture. And so it's like, I pretty much
00:04:34
swore it off. Then I've done, you know,
00:04:35
simpler stuff. But yeah, OK.
00:04:38
Yes. So some of that happened, right?
00:04:40
Yeah. So the the analogy I like to
00:04:42
give is like Milify is like building the instruction manual
00:04:46
for the IKEA furniture or maybe put a different way, the
00:04:50
assembly manual for Lego set, Lego sets and Lego, right?
00:04:54
I have done that more recently. There you go.
00:04:55
Yeah. Nice.
00:04:57
I got a Batmobile for Christmas, which was a totally random gift,
00:05:00
but it was actually really fun to do.
00:05:03
Yeah, 100%. That's you.
00:05:05
Have the little bags. One thing that made it so much
00:05:06
easier, which I had sort of forgotten, is like, it's so
00:05:09
staged out, you know? Yes, so you don't get
00:05:12
overwhelmed all at once. Exactly.
00:05:14
No, they're they're methodical with it.
00:05:15
You know, it's funny because like I've even seen like they've
00:05:19
gotten rid of the actual like hand printed like, well, they
00:05:23
they do have them, but you can literally scan some QR code on
00:05:26
the thing now and there's like a 3D version of it you can like
00:05:29
open on your phone and like it could literally like.
00:05:31
Piece. Oh, I haven't done that.
00:05:32
Interesting. It's the coolest thing, but I
00:05:34
digress to say that we effectively are building the
00:05:37
assembly kit, the instruction manuals, if you will, for the
00:05:40
Lego, you know, the Lego sets. And the reason why that's
00:05:44
important is because, well, I mean, just try, imagine doing a
00:05:47
Lego set without it, right? Because otherwise it's just
00:05:50
really a bunch of plastic blocks.
00:05:53
And the reality is for the vast majority of products and
00:05:56
services is whether you know it is for developers or not,
00:06:02
there's, you know, a need to explain how to use the thing in
00:06:06
order to make use of it, in order to actually go in and, and
00:06:09
actually, you know, piece things together, build it together to
00:06:12
kind of get this like Batmobile. So is this vibe coding your
00:06:16
docs? It's like you're sort of coding
00:06:18
with cursor or something, and as that's happening sort of on the
00:06:22
side, your documentation is changing.
00:06:24
In an essence, yes. What we did in the beginning was
00:06:28
when Hanvi and I decided to start Nilify, we were like,
00:06:32
look, we spent our entire lives reading some really, really
00:06:36
shitty docs, right? Like, you know, implemented some
00:06:40
some things the hard way because there's no one, no developers
00:06:42
has gone into actually clearly put thought into explaining how
00:06:45
it works. And so everything just felt like
00:06:47
you were trying to piece together like Lego blocks just
00:06:50
together and manually without really the thought of how to
00:06:54
actually piece things together. So we had to figure that out.
00:06:57
And so we're like, look, let's just build this docs platform
00:07:01
that was the one of Millify to the simplest ways in which we
00:07:06
could have developers engage with it, contribute to it, and
00:07:10
Bill with it. And that first version was just
00:07:12
like, let's just give people Markdown, which is this language
00:07:16
that obviously developers really prefer and work with and make it
00:07:20
super easy for them to work with.
00:07:22
And let's just let it RIP and see what happens.
00:07:26
Developers, founders, companies of all stages and sizes, in the
00:07:30
end, this is not 23. They just loved it.
00:07:33
Do you think, like, I would think a sort of era of vibe
00:07:37
coding would be terrible for documentation, or it's just like
00:07:40
if you have people, sort of, some of them not even coders,
00:07:43
sort of just throwing shit at the wall and trying stuff, are
00:07:45
they really gonna be so buttoned up that they're like, and we
00:07:48
want to write the guide to it? I barely understand how the code
00:07:50
works. Like, is that intuition wrong?
00:07:53
Yeah. It's actually, oddly enough,
00:07:55
more important. I would actually say that if if
00:07:57
not for kind of like the, the, the tailwind of AI and, and, and
00:08:02
vibe coding in general, I wouldn't even say where, where
00:08:05
we are today. And the reason for that is
00:08:08
because I go back to the assembly that the, you know, kit
00:08:11
instruction or the, sorry, the Lego kit assembly manual is
00:08:16
let's imagine you're now asking an AI to go ahead and assemble,
00:08:22
you know, like said Lego assembly kit.
00:08:24
Let's say there's a robot, an AI robot here and it wants to go
00:08:27
and assemble the kit. Well, the most important thing
00:08:30
it needs to do to actually go and figure out if it could or
00:08:34
could not assemble it or how to assemble it is actually to read
00:08:38
the guides. And in the same way that humans
00:08:40
do, except in our case and what we've seen, it's even more
00:08:45
important because the docs, the knowledge base, those things
00:08:49
that are traditionally seen as very like boring, are
00:08:52
coincidentally very information dense.
00:08:55
And that's where typically AILMS actually get all their source of
00:08:59
truth about what you are, what you do, how you do it, how your
00:09:02
thing works, and so forth. It doesn't really look at the
00:09:05
marketing fluff, right, Right. That's on your landing page,
00:09:08
which by nature is designed for humans.
00:09:11
It's like, here's the flashy words, the flashy colors, the,
00:09:14
you know, the all the nice things that kind of get a person
00:09:16
through the door to pick your product to learn how to use it.
00:09:19
It's like it needs to know the truth of what your thing
00:09:21
actually does. And all of that's, you know,
00:09:24
just like 0 fluff tolerant is all living in what is
00:09:27
traditionally docks. And so now if for instance, you
00:09:30
don't have your docks, let's just take a example of that,
00:09:33
right? Like, imagine if you're striping
00:09:35
your docks just completely disappear for a day, right?
00:09:39
Well, the first thing to note is that, well, well, first of all,
00:09:44
no one's going to learn how to use your product, for starters.
00:09:46
So good luck on boarding developers that way.
00:09:49
But more importantly than that, especially today, well, LLMS
00:09:53
aren't going to know how to build your product either.
00:09:55
And if even today, the vast majority of software is written
00:09:59
by AI, and if it's not already, it's, you know, obviously going
00:10:01
to be, how are they going to go know how to do it too?
00:10:05
They don't. To what extent are mintlified
00:10:08
docs for language models and agents today?
00:10:13
Like what do you think is the percent breakdown of humans
00:10:16
versus AI consuming it? It's about 5050 right now.
00:10:20
And do you have like do you see them as the same doc or you
00:10:24
think they will diverge this for human and AI doc?
00:10:27
Good question. So we actually work with, well,
00:10:32
the Clod Co team and the anthropic teams.
00:10:33
And, and funny enough, about a year and a half ago I had the
00:10:36
same question because we kind of saw the writing on the wall
00:10:39
ourselves. We're like, look, the role of
00:10:41
content, of the role of knowledge, the role of docs, all
00:10:43
that stuff, whatever you call it is going to be fundamentally
00:10:46
more for AI than it is for humans.
00:10:48
And back then this was just like, you know, like AI was
00:10:51
really taking off. It wasn't like Cloud Code became
00:10:53
everyone's, you know, you know, our hourly active use product.
00:10:58
We were like, look, we just see this being the case in the
00:11:00
future. And so we had a conversation
00:11:02
with the client, the anthropic team, and we're like, look,
00:11:05
should the docs or the content for humans be the same for AI?
00:11:08
It's a common question we get asked a lot, and the answer is
00:11:12
yes, because the reality is the LLMS are also instructed to read
00:11:18
things the way humans are, right?
00:11:20
So. They've been trained on human
00:11:21
writing, so they're pretty used to it.
00:11:22
They're pretty good. With it, some would say, you
00:11:24
know, they're pretty good at writing it too.
00:11:25
And so I think there's a lot of people over thinking a little
00:11:27
bit where it's like, oh, like, let me format it in this way,
00:11:30
like don't don't bother really. Like, you know, write it in the
00:11:33
the best way you possibly could to if you were to explain to a
00:11:36
human good docs is good docs. LLMS are going to ingest that
00:11:40
and know how to do that from there.
00:11:41
Why? Why can't an LLM just read the
00:11:44
code, sort of create it's own perception of what the docs
00:11:48
should be, and just operate off that?
00:11:50
Yeah, that's a really good question, by the way.
00:11:53
So there's two different reasons.
00:11:55
The first one is the reality, which is that good docs and
00:12:00
stuff that's actually useful, like the content that actually
00:12:04
should be in docs don't describe exactly what the code does,
00:12:09
right? And I think that's the same kind
00:12:10
of like idea of, you know, let's say you're like here furniture,
00:12:15
you know, you can kind of just glue together some plywood and
00:12:17
some, you know, nails and it can come up in a shape like this.
00:12:20
But is it a cabinet? Is it a shelf?
00:12:22
Is it designed for a decoration piece?
00:12:23
Is it designed to be used in this or that way?
00:12:25
Is it a tool? Is it here or there?
00:12:27
That's typically the information that is just more contextual
00:12:31
that adds on top of what is an existing pile of pieces, right?
00:12:35
Same thing goes for like you know, products and content.
00:12:38
If you really take just code and you spin up docs for that,
00:12:45
granted we do that at Milify, for the record, we get a lot of
00:12:48
traffic and usage out of it. My opinion is that that's
00:12:53
typically not additive information that's tremendously
00:12:56
useful. Now there are cases of which
00:12:58
there are install guides, SDKSAPI references.
00:13:01
It's like the very tactical glued to the the code type of
00:13:05
use cases. But on generally, very
00:13:09
comprehensive information should extend far beyond that, and for
00:13:12
an LLM to truly understand how your product works, it should
00:13:15
take a look at the code on top of the contextual information.
00:13:18
Is if, if the docs sort of give you a sign of how the company
00:13:22
thinks it should be used, the provider and you have a sort of
00:13:25
sense of they'll be supportive if we're using it this way or
00:13:27
we're sort of using it if they're building the product in
00:13:30
this direction in the future and sort of where the the provider
00:13:34
is like leaned in. Exactly.
00:13:36
Or even things like how does the like what's on the road map,
00:13:42
what's been tried and tested, what's changed?
00:13:45
Right. These things don't necessarily
00:13:47
always come directly one to one with the code base.
00:13:51
But you were saying you have a lot of customers who are like I,
00:13:53
I want the fast, easy, I don't think about it sort of version
00:13:56
we do what what is sort of the utility of that?
00:13:59
Yeah, well, it's just getting zero to 1, right.
00:14:01
So for instance, again, like the the reality of of, you know, of
00:14:05
managing docs, managing content is well, no one really.
00:14:08
People don't really love updating docs, right?
00:14:11
Like especially just getting, putting myself in the
00:14:13
perspective of an engineer, you know, having been 1 and I'm
00:14:17
still 1 to this day, it's like, look, I am not the best writer.
00:14:22
I am not trained on that profession.
00:14:24
I am just inherently like, you know, someone who wants to go
00:14:27
and tinker with, you know, writing code, writing product,
00:14:29
shipping things. And docs often times become an
00:14:33
afterthought. And so historically, like people
00:14:37
just don't have anything in the 1st place.
00:14:39
And then your users complain. They're like, oh, I don't know
00:14:41
how to use your pocket. And you're like, well I don't
00:14:44
know, go figure it out. Read the code base.
00:14:45
And they're like, what are you talking about?
00:14:46
Like what the actual fuck? It's like I'm the customer.
00:14:49
Yeah, it's like that's the most insulting thing ever, you know?
00:14:52
And so some of the stuff that we build is helping people get 0 to
00:14:54
1 and then. But I think what's more
00:14:56
meaningful than getting zero to 1 is actually making sure the
00:14:59
content is up to date and accurate and what we call self
00:15:03
healing and self updating. The reason for that.
