Latest / Agentic Conversations (formally mlops.community) / Investing in the Next Generation of AI & ML // Jill Chase & Manmeet Gujral // MLOps Podcast #143

Transcript

Investing in the Next Generation of AI & ML // Jill Chase & Manmeet Gujral // MLOps Podcast #143 · Agentic Conversations (formally mlops.community)

0:00
On Monday. I'm an investor at Capital G and
0:04
I usually drink cold brew with a little bit of I'm Jill.
0:08
I'm an investor capital G and I currently am enjoying single
0:12
shot. Hold of lattes.
0:15
What's going on? Everyone this is another edition
0:18
of your ml, Ops Community podcast.
0:20
I am your host Demetrios and super excited to talk to you all
0:24
today because wow, what a conversation with my me and
0:29
Jill. They went through the Paradigm
0:32
that investors are looking at right now when it comes to
0:35
developer tooling and the whole foundational, AI foundational,
0:40
model movement, I thought they had some incredible insights.
0:44
When it comes to the quality of Revenue around foundational
0:48
models, what does that look like for them?
0:51
And their investment thesis has. And how does that play into
0:56
machine? Learning operations and ml Ops
0:58
how much of em Phillips is going to be eaten by foundational
1:03
models in the idea of generative, AI large language
1:07
models and how much of ml Ops will still be standing in 10
1:11
years. So hope you enjoy this
1:14
conversation with Jill and me from Capital cheat.
1:23
All right by me and Jill you're currently investors at Capital
1:27
G. I know it is kind of related to
1:29
Google or it has something to do with Google.
1:31
Maybe you can break down exactly.
1:33
What cap she is. What kind of deals.
1:36
You're focusing on what you're looking at and how that all ties
1:40
into ml Ops. Yeah, totally well, thanks for
1:44
having us. We trusted that it's fun to hop
1:47
on a chat with you after after having listened to some of these
1:50
podcasts and maybe I'll start with a kind of a quick
1:53
introduction to myself and then ya bastards jail.
1:56
So she can chop a bit more about herself and a capital G.
1:58
But so I'm going to meet an investor capital G get to work
2:01
closely with with Jill and helping to lead Investments
2:04
around AI am L data and Source software and have been on the
2:08
team now for almost a year and a half.
2:09
So sometimes line prior to Capital G was that Tecton where
2:14
I got to meet to be chairs for the first time and helping to
2:16
set out our developer conference.
2:18
Apply and setting up an insect on East community.
2:21
So do you not go way back in the scheme of things?
2:24
But we got to work with some great folks on the tecton to you
2:26
in there and work on all things. Go to market, so wearing as many
2:29
notes as I could. And then prior to that, I was a
2:32
painting company. So working with, you know, large
2:34
cap tech companies on go to market and strategy and then
2:37
private equity and growth Equity investors on things like invest
2:40
in Tilden, so excited to be here.
2:44
Awesome. Yeah, thank you so much for
2:45
having us a super pumped for the conversation.
2:48
I would say, super briefly on, on my background.
2:51
Like meat said, we lead our AI am L Investments at Capital G.
2:54
I've been here about three years and prior to takeoff Gio's more
2:58
on the operating side. So I ran a YC Back, start up and
3:01
ran a private Equity back business, which two very
3:03
different operating experiences and then came over gabala G very
3:06
quickly. Just in case it's helpful and
3:08
I'll share a little bit about what cap G is and and sort of
3:10
where we focus our time and efforts.
3:12
So Capital G is alphabets independent growth fund.
3:15
So really two parts of that independent meaning we operate
3:18
entirely, like, sort of normal Financial investment fund.
3:23
So, just like and rusyns Goya red Point index etcetera.
3:26
It just so happens that our only LP is alphabet.
3:29
So we invest often out cheated off of it and then growth,
3:32
meaning, sort of anything really post product Market.
3:35
Fit is where we tend to spend most and that can be serious.
3:38
Be serious? See.
3:40
All the way through IPO recently, we've even started
3:43
Looking at sort of more series, a companies and specific
3:45
categories. And we could talk about how that
3:47
relates to a I am l in a second. And the only other thing I share
3:50
I guess is our general strategy for investing is we try to write
3:54
big checks and spend a lot of time with relatively few
3:58
portfolio companies. And so, you know, we'll even if
4:01
our first check size is 25, million will try to get upwards
4:04
of 150 million dollars into every business that we work
4:07
with. And we will spend a lot of time
4:08
with those businesses and as a result, you know, our strategy
4:11
tends to be Let's spend 18 months on a category.
4:16
Learn everything we can about that category and for the meat
4:19
myself, it's been AI am L. We've been living breathing it
4:21
for the past 18 months. Working with Folks at Google
4:24
brain and Google cloud and helping them figure out or
4:27
having them help us figure out our thesis.