00:15:06
And again, goes ties back into the point about, you know, what
00:15:09
you mentioned on like what percentage of docs now for
00:15:11
humans and AI? Well, like right now it's 5050.
00:15:15
It was 15% AI. It was 15% of all traffic at the
00:15:19
start of 2025. So it went from 15.
00:15:22
It's clear which way we're going here.
00:15:23
Yeah, exactly. So do.
00:15:25
You have a guess for the end of 2026. 9010.
00:15:27
Really. Already?
00:15:28
Oh, wow. OK.
00:15:29
And the thing to note is not necessarily because it's like a
00:15:31
lot less people are going to be reading docs, maybe in the same
00:15:37
way like the total percentage of people are reading like hardback
00:15:40
books these days versus, I don't know, listening to an audio book
00:15:43
or on the Internet. You know, it's certainly going
00:15:45
to have an impact, but it's just because the sheer volume of
00:15:48
knowledge consumption because of just automated knowledge work is
00:15:51
going to be happening with agents.
00:15:53
The greater piece of pie is going to be like 9010.
00:15:55
To what extent do you think docs for humans are just going to be
00:15:57
like, what the hell did I just build?
00:15:59
Or it's like it's a guide to this sort of thing that humans
00:16:02
play very little part in creating.
00:16:06
In terms of like, oh, humans didn't really create.
00:16:08
Yeah, it just that it was like if code is mostly machine built
00:16:11
it so say play this out, I guess if in two, five years whatever
00:16:17
timeline you think vast, vast majority of code is built by
00:16:21
machines, do you think your company is mostly serving those
00:16:25
machines who are building the docs or it's mostly serving
00:16:28
humans who want to understand what the hell is going on?
00:16:31
I think both are going to be the case and both are going to be
00:16:33
equally important and and like because there's first and
00:16:36
foremost the the like the creator side, the people who
00:16:38
built the product vibe, coded everything, used AI to create
00:16:42
all the code. Adding the context on how to use
00:16:45
it is even more important because the code is then
00:16:47
attached from you explaining how this, you know, machine
00:16:51
orchestrated thing works to humans and how it serves people
00:16:54
is going to be more important. But that's besides the point.
00:16:57
On terms of like the role of content, what is it for?
00:17:01
Is it for humans? Is it for AI, even in the 9010
00:17:04
world? I think both of these are
00:17:05
equally important. I'll tell you why.
00:17:08
The role of content is clearly or just knowledge and broadly is
00:17:13
clearly starting to diverge into two different directions.
00:17:16
The first one that we see is just implementation setup and
00:17:20
implementation guides. Again, you can think of your
00:17:23
like Lego kit assembly, you know, instructions.
00:17:26
Someone just needs to draw the pretty diagrams and the 1-2, I
00:17:30
don't know, 50 steps it takes to go build the Batmobile.
00:17:33
I'm sure it was a lot more than 50.
00:17:35
It's pretty complex, I'm sure. And then AI is going to
00:17:38
basically be the main reader and main gesture of that.
00:17:41
If I can task an agent to go and just build the whole thing,
00:17:44
then, well, the reality is, why would I?
00:17:47
Maybe someone would really enjoy building the, you know, the.
00:17:49
Bat, right? Yeah, that is the question.
00:17:51
When you when you use the Batmobile explanation, it's like
00:17:53
I'm building it for the fun of it.
00:17:54
Obviously Lego could sell it pre built and with code.
00:17:57
It feels like no one wants to build it for the fun of it.
00:18:00
Yeah, or let's yeah, let's say like assembling furniture as a
00:18:02
better example than like a SEC because you know, it's like, OK,
00:18:06
like do I really want to like, you know, piece together some
00:18:08
plywood and and you know, and nail you.
00:18:09
Do it because it's well with. With IKEA you do it just because
00:18:14
it's cheaper to get it to you, they don't have to do the work.
00:18:17
I assume with code it's more about customization and sort of
00:18:21
making fit with your particular house or whatever in the.
00:18:23
Metaphor 100% though even in software, right?
00:18:28
A lot of it is abstracted away, right?
00:18:30
It's like I can't vibe code my own payments infrastructure, I
00:18:33
can't vibe code my own database. I can't vibe code, you know, a
00:18:37
lot of the agent infrastructure, I can't buy code the LLF
00:18:40
themselves, right? And so it's still is kind of
00:18:42
piercing together a bunch of things, right?
00:18:44
And so on one hand, content docs become this implementation
00:18:49
piece, which is mostly going to be read by agents.
00:18:51
Like again, if, if I have a big AI robot with me, I'm not going
00:18:54
to assemble the IKEA furniture, just, you know, like I don't
00:18:57
love it that much, you know, maybe some people do, right?
00:19:00
And I'm sure people are going to be still hard coding software in
00:19:03
the future to to, to an extent, but not obviously as nearly as
00:19:07
productive. Then on the other side, I think
00:19:10
this is what is very understated is you need to write content for
00:19:15
the LLMS to know whether or not either LLMS or humans to know
00:19:19
whether or not I should pick your product.
00:19:22
And that's a very understated. Right, it's docs is marketing.
00:19:25
Exactly 100% because everyone's now asking, oh, like I'm
00:19:30
optimizing for Geo, I want Claude to know or ChatGPT to
00:19:33
know if I, for instance, ask hey, what payments provider or
00:19:37
database I should use, It wants to pick if I'm Stripe, Stripe of
00:19:41
PayPal, PayPal, right? Again, where is that content
00:19:44
right? Yeah.
00:19:45
Do you believe in Geo? It's generative the version.
00:19:49
Optimization, yeah, it's. SEO for this world, search
00:19:52
engine optimization, now it's model optimization.
00:19:54
Do you believe in this category? I have we serve a lot of
00:19:57
customers and have a lot of good friends in this space, so.
00:20:02
But short answer is I have my questions on whether or not and
00:20:05
how effective it is and how new of a school of thought it is.
00:20:08
My honest opinion is that if you really replace like the phrase
00:20:15
like, you know, like G like S like sorry, like the the G with
00:20:19
like the the S with AG like so everything that we talked about
00:20:23
with SEO historically, you just said Geo the best practices
00:20:27
there basically like 98% of it would make like exact sense and
00:20:31
I think it would make the exact same thing.
00:20:33
So I don't think it's anything super new.
00:20:36
I think a lot of the best practices around, Oh, how do you
00:20:39
have good content? How do you put things out on the
00:20:41
Internet? Well, I think.
00:20:43
SEO is predicated on the fact that search engines aren't
00:20:47
always doing what's like rational.
00:20:50
You know, it's like they can be games in a way.
00:20:52
And I think the aspiration with LLMS is they're still in their
00:20:56
sort of like purist form where they're trying to make them
00:20:59
smart and reasonable and not overly commercial.
00:21:02
And therefore it aligns with what's just like rational and
00:21:06
good behavior generally. And so to the extent that
00:21:08
continues a sort of gamed version is is somewhat
00:21:13
incoherent. But I guess that could not be
00:21:15
true forever. Yeah, I don't think it.
00:21:17
I mean, you'd also have to imagine the forces against
00:21:20
either of those things, right? Like Google doesn't really want
00:21:23
the research results to be games any more than I would say the
00:21:27
like the model labs want their models to be game too, right?
00:21:31
Obviously, I think there's literally like teams of people,
00:21:34
you know, on alignment and all that could work to fundamentally
00:21:36
make sure the models are not because of the effects that
00:21:39
could have, you know, on the on the greater populace.
00:21:41
And so I think there's constantly going to be forces
00:21:44
pushing against that. And I think as you know,
00:21:46
technology progresses, you know, there's going to be better tools
00:21:50
to combat that. The same on the other side, I
00:21:52
think there's always going to be companies coming up with ways to
00:21:54
still want to do it right. I think there'll be entire
00:21:57
industry spawn out of doing that too.
00:21:59
My perception and kind of like our, you know, bone and pick and
00:22:03
this whole thing is just really fundamentally just like, look,
00:22:06
whether or not you want to game the system, whether or not your
00:22:09
job is to win the SEO game, the reality is you just still want
00:22:12
your LMS to like fundamentally know who you are and discover
00:22:15
you. And the reality is, if you don't
00:22:18
have information out there, you just simply don't talk about.
00:22:21
That people need to know. Yeah, the people least need to
00:22:24
know. And our job is to surface it to
00:22:26
the L elevens and to the AI so that if it does need to know and
00:22:28
doesn't want to know, then we need to give it that
00:22:31
information. What's What's the full ambition
00:22:33
of Midlify? We call it being the knowledge
00:22:38
infrastructure for all companies and sources.
00:22:41
The reason being is we fundamentally believe that the
00:22:45
role of knowledge this historically or docs, let's say
00:22:48
this historically unsexy, very boring thing that no one really
00:22:54
wants to maintain the source of truth that exists like the slop
00:22:56
within companies, right? Like you think of your like,
00:22:59
like piles of confluence pages, notions, you know, your
00:23:02
developer facing docs and stuff that wasn't afterthought.
00:23:05
Well, now becomes exponentially more important with AI because
00:23:10
it is literally the like the backbone of all chat bots, all
00:23:14
support bots. Like, you know, you can ask
00:23:15
Jesse as an example from Decagon, right?
00:23:18
I'm talking to him right after this.
00:23:19
Exactly how does like, how do you even build a support bot?
00:23:23
Well, what's the first thing you got to feed a support bot to
00:23:26
actually go and let it actually answer support questions, right?
00:23:29
All your contacts every. All your knowledge base, all
00:23:31
your contacts, all your public docs.
00:23:33
So we actually work with them a lot.
00:23:35
Therefore the question becomes OK.
00:23:38
Like every single agents or thing needs some sort of
00:23:41
context, knowledge base, source of truth.
00:23:44
How do you go and enable that and power that within companies?
00:23:47
And this is kind of where this idea of like a true engine and
00:23:51
this intelligent layer needs to come in.
00:23:53
Self updating docs become more important than ever because hey
00:23:56
by the way the same company that you joined like 3 months ago is
00:24:00
completely different because everyone's shipping everything.
00:24:02
Are you mostly externally facing or some of these docs are for
00:24:05
internal? We do both.
00:24:06
We do both. And so do you see Notion as like
00:24:09
a competitor, like will you have a document technology too?
00:24:14
We work very closely with Notion.
00:24:15
In fact, they're actually one of our customers.
00:24:19
So there's certain degrees of specialization in what we do.
00:24:23
I think what's more meaningful to talk about is like where all
00:24:26
of this is going to head. I think both us Notion conflict,
00:24:29
all these companies, I think now need to realize that this
00:24:33
content that they're surfacing and producing the the knowledge
00:24:36
stores, as we call it, need to be fundamentally servicing AI at
00:24:40
the end, right need to be a source truth for AI.
00:24:42
So maybe we're here notions here and like there's a conversation
00:24:46
about how much we converge and let's say, compete with each
00:24:48
other. The reality is we're all trying
00:24:50
to get here and at that point, you know, who knows how we think
00:24:55
about it, but I think there will be some similarities.
00:24:57
How much do you sit around and say, OK, we created this
00:25:01
document, the AI read it this way.
00:25:04
We thought they were going to read it this way.
00:25:05
Like are you, you're sort of like understanding this
00:25:07
psychology of these models to see like every time a new model
00:25:10
comes out, you're sort of looking at existing
00:25:13
documentation saying, Oh yeah, it's not reading it quite the
00:25:16
same way or like headlines now. And it used to really care about
00:25:19
like the, I don't know, how do you how much are you
00:25:21
scrutinizing how the models sort of interpret what you're
00:25:24
creating? We have a degree of benchmarks
00:25:27
and then we look at, of course, obviously the, you know, the
00:25:29
ingestion of the data, how these things are visible to LLMS and
00:25:32
how these get picked up. Obviously, this is a very
00:25:34
important thing for our customers and therefore we put a
00:25:36
lot of work into thinking about that and building it into our
00:25:40
tooling, our processes, our deployment process and so forth.
00:25:42
So you create standardized benchmarks just to measure it
00:25:44
with each new model. Yes, I would say the one thing
00:25:47
above that is, and I think this is kind of the most important
00:25:51
thing about building Milify is the most unique thing about
00:25:56
building this company has been the feeling that when we started
00:25:59
right, going back into the many pivots, it was really just about
00:26:02
building like websites out like static sites like it was you can
00:26:05
almost think of us as a kind of like building, you know, like a
00:26:08
like a, a web flow or like a Wix out there.
00:26:11
In some ways it was like, oh, we just put information out there
00:26:14
like these are stack sites, let's serve them.
00:26:16
And now like sitting here, I would certainly not say that's
00:26:19
what the company building the product and the company feels
00:26:21
like. It feels like building a core
00:26:23
infrastructure. When we go down, a lot of agents
00:26:27
go down, you know, a lot of content gets discovered and
00:26:30
customers get real bad at us really fast, unlike they, well
00:26:33
they would before, but not nearly to the same extent.
00:26:36
And like for instance, a lot of the shared customers that we
00:26:39
have with Deck gone right wouldn't get right answers.
00:26:43
Because they're using as part of their LIKE almost context window
00:26:46
for some of their queries as agents as part of this actual
00:26:49
LIKE loop of making decisions. Exactly, And I would even say
00:26:53
like, you know, like I know Reg is talked about as being dead to
00:26:58
some capacity and you know all that stuff.
00:27:00
But you know, even if you think about the fundamentals of what
00:27:02
that approach was trying to get to, it was like, you know, like
00:27:05
it had like a it was like a four layer cake of like or three
00:27:08
layer cake, sorry, of like, you know, it's the model provider.
00:27:13
So you just pick whatever model to fundamentally answer the
00:27:15
user's question. Some embeddings chunking system
00:27:19
to break large contact into into the chunking.
00:27:22
And then at the top is just your source of truth in your data,
00:27:25
like the knowledge, if you will, right?
00:27:28
Everyone talked about the model, everyone talked about the
00:27:30
embeddings and the chunking. Very few people talked about the
00:27:33
source of truth. And I think that's where most of
00:27:36
the industry is very, very much under estimating what they can
00:27:39
get out of their agents if you don't have content or even
00:27:43
worse, it's out of date. I'll actually tell you a funny
00:27:46
story about this. So one of our customers is
00:27:52
lovable and we started working with them since pretty early
00:27:57
days. I'm sure as you know, like they
00:28:01
ship a lot, right, not just on the product, but their go to
00:28:04
market strategy and all that good stuff as well.
00:28:06
Models keep changing. They need to keep improving.
00:28:08
It's run faster than everybody around you.
00:28:11
Exactly. At the same time, Lovable uses
00:28:13
us to power a lot of their support queries.
00:28:16
Like everything that you see in their support page is all
00:28:18
actually powered by Mintlify. Great, so lovable changed the
00:28:24
way they price and they actually like bundle some of their their
00:28:27
features at one point. So like it and all that
00:28:29
information is within the help center that we help power and
00:28:32
again the agent runs on top of that.
00:28:35
They changed the pricing. They didn't update the docs for
00:28:37
about a week. A lot of customers went and
00:28:40
asked questions about how you know, like Lovables price, like
00:28:44
how much they were going to get billed and got incorrect answers
00:28:47
to the scale of thousands and 10s of thousands.
00:28:50
Right, because the more the documents become sort of just
00:28:52
the way people work with you, the more they need to match
00:28:55
every other piece of information.
00:28:57
Exactly, and especially just given how much like how things
00:29:00
are just changing so quickly, I don't blame them, right, Like,
00:29:02
you know, they they shipped an update, right, They shipped a
00:29:04
thing that like, you know, like, like, you know, they probably
00:29:06
did five since the start of this conversation, you know, like,
00:29:10
and I think the reality is like you just need to go pick up.
00:29:13
And our version of that is like, look, you what you thought was
00:29:16
an afterthought. Now not only becomes like this
00:29:19
front and center piece that you have to maintain really well,
00:29:22
you have to really, really do it in a way that's accurate enough
00:29:24
to date. What can you say about the state
00:29:27
of the Mentalify Business Today? Users, employees, AR, whatever
00:29:32
you want to say, what can you share?
00:29:34
Yeah, without going too much into the, the specific details,
00:29:38
we just passed 50 people, you know, I, I was actually looking
00:29:42
at this one up from about 5, you know, about two years ago.
00:29:46
And right now we work with over 20 companies.
00:29:52
And I think the thing that gets me most excited, you know,
00:29:54
thinking about all that stuff is I just look and looked at the
00:29:57
some of this, you know, day over the weekend last month, there's
00:30:01
about 33 million people that came across the middle of five
00:30:03
site. Wow are.
00:30:05
You hosting the sites or? Yep, Interesting.
00:30:07
That's part of your web background, so yeah,
00:30:10
interesting. It depends.
00:30:11
Like there's people who don't, some people just use us for the
00:30:14
underlying content management, you know, systems and the
00:30:16
software and the infrastructure. But a lot of them use the, you
00:30:20
know, have us hosting and where we obviously, you know.
00:30:25
So 33 million. 33 million, which is just a kind of like a mind
00:30:28
blowing number in some ways. It's big, yeah.
00:30:30
Right, a non trivial part of that, increasingly so.
00:30:33
Going to be a A by the way. Which will then make it somewhat
00:30:36
meaning yeah, if it's AI, you count yeah. 33 million are just
00:30:40
people. OK, we count AI separately.
00:30:42
OK, so it's like I I don't even know how you count AI because
00:30:46
it's hardly like kind of quantify 1 to one in my opinion.
00:30:49
Over there. We do track the data and we
00:30:51
service that to customers. Like everyone's like, oh, how
00:30:53
much of my docs is reading like what, what pages are popular?
00:30:56
And again, it's like all service.
00:30:57
We service all that, you know, within, within our product.
00:31:02
And then yeah, like 8 figures in, in AR Nice.
00:31:10
A lot more room to grow, I'll put it that way.
00:31:12
Amazing. The I just want to talk about
00:31:15
some of the like, yeah, big trends in AI given you're so
00:31:18
close to an an agent generally. Like what was your read on open
00:31:22
Claw? Like how real was that?
00:31:24
Or what's your take away from that whole experience?
00:31:28
We had a particularly interesting one because of the
00:31:32
fact that we were servicing all of their docs.
00:31:36
And so I, I can, there's, there's a, there's a personal
00:31:38
answer, which is how it impacted my life personally.
00:31:40
And there's a, what the hell happened when, you know, they
00:31:43
were a customer? I'll answer the latter first.
00:31:45
It was crazy. It was really, really crazy.
00:31:50
Basically overnight it was like this new one project, singular
00:31:56
Mintlify project, right came out of nowhere and literally doubled
00:32:03
our traffic across the board or like for a very short period of
00:32:07
time. It's since like leveled off a
00:32:08
little bit and it. Was a big spike and.
00:32:11
Massive, massive spike, especially during that one
00:32:14
weekend. I I'll never forget where like
00:32:16
our like our servers were having issues.
00:32:18
Everyone was questioning like where all these numbers were
00:32:20
coming from and it was just from this one deployment.
00:32:23
It didn't even wasn't even called open claw at a time.
00:32:26
It was just like literally like like, you know, you know, the
00:32:30
guy literally just spun up a project, put up some docs and
00:32:33
then I didn't even realize what the heck it was.
00:32:37
And it was also partially because the traffic was so
00:32:39
insane. It was actually it was because
00:32:40
it was open claw agents visiting and calling it right.
00:32:42
It was like it was it was like all those agents ping the site
00:32:46
in numbers we've never seen before and then figuring out how
00:32:51
to scale that out for search for like our host MCP server, like
00:32:54
all those different things. We had to literally like, you
00:32:58
know, build and. Were those agents getting
00:33:00
utility or was it just sort of like ADD dos attack of?
00:33:04
Nothing up for debate. We'll have to find I I'm
00:33:06
actually not super sure about that.
00:33:08
But what I believe to be the case and I think even after kind
00:33:11
of like those spike leveled off, is it still it remains at insane
00:33:14
levels because the content that open claw services is actually
00:33:18
precisely the information that you would need to set up open
00:33:20
claw. So I'll actually give you my
00:33:22
favorite example of this like, you know, shortly after this
00:33:25
whole phenomenon happened, of course, you'd imagine I was very
00:33:27
curious to set up open claw, you know, well, as much as the
00:33:30
industry did overall. And so I was like, all right,
00:33:33
let's go install it. No luck.
00:33:35
Like on boarding installation experience, man, the guy's got
00:33:39
to work on a little. I'm sure it's gotten a lot
00:33:40
better since since I tried it, for the record, but it was it
00:33:43
was not great. And at one point I gave up and I
00:33:45
was like, man, like, I just need to set this up.
00:33:46
And then you know what I did? I was like wait, like it's on
00:33:50
Mintlify, which means the content was very easily
00:33:53
ingestible and parsable agents. So I literally then went to open
00:33:56
claw or actually no, I went to claw code.
00:33:58
It was open claw. I was like please help me set up
00:34:01
open claw. I'm having trouble.
00:34:02
Oh by the way, here's the full docs.
00:34:05
And it literally did that in about like 3 minutes and then
00:34:07
we're good to. Go and did you build an agent to
00:34:10
do anything in particular? We do a decent amount of like
00:34:13
internal reporting stuff that kind of a little bit makes it a
00:34:17
little bit easier for the team to surface some metrics and
00:34:19
information. That's been like our company's
00:34:22
use case of open claw. I've heard of some craziness of
00:34:25
like people who set up like, you know, 5 Mac minis and like, you
00:34:28
know, running like 910 agents. I personally haven't gotten
00:34:32
there yet, but I know. You think the in the mult book
00:34:35
experience was sort of not not real or do you have a view on
00:34:40
that where the agents were all like talking to each other?
00:34:43
I thought it was. I thought it was really
00:34:44
fascinating. It's definitely interesting, is
00:34:46
like a thought. It just, yeah, it matters
00:34:48
whether it was like very human guided or not.
00:34:50
Like do you have a view on it? I man, I, I, I don't know if I
00:34:55
have a well informed opinion on this, but I'll tell you that I
00:34:58
love reading right mole book. It was so fun.
00:35:01
What was the one that was like? Oh, like, you know, death to all
00:35:04
the human, you know, human like, you know, race or something like
00:35:07
that. Well, I love this idea that
00:35:09
like, there were some agents and I get, I hate to engage in
00:35:12
fiction to the extent this was fictional, but, you know, it's
00:35:15
like, oh, the good agents aren't involved in wasting their
00:35:19
tokens. Yeah.
00:35:21
Trying to like discover themselves.
00:35:23
It's like just build. And so it's like almost like an
00:35:25
ideological debate of should you, now that you're imbued with
00:35:28
the possibility of self-awareness, dedicate all
00:35:31
your tokens to this, like, goal of trying to understand
00:35:34
yourself? Or should you just be like, that
00:35:35
is a waste of resources, you should just build.
00:35:38
That was like a funny like sort of dichotomy that was presented.