4:29
What we think is most interesting, and then we'll sort
4:32
of pick companies to invest in from there.
4:33
So very dramatically driven and very focused on a, I am L.
4:36
Excellent. So we should start with one
4:40
question that I have been pondering. during and maybe met
4:45
me all ask you this because I know we've talked quite a bit
4:48
about it in the past when we get on our our chats Is ml Ops going
4:56
to get sideswiped by the foundational models and like the
5:02
generative AI hype haha. Alright, starting off hot.
5:08
We just that star not found reversal any.
5:11
It seems to be on folks, minds. I think, in our opinion, there's
5:16
always going to be room for sort of building your own models and
5:20
in particular, maybe they won't Be sort of large models or
5:24
Foundation models. Obviously those are very
5:26
different about, you know, building and scaling
5:28
environments but as long as there's kind of proprietary data
5:32
smaller use cases, there's always going to be room, I think
5:35
for folks, building and deploying their own models and
5:37
how across different Industries across kind of Enterprises and
5:40
that's where ml option. So that's where the storing tool
5:43
chain. And the infrastructure, that a
5:45
lot of folks have built beautiful Source or kind of
5:48
proprietary helps, folks, do that announce, there's a lot of
5:50
hype around, you know, large language models Foundation
5:53
Models but that doesn't necessarily just, I think, blow
5:56
out the rest of the kind of work and momentum.
5:58
Folks have built it in ml of wrong.
6:01
Nice. Jill.
6:02
Do you have any other thoughts on that?
6:05
Or I would say, you know, we have long historical even before
6:10
sort of the the previous three months, when Foundation models
6:13
have blown up in the public eye, we've been looking at the space,
6:17
as sort of into categories, what?
6:19
We call build your own ml or an elapsed, I think so.
6:23
Of the scaffolding infrastructure for tools selling
6:25
into machine learning engineers. And then what we call machine
6:28
learning as a service which Foundation models would sort of
6:31
fit into, which is basically, like, abstracting way
6:33
infrastructure and just giving the output of machine learning
6:36
and we can talk about that at length, but the reason that's
6:39
relevant to your question is, I think our perspective has been
6:43
there will be large businesses built in.
6:45
Both of those categories, they just are really different and
6:49
consumers and so you can imagine in sort of the Ops category.
6:54
You're likely to be selling more into large Enterprises that have
6:58
really sophisticated machine, learning teams, and machine
7:00
learning engineers, and house. And you still need a lot of the
7:03
tools that the ml Ops Community is building in order to make
7:07
successful sort of In-House models that are very
7:09
specifically oriented towards the use cases that those
7:11
Enterprises need and the only other and actually I was just at
7:15
the fortune AI brainstorm conference earlier this week
7:18
which was like with basically everyone there was like the CTO
7:21
of Walmart and CDF. Apple one and all this sort of
7:24
like technical leadership of many of these large Enterprises
7:27
and I would say there was still a lot, you know, a lot of
7:30
excitement and interest around foundation models.
7:32
But more excitement and interest around ml Ops and understanding
7:35
like, hey we still have a team of ml.
7:37
He's and we still have this massive proprietary data set and
7:40
we still have this model use case and it's not immediately
7:43
clear how we would use a foundation model.
7:46
But it is immediately clear what models we want to build and we
7:49
still need the tool, get from ml Ops to do that.
7:53
Oh God, so good. So the big thing of course you
7:57
both being investors. I think the question that comes
8:01
to my mind instantly is when you're looking at the space of
8:05
ml Ops and the companies that are coming up in ml up space and
8:10
the revenue that they've been generating in the market that is
8:13
there that has developed compared to the foundational
8:17
model space and all the hype and the revenue that these companies
8:20
have been generating it. It seems like when we look at
8:24
those self-serve or ml as a service type companies, they're
8:30
generating a ton of Revenue and it is very much what an investor
8:35
would love to see. I guess that's why there is a
8:38
lot of hype, and a lot of eyeballs and attention going
8:42
towards that space. And I'm wondering because I
8:46
spoke with saqib about a week ago, who is an investor at
8:51
Bessemer? And he was saying, he still is
8:54
very confident in the mo up, space taking off and the market
8:58
growing to being very large but he doesn't see it being this
9:05
like straight up or vertical type growth from the company's.
9:10
It's more of that Slow Burn. I'm wondering, do you feel the
9:14
same way? Do you have strong opinions
9:16
about that? And if so what?
9:19
Yeah, I'm happy to start and then the meat you can chime in I
9:21
think so. I think broadly.
9:23
Yes, agree with that point. I think the, there's two things
9:26
I guess there's like the slope of the revenue and then there's
9:29
the quality of the revenue and the quality of the business
9:32
model and maybe I'll hit on both the slope of the revenue.