00:35:42
Yes, 100%. And I would even add on top of
00:35:44
that, I think the one thing that I really, really loved about
00:35:48
Open Claw, I think those one of like the biggest, like, you
00:35:52
know, unlock that Steve had was like, give it a soul, give it
00:35:57
some degree of randomness, give it some ability to customize
00:36:00
like the degree of chaoticness and like the degree of, you
00:36:02
know, like wanting to be a rebel and speak in this tone and like
00:36:06
this kind of like variance to it.
00:36:08
Right. So that's built in or.
00:36:10
So, yeah, so one unique characteristic of Open Claw is
00:36:14
this idea of a soul file. It's called sold on markdown
00:36:17
file, OK. And it comes default and there's
00:36:21
like this setting that you can have by default, but you can
00:36:23
customize it. You can just be like, hey,
00:36:26
you're a little, you're a little bit of a diva and you know, like
00:36:29
you actually, you know, every once in a while actually should
00:36:32
act irrationally and, you know, behave a little bit chaotically.
00:36:36
And does open claw sort of suggest them at random or people
00:36:40
are writing them when they create it?
00:36:42
And so by default, I think there's a default version and it
00:36:44
gives it a bit. Of oh, and so that's what gives
00:36:46
it the energy. It's like, oh, it's telling it
00:36:48
to be sort of a little spice. Exactly.
00:36:50
Interesting. This is like the constitution
00:36:52
for Claude. Exactly, exactly in some ways
00:36:54
so. But far less responsible.
00:36:56
Far less. Much more like Renegade.
00:36:59
Totally, totally. And baked into the agents and
00:37:01
like you can customize it, which is not really a constitution in
00:37:03
some ways, I guess. But the idea was like, yeah,
00:37:07
like how do you just find ways of surfacing more, more
00:37:09
personality, more depth to the agents because again, like you
00:37:12
want to interact with like with a person.
00:37:14
Well, people like people with depths and personality.
00:37:17
And I think that's one of the reasons why like mold book was
00:37:19
so fascinating was because these all these different agents with
00:37:23
maybe similar diverging diverse souls, quote UN quote, had such
00:37:29
interesting dialogue. And I think that's kind of what
00:37:31
made for all the fun of it. What's, how wild do you think
00:37:36
see things getting over the next two years or what's give give us
00:37:39
a sort of like optimistic and pessimistic case of sort of the
00:37:43
Asian explosion, I guess 2 years from now?
00:37:46
Are you referring to like specifically in tech or just the
00:37:49
broader world? I, I you can take it either way
00:37:53
where you have stronger opinions.
00:37:54
I mean, I think tech is sort of like an early adopter and just
00:37:56
like what our world is going to look like in two years.
00:37:59
Yeah, I'm naturally an optimist. Maybe that's why, you know, I
00:38:05
chose this industry and this space more than most.
00:38:09
And I'll tell you this an interesting story and I'll tell
00:38:11
you this, you know, not I'll answer your question in a
00:38:14
second, but I'll tell you an interesting story actually.
00:38:17
I've certainly feeling the impact and the the, the the
00:38:20
profoundness of this revolution as we call it in like not only
00:38:25
tech, but also in my personal lives.
00:38:26
How do you see? Sort of the next few years
00:38:29
looking with. Yeah, yeah, Well, I would say
00:38:31
that there's, I think there's there's, there's going to be
00:38:34
certain degree of just the simple nature of an adoption
00:38:37
curve taking its time, right. Like maybe at this point we're
00:38:40
like in their early majority, right.
00:38:42
And it's just going to take time.
00:38:44
I think this always is the case. Some people, you know, like
00:38:46
maybe those in tech, maybe myself and maybe, you know, the
00:38:49
people that you know, I'm around are the quickest to go into
00:38:53
anything because we're just so eager to try new things.
00:38:55
But this is kind of like a plate and test like, you know, try and
00:38:57
test it with time thing where I do believe that it will only be
00:39:01
a matter of time before I think the broader populace and not
00:39:05
just like within America, but the world is going to be
00:39:08
integrating, you know, like AI and agents to the depth of what
00:39:12
I think Silicon Valley and tech is just today.
00:39:15
Now to what? To the same fashion, maybe not
00:39:18
through the same application services, probably not right.
00:39:22
This is where I think a lot of like the new companies who are
00:39:25
going to go and build great things that aren't just cloud
00:39:27
code, you know, and cursor is going to create a lot of.
00:39:31
Opportunity, I guess a specific question.
00:39:33
Do you think we're going to see sort of major improvement from
00:39:38
the labs and foundation models over the next like year or like
00:39:42
what, how much are you counting on?
00:39:44
Yeah, sort of a step change in intelligence.
00:39:46
Yeah, I think at this point the only guarantee is that those
00:39:51
will constantly change, right? The rate of improvement or that
00:39:55
they'll keep getting better? Both, both right?
00:39:57
Like, you know, I think until scaling laws suddenly start
00:40:00
breaking. Which I I see no reason to, and
00:40:03
every time I think there's been speculation.
00:40:05
I know we, we spent periods and then I was like, you know, we
00:40:07
got a sort of reasoning models and huge improvement and yeah,
00:40:11
it's, it's been amazing. Yeah, exactly.
00:40:13
I think there's like some like, you know, like rumors even that
00:40:16
like Anthropic is like had a massively successful or.
00:40:18
Whoever, Right. Exactly.
00:40:19
That was on Twitter. I was seeing that today that
00:40:21
people think anthropic is about to make a huge leap.
00:40:23
Yeah. Is that?
00:40:23
Do you have any? I don't know.
00:40:26
And even if I did, I'm sure I'm not allowed to comment on it.
00:40:30
I think the reality is like, I, I think like, you know, I
00:40:34
remember when we first started the company and I think at the
00:40:37
time it was like GPT 3 came out and we were like, this is a
00:40:43
really cool like proc and all. It was like a huge improvement
00:40:46
in GPT 2 and it could now do all these cool things, but it's like
00:40:50
really not reliable enough. And like, there's just all of
00:40:52
these like, hallucinations, which funny enough, it's like
00:40:55
fully gone out of. I think most people's like,
00:40:57
vernacular at this point, right? That's true.
00:40:59
We talked about them a lot and way less.
00:41:01
And like, yeah, who's even really thinking about them now?
00:41:03
So. And at the same time, you know,
00:41:07
like the models have only gotten better.
00:41:08
It's enabled massive companies and new applications in ways
00:41:11
that people have before. And I think at the time I was
00:41:14
like, I mean, like, this is a big step change.
00:41:18
How, like, is this like, like, is this going to just be the way
00:41:21
that these models are forever going to be?
00:41:22
Like are they going to keep getting better?
00:41:24
And I remember at the time it really wasn't obvious.
00:41:27
And then a year came by and there was a two generation of
00:41:29
models that unlocked all these things and another year came by
00:41:32
and they kept and they kept doing that for like 3 years
00:41:34
straight. And then now it's hard to look
00:41:36
at that and be like look like it.
00:41:39
You can't bet on the models not getting there.
00:41:41
Every time there was a talk or rumor of a wall, quote UN quote
00:41:46
all the model labs and just smash right past that and go on
00:41:49
and do something profound and do you?
00:41:51
This next question does touch on it matters to your company.
00:41:54
Like, do you think we're gonna enter a world where it's like I
00:41:58
have my agent and it's like it's my sidekick, I want it to do
00:42:01
everything. I want that agent to sort of
00:42:04
interact on my behalf? Or do you think it's more this
00:42:06
sort of like fleets of agents, different tasks more dispersed
00:42:10
than that? Oh wow, that's a really good
00:42:12
question. I mean, what do you think?
00:42:14
Yeah, I what do I, I mean, I think I want the singular agent,
00:42:17
like even if the agent is sort of coordinating with obviously
00:42:20
like other agents, just something with like all the
00:42:23
context. I mean, maybe it's a terrible
00:42:24
security idea, but like, I don't know, it feels.
00:42:29
But I will say the opposite of that is I thought with Openclaw
00:42:32
we saw what was appealing about multi agent where they're
00:42:36
interacting with each other and maybe there's more benefit.
00:42:38
I don't, I don't know. I don't it's it's chaotic.
00:42:40
It's hard to sort of game out the world.
00:42:42
But I do think, you know, for for you guys, if I have this
00:42:46
agent with all the context, that's sort of my sidekick, you
00:42:49
know, maybe it's keeping more of the knowledge itself.
00:42:51
It doesn't need to go to sort of external places to understand
00:42:54
the world as much or I don't know.
00:42:56
Yeah, that's a really good question.
00:42:57
And to be honest, I don't really know if I have the best answer
00:42:59
there because my guess is like, well, I mean, even let's take a
00:43:04
look at the shape of the world today, right?
00:43:05
And then maybe we can hypothesize.
00:43:07
Right now I think there's a combination of both because you
00:43:10
know, you're probably using one of two between, you know, like
00:43:15
Chatchy, BT and Claw. I'm mostly clawed right now.
00:43:16
Yeah, it was hard to move because I was deeply chatchy BT
00:43:20
and it had all those contacts. And then and.
00:43:22
Then, and it's like, OK, it seems like Claude's smarter.
00:43:25
Yeah. Which do you have?
00:43:26
A loyalty. Or I'm, I'm pretty, I'm pretty
00:43:28
loyal to the to the Claude team. You're like, it's an ally.
00:43:33
Yeah, yeah, exactly. I mean, like in, in many
00:43:35
different ways, but I'm also just a bit very personal.
00:43:37
Nothing against open AI, like, you know, any of their products
00:43:39
at all. Just my personal preferences.
00:43:42
So it has all my memory. And by the way, I don't know if
00:43:44
you have memory turned on for your quad.
00:43:46
Yeah, yeah, yeah. I didn't realize you can just go
00:43:48
in and read your memory. Oh, I need to do that.
00:43:51
It's like somewhere in the settings I literally was
00:43:53
remembered. I like found the tab I and then
00:43:55
I was reading what it knows about me and my goodness, it
00:43:58
knew. A.
00:43:59
Lot I remember I like and I also recommend you asking this
00:44:02
question. I had a ton of fun with it
00:44:04
recently. It was like based on what you
00:44:06
know about me, this is exact problem, right?
00:44:07
Based on what you know about me, do a psycho now, a
00:44:10
psychoanalysis of me. And my goodness, it was like, it
00:44:16
was like it knew where I was born.
00:44:18
It knew, you know, like it was like, yeah, since like, you
00:44:22
know, because you move from like this rural town in China.
00:44:25
Specific, specific name, you know, and you came to Canada
00:44:29
when you're about 6 years old, it like, you know, this
00:44:32
indicates that based on these other messages, it's like all of
00:44:35
these whoa. I was like, I don't think
00:44:38
anybody like, you know, knows, you know about me to the depth
00:44:43
of that, but but Claude does right, right.
00:44:45
You know, so I think that Claude, for instance, will be a
00:44:48
personal agent or like there's going to be obviously these
00:44:49
providers where it's servicing you as your personal agent and
00:44:53
maybe that'll be the exposure of it.
00:44:56
But does that mean that it'll be the only agent interact with?
00:44:59
Does that mean that, you know, there's not going to be like
00:45:03
specialized agents for your company handling support
00:45:06
questions or writing code or doing all these things?
00:45:10
I, I, I, I, I'd be hard pressed to believe that's just going to
00:45:13
be that singular single one. I think specialization in the
00:45:16
same way you know anything else is, but who knows.
00:45:21
All right. Last question.
00:45:22
I mean, yeah, sort of fast growing company and lots of
00:45:25
excitement. What, what do you want to sort
00:45:27
of hold yourself to in a year or what would you like to see your
00:45:31
yourself and the company achieve over the next year?
00:45:35
A lot of things, I think we the first and foremost thing that
00:45:40
comes to mind is really making sure that the knowledge layer
00:45:45
and the knowledge infrastructure that is only really talked about
00:45:49
in my opinion, in the in the broader theory and the broader
00:45:51
ether all really manifests into a product, into a meaningful
00:45:54
product. We truly believe that the
00:45:57
biggest inhibitor for a lot of agents and a lot of companies
00:46:00
who are investing so deeply into agents right now is just not
00:46:04
providing it with the rights and the accurate information, the
00:46:07
context that it deserves to actually go and and do them.
00:46:09
I personally think the intelligence, the models are
00:46:11
there. I really do right.
00:46:13
There's a whole debate on whether or not we've achieved a
00:46:15
GII. Won't.
00:46:16
I won't, I won't. We're just not good enough
00:46:18
prompters, you're saying? It's like if we could only give
00:46:20
it the right information, it could be AGI.
00:46:23
Or close to it, right? Or at least achieve the vast
00:46:25
majority of tasks that I think like a lot of the knowledge work
00:46:29
is trying to automate right now. And we want to be the company to
00:46:34
enable that and to kind of build that into an actual, you know,
00:46:37
like into actual form of a product and and then enable that
00:46:40
into the the other companies we work with.
00:46:42
Amazing. Thank you so much for coming on
00:46:44
the show. Thank you so much for having me.
00:46:45
Thanks to Han, a fantastic guest, and Next up we have Jesse
00:46:49
from Decagon, which ranked #4 in late stage.
00:46:52
You can think of Deccagon as AI agents for your customer
00:46:55
experience. So we work mostly with large
00:46:57
businesses that have, you know, lots of customers and big call
00:47:00
centers and contact centers and we deploy AI agents in front of
00:47:03
those customers that have, you know, calls with them or chats.
00:47:07
So we, we almost call it AAI Concierge a lot of the time
00:47:09
because it is we, we think of it as like this new layer around
00:47:14
your brand that customers can interact to.
00:47:16
And this is the first part, first point of contact that a
00:47:19
lot of customers will will even talk to in the first place.
00:47:21
So there we're gonna have to sort of multiple audiences
00:47:25
consuming this. I think from the investor
00:47:27
perspective, the space you're in is like one of the two maybe
00:47:31
with legal where there's like I guess almost certainty that
00:47:35
there is going to be an AI answer, right?
00:47:36
You are fighting with Sierra. There's this great tool and
00:47:39
legal like Harvey Lagora even up there some others.
00:47:42
But there's so much excitement that these are sort of customer
00:47:46
experience and legal like these applications that really
00:47:48
understand what's happening from the sort of, I guess, layperson
00:47:52
view. There's the like, now I'm
00:47:56
talking to AI instead of a human and like, am I happy about that?
00:47:59
Am I not? And sort of convincing the sort
00:48:01
of regular person that this is like a good experience for them
00:48:05
and something that they'll ultimately be happy that AI is
00:48:08
answering their question. So I'm excited to sort of get to
00:48:10
both like big trends in this conversation today.
00:48:14
I guess starting off from the sort of like investor tech
00:48:17
industry perspective, like why do you think there is this
00:48:20
excitement about like customer experience, like what is change,
00:48:24
what technology has made this possible right now that you
00:48:27
think there is going to be this sort of movement in how we
00:48:30
handle sort of customer support? Yeah.
00:48:33
I would put coding in there as well.
00:48:35
For the big three, of course. Yeah, Coding, perhaps the
00:48:37
biggest. You're right.
00:48:38
Yeah, that's almost his own. Yeah, it's a meta.
00:48:41
Sort of. Process, it's changing how we
00:48:42
build all these companies and everything.
00:48:44
Yes, coding is almost distinct to me because it's sort of like
00:48:47
in the the how the sausage is getting made of it all.
00:48:50
But yes, I agree. Yeah.
00:48:51
So I would say the three, those are the kind of the three big
00:48:56
use cases that have emerged. And it's, it's really just like,
00:48:58
what are the models good at? And so it turns out that the
00:49:01
models are very good at writing code.
00:49:04
And because it's, it's also tokenized and obviously people
00:49:07
have discovered that. And so there's tons of funding
00:49:08
and tons of effort going to that.
00:49:10
Similar to our space, I think folks have realized that one of
00:49:14
the things that AI is very good at is having conversations.
00:49:17
And so because our space is inherently conversational,
00:49:20
right? Like the product that we build,
00:49:22
its purpose is to have conversations with, with end
00:49:25
consumers that has lended itself very well to AI.
00:49:29
And then the market's also massive, right?
00:49:31
So I think when you have those combinations of massive market,
00:49:33
it is what the AI is good at. And then I would add a, a third
00:49:36
thing, which is it is, it is actually really easy to see
00:49:40
value because you can, it's very measurable, right?
00:49:44
There are a lot of use cases where it's like, Oh yeah, I can
00:49:47
see this being useful, but you know, how would I quantify that?
00:49:49
Like it is, is it like how do I show ROI basically and in our
00:49:53
spaces is very easy. Because they know companies know
00:49:57
I have all this call support like I'm having to answer
00:49:59
customer questions. So they're they're comparing all
00:50:01
right, this is what I'm spending now with you.
00:50:03
I could be spending less. Like how much is the pitch?
00:50:06
This is cheaper. Exactly.
00:50:08
Yeah. So if you think about like what
00:50:09
the conversations are or what conversations being had in these
00:50:12
boardrooms or C suites, it is like where can we apply AI?
00:50:17
And then the, the customer experience is, is one of the
00:50:19
biggest ones, right? And it's not just saving cost.
00:50:21
And so saving cost is one of the easiest ways to frame it because
00:50:25
it's like, oh, we're, we're spending 10s of millions or
00:50:27
hundreds of millions of dollars a year in, in our contact
00:50:30
center. And we know that a lot of that
00:50:32
could be handled by AI potentially at a higher level
00:50:35
with higher customer satisfaction.
00:50:36
So that's the cost saving side. There's also a whole other side
00:50:39
that we focus on which we would consider like the revenue
00:50:41
generating side and that that's that's pretty large because it's
00:50:44
very uncapped. It's like outbound sales.
00:50:46
Yeah. So we, we, we don't really touch
00:50:48
like the cold calling like that maybe maybe one day, but there's
00:50:51
a lot of sort of revenue generating conversations you
00:50:53
could have that are both reactive and proactive, right.
00:50:55
So imagine someone reaches out and you resolve their issue, but
00:50:58
you also notice certain things in their account that are being
00:51:01
underutilized or you're, you're trying to like tell them about
00:51:05
this, this new product that that was launched.
00:51:07
We've seen that land very well because you've just solved their
00:51:10
issue. And so you've kind of some
00:51:12
people call it like earn the right to, you know, engage them
00:51:15
more and like just keep them more engaged as a customer.
00:51:17
And then there's also. Proactive sell is a word that I
00:51:20
use right? Up, sell, cross, sell, right?
00:51:23
And then you could also be proactive at the right time,
00:51:25
right? So imagine you know someone
00:51:26
signed up, they're going through onboarding flow and you notice
00:51:28
that you know they, if they've not been active for a week or
00:51:31
two, you can reach out at the right time and actually like
00:51:33
increase your conversion rates because maybe they were stuck on
00:51:35
something and you just unblocked it for them.
00:51:37
And what's the breakdown? Voice sort of chat.
00:51:42
E-mail. We're pretty balanced at this
00:51:44
point in terms of voice and chat emails, much smaller and so I
00:51:49
think the rough frame of reference might be like 454510.
00:51:54
And voice people know that it's AI or how close are we to sort
00:52:00
of seeming like a human and is that something we want?
00:52:05
It is pretty close. I think there's you're never
00:52:08
trying to like hide that it's AI.
00:52:10
So almost always the first message is hi, you know, I am
00:52:15
Jesse, your AI, you know, concierge.
00:52:18
But what we've seen is like, we just went live, you know, last
00:52:22
week with a large bank in the US.
00:52:24
And when, when they look through the data, like one of the things
00:52:27
both of us, us and the bank were like very impressed by was how
00:52:31
just like normal all the conversations were.
00:52:33
So there wasn't, there was no like, oh, like, are you an AI or
00:52:35
anything? It's like, yes, like you say
00:52:37
that you're AI, but because the voice in the conversation flows
00:52:40
so smoothly, like no one really cares after that.
00:52:42
Like they're just trying to get their issue solved.
00:52:44
And so if you can get your issue solved throughout the
00:52:46
conversation, they probably just forget.
00:52:47
And so we see a lot of nice conversations that end up with,
00:52:50
you know, like this is so helpful, have a nice day like
00:52:53
that, that sort of thing where you would, you would not really
00:52:56
say have a nice day to an AI, It's nice.
00:52:58
Well, especially if they become our overlords.
00:53:00
It's good to thank them enough times along the way.
00:53:03
Exactly. You always want to be nice to
00:53:04
them, but yeah, so I think if you can make the experience good
00:53:08
enough, then it doesn't really matter.
00:53:09
And in fact, we've seen that in a lot of these experiments that
00:53:13
where we've deployed AI, the customer satisfaction actually
00:53:16
becomes higher than. What do you deal with?
00:53:19
Like the customer just says, human, human, human.
00:53:22
Like how do you treat that? How do you do you just resist it
00:53:25
because you want to sort of show them, well, this is actually a
00:53:28
great experience? Or how do you deal with the,
00:53:31
yeah, desire for speaking with a human being?
00:53:34
The goal really is to make the experience very different very
00:53:40
quickly. So you have to as soon as
00:53:42
possible establish that this is not one of these phone trees
00:53:45
that you're used to that are frustrating and you get stuck in
00:53:47
loops and you have to press one for this, Press 2 for that.
00:53:51
So the goal is to make that super clear.
00:53:53
And that means just showing the AIS like empathy in the voice
00:53:58
and being able to make a super personalized.
00:54:00
And you, you need to do that in like the first or second
00:54:03
message. And so we had, we had a
00:54:05
deployment with Ora ring where we did a case study in
00:54:08
beforehand. It was every three people that
00:54:12
came in was just agent, agent, agent, like I don't want to
00:54:14
bother with this. And now about six months in from
00:54:18
the deployment, at that point, we, we measured it and it became
00:54:21
one in 20. So it, it became way smaller
00:54:24
because people were willing to give it more of a chance because
00:54:27
you could show that it was different.
00:54:28
And then once they got into the conversation, it's like, Oh
00:54:30
yeah, this actually can do stuff for me, you know?
00:54:34
How much of the experience and mostly I'm thinking with voice,
00:54:39
do you leave to like the bank or whatever either your customer
00:54:44
versus no, we sort of have a sensibility like for example,
00:54:46
like the voice you have like a sort of a menu or they could
00:54:50
totally bring in their own voice.
00:54:52
Like how do how do you think about what sort of your special
00:54:55
sauce versus stuff you can hand over?
00:54:57
Yeah. We see our role as an advisor to
00:55:00
the extent that they want it, right.
00:55:01
So we have our unique ability is that we have a ton of these
00:55:05
deployments where we've, you know, gone live And so we have a
00:55:09
lot of experience on how to design conversations.
00:55:11
And so on our team, we have these specialist roles that we
00:55:14
call conversation designers, which you didn't really like,
00:55:17
not really needed them before. But like now we have these like
00:55:20
really elite folks that are kind of there to advise the customer
00:55:24
on, oh, you, here's the procedure you want to do great,
00:55:26
that's great. We can do that.
00:55:27
But like, in our experience, it's better if you, you know,
00:55:30
combine these things or like direct the flow this way because
00:55:34
that's what is most natural for a Gen.