9:36
The reality of what we've seen in the ml op space is because
9:40
and a lops is uniquely powerful in large.
9:44
Enterprise use cases where again, there are teams of
9:47
machine learning Engineers, who are sophisticated and there are
9:50
very clear, use cases. And lots of data in these
9:54
Enterprises, the this Revenue rep tends to be slower and the
9:59
go-to-market processes tend to be a little bit less efficient
10:03
in the early days because it ends up looking a little bit
10:06
more like one or two million dollar contracts with like large
10:10
Enterprises with a little bit of like Consulting Styles sales
10:15
because you're kind of trying to fit the sail into what the use
10:17
case is. But it's a big contract so you
10:19
want to do it and like it makes a ton of Of sense.
10:22
And I think there's a lot of Market pull for those types of
10:25
contracts but it does make it such that having like true
10:29
go-to-market velocity like what you're seeing on that sort of ml
10:32
as a service Foundation model side.
10:34
It's just a lot harder. It's just a different cell that
10:37
said I would agree with you mr. From from Bessemer in that like
10:41
they're still a big Market there to capture.
10:44
It's just a very different go to market model.
10:47
Like it's lot more Enterprise sales.
10:49
It's going to be a lot slower but the contracts are Be a lot
10:52
bigger and ultimately I think it will be very tied to Roi and
10:56
extraordinarily herder about. And so that gets this like
10:59
Revenue quality business model quality piece where I do think
11:03
the ml upside as long as you're being incredibly thoughtful
11:06
about not doing a true Consulting project and offering
11:09
in sort of a scalable product has way you end up with like you
11:13
are selling a product and the margins are going to look really
11:17
good and at a certain point like the payback periods are going to
11:21
look really good because The contracts are much higher and
11:23
it's very hard to rip out and it's very differentiated it's
11:27
different on the analyze, a service or Foundation model, the
11:30
side, where I think, we also have a lot of excitement because
11:33
the revenue trajectories are a lot higher because you're
11:35
hitting on like pretty clear Market pull.
11:38
I would say in market, demand, and it's often times you're
11:41
selling into like, consumers or smbs or prosumers, and so, the
11:47
go-to-market motion just enables a lot faster, Revenue growth,
11:51
which is Really exciting. I think the hard part there is
11:55
on the sort of like quality, but this is all you really want to
11:58
sort of sharply, look at what the differentiation is Because
12:01
unless and we can talk about this more but unless you're
12:04
building your own model or I've some tech mode or have some you
12:07
I'm oh, like a lot of these companies that are all doing the
12:10
same thing are all going 1230, you know?
12:12
And like that's that is exciting and a little scary.
12:15
Now I meet, I will totally agree with John and think like they
12:19
sort of go to market motions and and serve like and more.
12:22
Cats are very different for those two categories I think
12:25
zooming out to like it's almost like the saying of teaching
12:30
someone to fish versus giving them a fish where you know the
12:34
large language model companies that have kind of taken off,
12:37
have found a really kind of clean use case, or in the
12:41
process of finding a clean, use case with a clear End Market.
12:45
And often like consumer style demand where individuals will
12:49
down the platform And subscribe and what's different Giant true.
12:53
When you do that versus ml, Ops tooling infrastructure, tooling
12:56
developer tooling is that is even if you pick the best
13:00
company in the ml Ops landscape in the long run, sort of an
13:04
index ml workloads, right? It's a at a crate on ML compute
13:09
and how much ml adoption is. So it's teaching the Enterprises
13:14
to fish or the the End customer to fish.
13:16
And so I think that ends up being the distinction where you
13:19
know if you are you sure a building the tooling, it can be
13:21
a A longer Journey, But ultimately there's much bigger
13:25
pool of upside potentially, it was, you know, all the potential
13:28
ML workloads and all the potential and all developers and
13:32
all the potential ml use cases versus ultimately building a
13:36
really good and application that can take off with sort of bag
13:39
consumer style, scan suggests point.
13:42
So I love this idea. Around the foundational models
13:46
are almost more like how I Envision it as like it's b2c on
13:50
style where you go, you grab it. And it's lightweight, you're up
13:56
and running within 10 minutes, potentially of just putting your
14:00
email into a website but then later, that quality of that
14:05
revenue is not so clear because we are still very much in the
14:10
early phase and we don't know. A what is the mote?
14:13
That is there that these foundational models have
14:16
considering everyone has access to these models unless you're
14:21
training your own. And so the models are almost
14:23
commoditized in that way. And then when you compare that
14:28
to the ml, Ops side of things and these devtools, it's a much
14:33
slower cell and it's a much more cumbersome process.
14:38
But once you're in, you're in, and it's gonna be a real pain.
14:42
To rip that out because of all of the headaches that you went
14:46
through to just get it working and get the team on board, get
14:50
every word up to speed on how to do things and so I can see that.