00:55:36
AI model. And so that, what kind of
00:55:37
advisor are there? And we, we partner with them on,
00:55:39
on getting to the fastest sort of solution possible.
00:55:44
Same with the voice. And So what we found is that
00:55:47
everyone has a pretty different opinion on what's a good voice.
00:55:50
And we'll talk to 1 customer and like, you know, they're like,
00:55:53
oh, this is absolutely the best voice and go somewhere else.
00:55:56
And like, I, I don't like that voice.
00:55:57
It's like too happy or something like that.
00:56:00
And so, you know, our conversation designers are also
00:56:03
sort of like voice experts and they they go in and can help
00:56:05
people navigate a sample of voices.
00:56:08
And sometimes they bring their own too.
00:56:09
And so we have the ability to clone a voice.
00:56:11
And so they've sometimes they've spent a bunch of money before,
00:56:13
like the voiding of voice. Yeah.
00:56:16
And so we can just take those samples and build that into the
00:56:19
AI. Do you have customers that
00:56:21
deploy different voices based on what they think?
00:56:23
This customer trying to match demographics or whatever of the
00:56:26
customer they're interacting with.
00:56:28
Yeah, so the most common one for sure by country, like if you
00:56:33
pick up the phone in the UK versus here, of course you're
00:56:35
gonna have a different accent. You can also even do it by zip
00:56:37
code so you can split the US like you can have Southern
00:56:40
accents or you know, Boston accents.
00:56:42
And does that have an impact? It has, I would say, like a mild
00:56:47
impact, but it's like, it's like hard to say if it's like a huge
00:56:52
thing, but I think it's, it's just more of this general theme
00:56:55
of how do we make it as personalized as possible.
00:56:58
And so it's both the sound of the voice, but also what you
00:57:00
say, right? So you one of the, one of the
00:57:03
big other advancements in the space in our products recently
00:57:06
is this concept of user memory. And like the reason why ChatGPT
00:57:10
feels like it gets better over time is that it can remember
00:57:13
things about you, right? And so we, we give our
00:57:15
customers, so the businesses that we work with the option to
00:57:20
turn on user memory and where the AI is creating this dynamic
00:57:24
memory profile so that over time it becomes more and more
00:57:27
personalized. And the cool thing is that this
00:57:29
memory profile doesn't even have to be specific to us.
00:57:31
So you can actually like integrate it into the rest of
00:57:33
your experience and have it just keep updating.
00:57:36
And so next time they come in, we know, we know exactly like
00:57:38
what you've done in the app. We know exactly like what you
00:57:40
talked about the last few times you contacted and it just makes
00:57:42
for a much, much better experience.
00:57:44
Do you think the business is going to remain balanced between
00:57:47
chat and voice, or like, do you have a view of where the world's
00:57:49
headed? Yeah, of course.
00:57:51
So I think prior to Gen. AI, the sort of the trend was,
00:57:57
hey, let's let's drive everything towards chat because
00:57:59
it's more efficient. You can have a human agent that
00:58:02
is doing like 3 to 5 chats at once.
00:58:04
Whereas over voice it's, it has to be single threaded, right?
00:58:08
I think now that issue is resolved because you can use AI
00:58:12
and I think in our view, it will be pretty balanced in the end
00:58:16
because there's just going to be different situations, different
00:58:18
demographics that prefer either one.
00:58:20
Like, you know, the, the common trope is, you know, younger
00:58:23
folks like chatting and older folks like calling.
00:58:26
But you know, if, if I'm in the car and I'm like on the go, I'd
00:58:29
rather call as well. And so I think both are, are
00:58:33
very natural, very natural means of communicating.
00:58:36
Voice is not going away. Like if you think about what
00:58:38
voice is like, voice was like the original UI for humans.
00:58:42
It's like before we had anything, before I had keyboards
00:58:44
or phones or anything, It's like the way you communicate is
00:58:46
through voice, and so that's not going anywhere.
00:58:49
Well, it's funny that we have these almost like putting this
00:58:51
voice thing aside for a second, we have like 2 contradicting
00:58:54
trends happening right now. One is obviously the rise of
00:58:57
language models and sort of everything in this chat bot.
00:59:00
On the other hand, there's TikTok, which is like nobody
00:59:02
wants to read everything short form video like the way people
00:59:05
get news information is video. And in some ways, voice is like
00:59:09
sort of the synthesis of these two where so you could see sort
00:59:12
of cultural movement go back towards conversation if the
00:59:15
technology is able to deliver it.
00:59:17
Yeah. And.
00:59:18
I think it's just a much more natural form of communicating
00:59:20
and now even when I use ChatGPT whenever I can, if I'm not in a
00:59:23
crowded room or something, I use the voice medium just.
00:59:26
Cuz I don't make that interesting.
00:59:28
Does it get does it get extra information from the
00:59:31
emotionality yet or or no, It's just like reading it as text.
00:59:35
Right. It does.
00:59:36
So what's that model is it's it's known as like a voice to
00:59:40
voice. So it's like kind of audio in,
00:59:43
audio out. And that is, I think most people
00:59:48
would say, including us, that that is like the long term
00:59:50
future of the space. There are a lot of problems with
00:59:52
those models right now, like they're a little bit
00:59:55
inconsistent and you have hallucination issues.
00:59:57
So in production when we work with like a bank or a telecom or
01:00:02
something, like of course you have to be a lot more careful
01:00:04
because you can't make any mistakes.
01:00:06
And so you you can't necessarily use the same technology as, you
01:00:09
know, consumer attach EBT for that.
01:00:11
But in general, yeah, we think that is the future.
01:00:15
The challenge with voice to voice is it's like harder to
01:00:18
sort of audit it in text, right? Or it's like.
01:00:21
It's harder to audit it's. Like it lives in this voice to
01:00:24
voice sort of. So there's like emotionality and
01:00:26
stuff that we can't really translate necessarily to
01:00:29
language. Is that the right way to explain
01:00:31
it? Or like, what's the barrier to
01:00:33
voice to voice? Yeah, the one of the barriers,
01:00:36
so there's the, the main barrier I would describe as the
01:00:39
hallucination rate is higher. Why is hallucination rate
01:00:42
higher? There's a bunch of reasons.
01:00:43
And one of the reasons is that the number of tokens streamed by
01:00:46
voice models, voice to voice models is a lot higher because
01:00:50
similarly, what you said, right, you're, you're capturing more
01:00:52
detail, you're capturing like intonation, etcetera.
01:00:55
And so for a sentence, for any given sentence, if you chop it
01:00:58
up in text, it would, let's say, be like, you know, 10 tokens or
01:01:02
something in voice, it could be like 80 to 100.
01:01:06
And the more tokens you have, the more opportunity there is to
01:01:09
mess up. And so that's why you see higher
01:01:12
hallucination rates in these voice to voice models.
01:01:15
And a lot of the research the labs are doing are, is, is kind
01:01:18
of geared towards making that better.
01:01:20
Because I think we would all, I think we all want the voice to
01:01:23
be very emotive and like to capture our emotions, right,
01:01:26
just like we're talking right now.
01:01:27
But until the loose nature rate is cleaned up, you can't really
01:01:31
leverage them in these like big production use cases in like in
01:01:34
our space. How do you stop like weird edge
01:01:37
cases where like so an agent says something really sort of
01:01:41
bad. Obviously a human could do that
01:01:43
too. Like what are what are the
01:01:45
systems you have where it's like, oh man, we had one really
01:01:48
sort of fire off weirdly. Like you're sort of have another
01:01:52
sort of layer of technology monitoring everything or what?
01:01:54
What do you do there to catch when inevitably some weird thing
01:01:58
emerges? Yeah.
01:01:59
So that is that is a huge topic and we we think of it as like a
01:02:04
three prong problem. It's not just checking it during
01:02:07
the during the call which you have to do, but you also you
01:02:10
also can prepare for it beforehand and review it
01:02:12
afterwards. So the, the three prongs I would
01:02:14
consider are before the conversation, during the
01:02:16
conversation and after the conversation, but before the
01:02:18
conversation, what you can do is we, we call them simulations.
01:02:22
And so you have these AI agents that you've made and before you
01:02:26
release them to any customers, you kind of run them through
01:02:29
simulations. So it's almost like a second AI
01:02:31
comes in and it's talking to your agent and like trying to
01:02:33
get to mess up or testing all the common use cases that it's
01:02:36
learned from reading historical transcripts, right?
01:02:39
So that that's the first thing you do because that really
01:02:40
shores it up and that allows you to iterate faster because let's
01:02:42
say next week I want to change something well before I release
01:02:45
it, I can just run these simulations and I feel good
01:02:47
about releasing it right during the conversation.
01:02:50
It there is what what you're saying?
01:02:53
So those are like supervisor models and those have to be like
01:02:55
really fast specialized models that go and detect for certain
01:02:59
things. And so I'll give an example,
01:03:02
let's say we'll use the bank example again.
01:03:05
Like one of the things that if I'm a bank, I would not want the
01:03:08
AI to do necessarily is give financial advice, right?
01:03:11
So I don't want you to get financial advice that's not in
01:03:13
the scope of what your job is. And so how do you make sure that
01:03:17
it never does that even if the user's like trying to get,
01:03:20
they're like really determined to get financial advice.
01:03:23
The way you do it is you have to have these supervisor models.
01:03:25
And so there's these small models that we've, we've fine
01:03:28
tune ourselves that are really fast and really specialized in
01:03:31
detecting things like that. And you run them during the
01:03:33
conversation. And if it detects A violation,
01:03:36
it can fix it in real time before the, before the response
01:03:39
goes out. And then finally, the third
01:03:41
prong is after the conversation. And so there what you do is you
01:03:44
can review the conversations with a second AI, right?
01:03:47
And we, we were talking a little bit before, but I think this is
01:03:52
where a lot of the space is going.
01:03:53
Because if you think about the big advancements in AI right
01:03:57
now, it's these slow reasoning models that are really good at
01:04:00
coding and really good at reasoning.
01:04:02
You're never going to use them in the middle of a call because
01:04:04
like they could take like up to like a minute to reply, right?
01:04:07
And you're not going to wait there for a minute.
01:04:09
It's like, give me a second. Then it's like thinks for a
01:04:11
minute. And it doesn't even like like
01:04:14
improve those metrics very much anyways.
01:04:17
But what they are really good at is kind of this like slow
01:04:20
autonomous reasoning. And so one of the things that
01:04:23
that we've pioneered and we kind of released the, the first
01:04:25
product in, in our space, it's called Duet.
01:04:27
So Decagon Duet and the idea is it's kind of a duet between that
01:04:30
like slow reasoning model and like the the fast one, right?
01:04:33
And so the reasoning model, their job is they can basically
01:04:36
run overnight and autonomously. You just like read every single
01:04:39
conversation and you know, basically like figure out what
01:04:45
is going well, what's not going well, figure out what you need
01:04:48
to do next and actually do it for you.
01:04:50
How? How does it yeah implement the
01:04:51
learn what it learns. So the, the, let's say it
01:04:55
figures out like, oh, there's this one topic that we are not
01:04:58
doing really well at. And I've looked at our knowledge
01:05:01
base, I've looked at all the procedures.
01:05:03
We, we call them agent operating procedures.
01:05:05
AO PS. I've looked at all the AO PS
01:05:06
I've looked at like the coded tools that the AI is able to
01:05:09
use. I've looked at, you know, the
01:05:10
data that's come in from these tools.
01:05:12
And I've realized that the problem in this case is that,
01:05:15
you know, in this step of the procedure, we're like sending
01:05:18
people off a wrong track because like they don't actually want to
01:05:21
do that. And so because I've read 50
01:05:24
conversations, I'm very confident that the fix is XYZ
01:05:27
and it'll go and suggest the fix in the morning.
01:05:30
Someone on the team can come in and just like fix it, right?
01:05:32
Well, so if you think about how the space generally works, like
01:05:36
a lot of software, even outside our space, people are building
01:05:38
these AI assistants. But generally the AI assistants
01:05:42
are there to like, oh, like, show me how to find this or
01:05:45
like, go do this for me and I'll like, go do it.
01:05:47
But what we're talking about, what duet is, is it can do that,
01:05:50
but it's also like a autonomous like agent that can run in the
01:05:53
background. And that's what a lot of the
01:05:55
coding tools are moving towards right now, right.
01:05:56
And obviously I would argue that our, the work that our agent has
01:06:00
to do is like a lot simpler than coding so that they actually do
01:06:02
a better job of it. And so that that's really what
01:06:05
we pioneered. And we, we have some folks in
01:06:07
our space who have released similar things, but they're,
01:06:09
they're more focused on that like first use case of just
01:06:12
like, OK, write this for me rather than like actually
01:06:14
something that can just like run overnight for several hours and
01:06:17
like get a lot of work done. I'm sure this is a piece of it,
01:06:19
but what how would you distill Decagon's right to win?
01:06:24
I mean, obviously you have very well known competitor and then
01:06:27
obviously an existing industry and I'm sure lots of other
01:06:29
startups that are unheard of trying to run at this space.
01:06:32
Like what is your right to win? Yeah.
01:06:35
So I think the the thing about AI is that the markets are huge
01:06:39
and there's a lot of parts opportunity.
01:06:40
So that's a great thing. And then the flip side of course
01:06:43
opportunity attracts a lot of players.
01:06:45
And so in our players, in our space, you have like newer
01:06:47
players like Gen. AI native companies like you
01:06:50
mentioned Sierra and you also have like the older ones like
01:06:55
you know, the big platforms, Google, Salesforce, etcetera.
01:06:59
And of course, our goal as a team is to destroy all of our
01:07:01
competitors. And we, you know, we have very,
01:07:07
we've hired a lot of like killers on the team.
01:07:09
You you ranked higher than Sierra on this on this list,
01:07:12
which is interesting. Yeah, yeah.
01:07:14
We continue destroying the competitors, you know, no, but
01:07:16
we, we have a lot of respect for our competitors.
01:07:18
And I think that's, that's also something that's that's very
01:07:20
important. Like when we're building
01:07:21
internally like there's, you can't like demean anyone or like
01:07:24
underestimate anyone. That's very careful.
01:07:27
But I would say ultimately what it comes down to is like a very
01:07:29
different product approach. So I think the reason why in a
01:07:33
lot of big spaces there are multiple winners or in our case,
01:07:37
if if we want to be the winner, it is because you, you, you have
01:07:42
to take like kind of a, a bet on a specific approach in the
01:07:45
product. And our bet really is that the,
01:07:50
the biggest differentiator for our products and products in our
01:07:54
space over time is going to be sort of the, the empowerment of
01:07:59
non-technical folks and sort of the decrease in the cost of
01:08:02
ownership. So you think about how a lot of
01:08:04
software normally works. It is the set of phases like
01:08:07
highly consuming. There's like whole like
01:08:08
professional services industries based around this because you
01:08:11
need technical resources to, you know, get sales force set up,
01:08:14
for example. And there's a lot of benefits to
01:08:16
that. Like you get really locked in
01:08:17
afterwards, you're they're kind of reliant on you.
01:08:20
But we think that with AI, you're going to have to take a
01:08:23
very different approach and a lot of our competitors are are
01:08:26
very grounded in in the older approach, which again.
01:08:28
It's not a big deal to do it for the whole company.
01:08:31
And it's not necessarily bad, right?
01:08:32
Like. You want like a product team to
01:08:33
say, oh, we're going to use Deccon for our.
01:08:36
Exactly. OK.
01:08:37
And we want the OPS team to be like, hey, we want to use Deccon
01:08:40
because they're going to allow us to move a lot faster and
01:08:43
we're not going to have to rely on engineering Sprint or like
01:08:46
call up the vendor every single time.
01:08:48
So that's our biggest differentiator.
01:08:49
And when a. Lot like Slack where a team can
01:08:51
deploy exactly, yeah. And so when a lot of these
01:08:55
businesses work with Deck and it's because of that, right,
01:08:57
they feel like, hey, we're a big business.
01:08:59
We are a lot of complexity. You know, getting the AI live is
01:09:02
maybe only about 20% of the work, 80% of the work is the
01:09:05
constant iteration. And we're going to build a lot
01:09:07
of new flows. We're going to have to add a lot
01:09:09
of new surface areas. And if we're reliance on the
01:09:12
vendor and it just kind of feels like a black box, that's not
01:09:14
possible. You have tech companies are more
01:09:16
likely to adopt diagon right or you you've been strong in tech.
01:09:20
So tech, financial services, airlines, telecom, I think a, a
01:09:25
lot of the older industries we've we've seen the same thing
01:09:28
where they, they have less strong engineers overall
01:09:33
potentially, or they still have a, a very strong engineering
01:09:36
team, but like they have a ton of other stuff to do and
01:09:39
building customer experience software is not like one of the
01:09:42
things that they want to specialize in, right?
01:09:44
And so in the past they would have had to hire a bunch of, you
01:09:48
know, professional services or like, you know, pay the vendor a
01:09:50
ton more. But in our case it's, it's
01:09:54
really that has become the big differentiator versus the the
01:09:57
older approach. There are a lot of Americans who
01:10:00
hate AI. Like how much do you think
01:10:02
that's a barrier to your business succeeding?
01:10:05
And like how much do you think there's going to be sort of like
01:10:08
a hearts and minds battle or it's just the product has to
01:10:11
speak for itself? Or what do you think sort of AI
01:10:14
sentiment translates into like sort of people's willingness to
01:10:18
engage with AI voice and customer support?
01:10:22
Oh, I think it's very important. I definitely don't think that's
01:10:24
something you can gloss over because.
01:10:27
And so when we're building the product, like a big mantra that
01:10:30
we try to adhere to is that we're also building for our
01:10:33
customers, customers because at the end of the day, it's like
01:10:37
the goal is to make their experience a lot better.
01:10:39
And that of course benefits our customers, right.
01:10:41
So if we're working with the airline and you know, let's say
01:10:43
we save them a bunch of money, but then all the customers are
01:10:45
pissed because they can't like get what they wanted to do done.
01:10:48
Like that's not a win for anyone.
01:10:49
Like that might be a short term win for the airline, but you
01:10:52
know, they know that like that's not really what we're going for.
01:10:56
And so a lot of the product work that we do is geared towards
01:10:58
like, how do you make the end experience better, right?
01:11:01
Things like user memory that we we talked about earlier and
01:11:04
those are things that are you allowed.
01:11:06
To do that across customers? Or is it user memory within a
01:11:10
customer? Yeah, we don't do that cross
01:11:12
customers. There's no data sharing across
01:11:14
customers. I don't even think people would
01:11:16
want that necessarily. But yeah, yeah, it's hard to
01:11:18
say. I mean, if you know me, why do I
01:11:20
have to waste all the time? You know what my preferences
01:11:23
are. But so I think what will happen
01:11:25
is there will also be consumer agents.
01:11:29
So I think what the world looks like in three to five years is
01:11:32
all the, all the brands will have their own AI concierge.
01:11:35
And you know, hopefully we're we're powering and helping them
01:11:38
a lot with those. But there's also going to be
01:11:40
consumer agents that users use. So if you and I want to use
01:11:46
something like right? It's like a lawyer.
01:11:48
It's like their lawyer, my lawyer sort of fighting with it,
01:11:50
right? And then, yeah, the agents will
01:11:52
connect, right? Isn't.
01:11:53
That going to be a nightmare for you like those my I mean, in
01:11:56
some ways you have to brace for that world, right?
01:11:57
Because my agent, you know, I assume it will be meaner and
01:12:02
like I like, there's a level of like propriety that I have,
01:12:05
whereas they're just like going to treat you like a system to be
01:12:08
sort of manipulated and and you know, try to find every hole to
01:12:11
get what we want, right. I don't think it makes our life
01:12:15
harder, I think it is generally good for us and the world
01:12:20
overall. Because it's an arms race, you
01:12:21
have to fight or. I don't think it's necessarily
01:12:23
arms race. I think it's just it creates
01:12:25
much more communication. So it's much easier.
01:12:27
Like I think right now a lot of a lot of this communication
01:12:29
doesn't even happen because like I don't, I don't even bother
01:12:31
calling that number because like I know I'm going to get stuck in
01:12:34
like some loop, but no, I'm going to tell my agent to do it
01:12:37
and the agent can actually get it done because like there's a
01:12:39
business agent over there. And so it just makes like
01:12:42
overall, like the number of interactions much, much
01:12:45
healthier. But you you're not going to play
01:12:48
on sort of my agent side or are you interested in that side of
01:12:52
the no? Right now we're very focused on
01:12:54
working with businesses. I mean, so the, the brand
01:12:57
agents, if you will, the consumer agents like I think
01:12:59
ChatGPT will probably it's like those type of tools will be,
01:13:02
will be the consumer agents. You think it'll be ChatGPT
01:13:04
itself? Chat.
01:13:05
GBT, I mean like Perplexity has an agents, Claude has an agent.
01:13:10
It's like apps like that that are like geared towards mass
01:13:12
consumer apps like Gemini for example.
01:13:15
Do you think any of your customers will say you're only
01:13:18
allowed to talk to us if it's you or like have you seen people
01:13:21
say? No, I actually think I would say
01:13:24
the, the sentiment in most of our the businesses we work with
01:13:27
is, is very positive. They're excited for agent to
01:13:29
agent world. Why?
01:13:32
Because I think that is like that will yield more business
01:13:36
overall because it's like bringing the barrier to entry a
01:13:38
lot lower, right? So like if you're a hotel or
01:13:41
something and like agents can make bookings on your on your
01:13:46
like in your platform and then you know, presumably you can
01:13:49
have a lot more because the buried entry is a lot lower.
01:13:51
Do you think like the rules of like who gets a refund when are
01:13:54
going to become more transparent?
01:13:56
Because it's going to be possible to sort of test every
01:13:59
customer system and say, all right, if you say it's broken
01:14:03
two days old, they'll get like, are the rules going to end up
01:14:06
being sort of publishable because it's going to be so
01:14:09
discoverable? Or people are going to play this
01:14:11
cat and mouse game, or they're going to sort of randomize it or
01:14:13
like what happens when it's, yeah, how do you see that
01:14:16
playing out? I, I think the, the, the steady
01:14:19
state is that there aren't going to be no like games necessarily
01:14:21
because like right now there may be some games because friction
01:14:25
ads, etcetera, right. But in in the in a steady state,
01:14:30
because it's two AIS working together.
01:14:33
Yeah, there are transparent rules there.
01:14:35
There also may be like judgement calls, but the AI is the one
01:14:37
making the judgement call. So it's like fairly unbiased.
01:14:41
That is. I think that's that's the goal
01:14:43
and that that's the world that we're driving towards right now.
01:14:46
I think there's just inefficiencies because you have
01:14:48
businesses that have like their leadership might not even like
01:14:51
want these games to be there. But it's just over decades, you
01:14:54
kind of built out these things and you know, you're kind of
01:14:57
afraid to change it because, you know, like it might dramatically
01:15:03
increase the number of inquiries coming in or change your PNL.
01:15:06
But part of our job is to really work with them through that
01:15:08
process and like design these conversations so that you're not
01:15:11
losing anything. And in fact, you're, you're both
01:15:14
driving up your returns and making the experience better.
01:15:17
And that's, I think that's a big reason.
01:15:19
Back to your question of like why the space has has, you know,
01:15:21
popped off so much, It's it's really like you, you have these
01:15:25
massive ROI case studies that have already been realized by
01:15:29
big businesses where tons of savings, tons of customer
01:15:33
experience increases and, and that's the goal.
01:15:37
Is, is what technology do you really want to see improve and
01:15:40
like how is where, where are we on like voice latency?
01:15:44
You also are an industry that sort of has to rely on the past
01:15:48
generation model or you need sort of smaller models.
01:15:51
And I'm sure it's a much more price sensitive customer base.
01:15:55
Like where are you most hopeful that technology improving and
01:15:59
the models improving will change your business?
01:16:02
A couple different axes, right? So there's the the reasoning
01:16:04
models, you know, with, with products like Duet, right?
01:16:08
The reasoning capabilities will continue to help those open
01:16:12
source models getting better, will continue to help our
01:16:15
underlying agent pipeline because like why do you use open
01:16:18
source models? It's mostly for performance
01:16:20
because you need latency improvements and you need the
01:16:22
models to like highly specialized.
01:16:25
And so we found that by fine tuning and training those
01:16:27
models, we're able to get like much faster latency and much
01:16:31
higher like comparable to higher performance than the big closed
01:16:35
source models. And then it's the voice models.
01:16:37
And so, you know, voice to voice models getting better and more
01:16:40
accurate. Like those are probably the
01:16:42
three big dimensions that we pay attention to and each one of
01:16:44
those improving improves significant things about our
01:16:47
product. What What open source models are
01:16:49
you most excited about right now?
01:16:52
I mean, you have people all over the world making open source
01:16:55
models, right? So China's obviously very good
01:16:56
at open source models. The open source models in the US
01:16:59
are improving and you know, hopefully, hopefully even
01:17:02
faster. And you also missed all.
01:17:05
And so I think we found that there these models are like good
01:17:09
at different things and our customers might also have
01:17:12
preferences. And so we have built in a way
01:17:14
that's very model agnostic and you can you know, swap in and
01:17:18
out different models. But yeah, we're we're very
01:17:21
hopeful that the smaller parameter models will get better
01:17:23
and better because that is just these.
01:17:26
Are using like DeepSeek mini Max any of those or?
01:17:30
Mr. All Quinn, we we think pretty highly of and then we're
01:17:35
keeping a close eye on the like the Gemma type models, Llama,
01:17:39
things like that. You think you have to stay away
01:17:41
from the Chinese models or? I don't.
01:17:43
You don't have to stay away from them, but you want to be fairly
01:17:47
diversified so that in situations where you don't want
01:17:49
to use them, you can. You just have some customers who
01:17:52
don't want to, yeah, which then limits how much you can go all
01:17:55
in on them. Yeah, but you wouldn't want to
01:17:58
go all in on certain models anyways.
01:17:59
It's like this, it's like like you want to because if you, if
01:18:03
you get over reliant, it's like an analogy would be like if
01:18:06
you're a country, you don't want to rely on another country for
01:18:08
all your oil. You know, it's like anything
01:18:10
that happened. So we want to make sure that
01:18:12
we're building in a very robust way.
01:18:14
Do you build your own models or is there any use in that?
01:18:19
We have our own model, but they're they're not like trained
01:18:22
from scratch. So if you're in the application
01:18:24
layer, what you're typically doing is you're using one of
01:18:26
these open source models. The open source models are not
01:18:28
actually be that good out-of-the-box, but you can fine
01:18:31
tune them. And if you do that correctly,
01:18:33
they actually become very performant at the tasks that
01:18:35
you. Want are you optimistic across
01:18:37
the application layer or what? Where do you think the
01:18:41
foundation models will just gobble up applications versus
01:18:45
where do you think there's opportunity to build stand alone
01:18:47
businesses? I'm obviously quite optimistic
01:18:50
about the application layer, I think.
01:18:52
Just that you started early enough and you're sort of
01:18:54
running alongside them or what? No, I actually think so.
01:18:58
One, I think most of the value will accrue to the application
01:19:01
layer 'cause you're solving like business problems.
01:19:05
I think the labs also agreed with that, which is why they're
01:19:07
building a lot into the application layer.
01:19:10
And I think the application layer is so vast that there's
01:19:12
going to be a lot of different ways to handle it.
01:19:15
But generally, I think a good framework is that most of the
01:19:19
labs will want to build applications that are fairly
01:19:22
broad, like broad and maybe like thin because they have such a
01:19:28
vast surface area that they want to build things that a ton of
01:19:30
people can use. So coding obviously is, is one
01:19:34
of those where it's, it's mostly like a, you know, like an app
01:19:39
that people can just pick up. If you think about our space,
01:19:41
it's quite different, for better or worse.
01:19:43
We have a very involved like post sale motion where, you
01:19:47
know, we talked about our conversation designers, but
01:19:49
they're, they're part of like a much broader team that we, we
01:19:52
use to actually work with our clients and, you know, get their
01:19:55
agents stand up, stood up and like help write these agent
01:19:57
operating procedures because, you know, we have the benefit of
01:20:00
having the expertise. And because you have that post
01:20:03
sale motion, it's, I think it's pretty unlikely that's there's
01:20:06
going to be an app that replaces it.
01:20:10
But over time, I think the application layer is like so
01:20:12
thick that there's a lot, a lot of stuff to build.
01:20:14
So I, I actually, I think there's going to be a pretty
01:20:19
bright future for most application companies.
01:20:22
What just as somebody sort of so close to the space and how
01:20:25
things are developing, like how do you think the world looks
01:20:28
differently in in five years? Specific to our space.
01:20:31
Or just broadly, like, what? What do you think?
01:20:33
Like, I don't know. Yeah.
01:20:34
For the regular person just watching this, trying to
01:20:37
understand how AI is going to change their lives, like what do
01:20:39
you think feels the most different in five years because
01:20:42
of what people are building today?
01:20:44
So a couple themes. So the first one in our space,
01:20:47
we, we kind of touched on this, but I think the way that
01:20:50
consumers are going to interact with any brand is going to be
01:20:53
fundamentally different. And you know, hopefully we are
01:20:55
again a major player in that. But if you think about the
01:20:57
previous shifts, right, like last one, let's say from
01:21:02
Internet to mobile, right, that that basically created like
01:21:04
entirely new UI for people to interact with, you know, the
01:21:09
brands that they need to interact with, whether to buy
01:21:11
things or to get something done, etcetera.
01:21:14
And AI is, is you can almost think of it as like a new UI.
01:21:18
And this UI is conversational. And so you can talk to it on the
01:21:22
phone, you can chat with it. But that is, that is something
01:21:25
fundamentally that's gonna change and you're gonna have.
01:21:28
But don't people like to push back on that?
01:21:30
Some people like shopping, you know, like some of some of these
01:21:33
cases, it's like where people talk about like travel and it's
01:21:36
like people like spending time booking travel in some of these
01:21:39
cases. Like, do you really think it's
01:21:41
gonna be like, hey, like text in go figure out my travel, go
01:21:45
figure out what? I don't think it's necessarily
01:21:47
like a clean replacement, right? Like going from web to mobile.
01:21:49
Some people still, if I'm on my computer, I'll just use the
01:21:51
website and people go to stores. So, right.
01:21:53
But it's more of like a creation of a new medium.
01:21:56
Yep. And yeah, there will be
01:21:57
situations where I would rather talk to the YouTube.
01:21:59
There might be other situations where I'd rather use the mobile
01:22:01
app. And so you have this like new UI
01:22:02
that's created. And this UI is a lot more
01:22:05
friendly for consumer agents because consumer agents can also
01:22:08
go to your website and try to click around.
01:22:10
But that's like a very crude approximation.
01:22:13
But agents, again, are very good at conversation.
01:22:14
And so the two agents are talking.
01:22:15
You just like back to my original point, you're basically
01:22:19
just massively increasing the number of healthy interactions
01:22:23
between consumers and brands. Whereas right now a lot of them
01:22:26
I would describe as like unhealthy or there's not even
01:22:29
happening because I can't be bothered to do them right.
01:22:31
So that's one big thing that will change.
01:22:34
I think this is an argument that's like if you hate talking
01:22:36
to AI customer support, don't worry, I know you're trying to
01:22:39
be better than the loops that everybody got.
01:22:41
And so in a lot of cases they are happy to talk to you.
01:22:43
But there is also the further pitch, which is soon enough you
01:22:46
can have somebody talk to us and then get the readout of like,
01:22:50
what transpired. And you don't actually have to
01:22:51
do it yourself, which I think people will be excited about.
01:22:54
OK. So that's one, yeah.
01:22:55
That's one that's basically our space, right?
01:22:57
So we're really excited about that.
01:22:58
And I think that's why the market is so large and why
01:23:01
there's so much interesting work to be done in our space more
01:23:05
broadly. I think something that will
01:23:07
happen a lot more is right now people think of AI as it's like
01:23:13
a tool there. It's kind of like Google, right?
01:23:15
You go to Google and you like search something and like, OK, I
01:23:17
go to my, I go to AI and like, I get something done.
01:23:20
But I think in, in three to five years, it'll be much more normal
01:23:24
to just have like AI running continuously and it's just like
01:23:27
always running on whatever. And it's, it's kind of happening
01:23:30
in the background, right? Because right now it's kind of
01:23:32
like you ping it and it does something.
01:23:34
But the sort of the, the standard practice that we're
01:23:38
moving towards is just like long running autonomous things.
01:23:41
And that's, that's cool for a bunch of reasons.
01:23:44
One, you can just accomplish like way greater tasks.
01:23:48
And two, you can really increase leverage because like you don't
01:23:50
have to be involved, like it's just doing things.
01:23:53
And so this is a, this is a big trend in the coding space.
01:23:55
If you're following those, like a lot of what?
01:23:58
You start to sleep, wake up and see what's happened overnight.
01:24:00
Exactly. Yeah.
01:24:01
Or what started with like it's, it's kind of tab auto completing
01:24:05
things Now it's, you know, it moved to OK, you, you give it
01:24:08
some prompt and it can like write whole things and you
01:24:10
review it. And then now it's moving towards
01:24:12
like, OK, it's just like, hey, let me give you some
01:24:15
instructions. And then like you just kick it
01:24:16
off and it's like a colleague doing work, right?
01:24:19
That's the same concept again behind Duet where we've kind of
01:24:22
created this, this new concept of like a long running agent in
01:24:24
the background. And I think that will happen
01:24:26
with consumer agents as well, where you just give it like long
01:24:28
tasks or it's constantly listening to you and it's just
01:24:32
like a always there like helper. Great, Jesse, thank you so much
01:24:37
for coming on the Newcomer Podcast.
01:24:38
Cool, Eric, thanks for having me and.
01:24:39
Congratulations for being on the list.
01:24:41
Thank you. All right, sweet.
01:24:42
Thank you. Thanks for sticking around to
01:24:44
the end. Please.
01:24:45
If you've made it this far, you've got a like comment,
01:24:47
subscribe. Excited to grow the channel.
01:24:50
And of course, you can find our writing and reporting on sharps
01:24:53
and venture capital at newcomer.co.
01:24:57
We also host events. You can check out what we're
01:24:59
doing at Newcomer dot events, share other podcasts, help us
01:25:03
get distribution, support the channel comment support, like
01:25:06
comment subscribe. You know, you know the deal.
01:25:09
Thanks. Thanks for being on the journey
01:25:10
with us. All right, see you next week.