14:55
See that the big question that I've got in my head is how much
15:01
of what we currently call ml Ops?
15:05
Do you feel like could be verticalized into these
15:09
foundational model type? Scenarios.
15:13
So how much of ml Ops is now going to not be necessary?
15:18
Because you have foundational models that are so much faster
15:22
and you just boom, you're up and running in a day and you don't
15:26
have to worry about any of that, white glove treatment talking
15:29
with the vendor, having to ask for special features and all
15:34
everything that we know that comes with the Enterprise sales
15:37
motion. So how much is going to be
15:39
verticalized by a foundational? Model.
15:43
All right, I think one thing I would call out is like the, the
15:46
companies that have popped out in leveraging large language
15:50
models or Foundation models so far and there's obviously we're
15:54
they were in the, you know, the top of the first inning.
15:56
So there's a lot that's going to count.
15:58
But they mostly been around to your point kind of like consumer
16:02
style and application, but not really tied to a proprietary
16:07
data, proprietary environments, right?
16:10
It's it's things that folks can adopt like they would adopt
16:12
notion Start typing their notes into or start typing text into
16:17
it. So I think that's like a scale
16:18
and or like a vector that we think about a lot to for what
16:22
kind of remains within the boundary of ml Ops and building
16:25
your own ml is the extent to which you're working with
16:27
proprietary data, the extent to, which you're comfortable sending
16:30
things out to an application or training, Piaf, or even.
16:33
In some cases, you know, if you think about dental services or
16:36
health care, where your proprietary data is about where
16:38
it's a, it's a competitive advantage to keep that in-house
16:41
and to train, your own models on that.
16:43
So totally agree. Like there are a number of and
16:46
use cases and verticals were large models.
16:48
Foundation models are going to build, you know, really exciting
16:51
businesses and what that end state is, we can talk about
16:54
whether, you know, it consolidates, certain certain
16:56
number of companies or Weber structure goes, but it in my
16:59
opinion, I think there's always going to be this domain sort of
17:02
building models in house using proprietary data for use cases
17:07
that already exist or in this domain that art really
17:10
generative or Foundation model-based.
17:13
You call sort of analytical and now, you could even call
17:15
operational and now, for for a lot of the kind of lives in
17:18
deployment and application use cases.
17:20
Yeah, I think that's well said, I mean, the only thing I'd add I
17:23
guess is like, I must mention this at the beginning.
17:25
I think foundational models have just I think have opened up sort
17:30
of a new type of use case or feature or whatever you want to
17:34
call it for ML Ops which is this like fine-tuning.
17:38
And so as an example I think like we said a ton of time with
17:41
Alex Ratner at snorkel. He's He's like an amazing guy.
17:44
I don't know if you've spent time in that, they've built out
17:46
a really interesting platform where they're doing data
17:49
labeling, they've been doing data labeling, they have snorkel
17:52
flow, which sort of expand across the entire value chain.
17:54
And then, with the Advent of foundation models, they released
17:58
a new feature fairly recently, which is this sort of ability to
18:01
like really go into the snorkel flow platform and fine-tune the
18:06
foundation models, such that they are effective for your use
18:09
case. And so I think even in a state
18:11
where I wouldn't if I Random Ops company.
18:14
I wouldn't be overly concerned about Foundation miles, to be
18:16
honest, but I do think it's very smart to take advantage of the
18:19
hype associated with those models and assume that for many
18:22
of your customers. Like they will probably try to
18:25
use gbd3 at some point and it will be relevant for them at
18:28
some point. And if there's a part of your
18:30
products that can allow them to more effectively, use the
18:33
foundation models and tune them so that they're more
18:36
appropriately. Designed for whatever use case
18:38
you're building for, that's a great way to take advantage from
18:40
the hype without sort of like completely pivoting your
18:42
business. Model.
18:45
Oh, I love that. Yeah, that's so true.
18:48
It's like being able to almost use the demand.
18:53
That is there in the market and say here, we'll take some of
18:57
that demand. But in reality, we know that you
19:01
really are looking for something a little bit more sophisticated
19:05
or something a little bit more in the style that we like to do
19:08
it. And you said something
19:10
interesting there, Joe, and maybe you can jump into it a
19:13
little bit more. And you said, if I were an ml
19:15
Ops company, I wouldn't be afraid of foundational models
19:19
taking over anytime soon or something like that.
19:22
Q dig into that a little bit more.
19:24
Yeah, I know. I said, I wouldn't be overly
19:26
concerned about them, and the reason for that is I just think
19:29
they're really different. We spent a lot of time.
19:31
For obvious reason, talking about the envelopes category, I
19:34
am extraordinarily bullish on Foundation models.
19:36
And we could talk about that at length as well.
19:39
I do think like they're really different.
19:41
It's just different use cases. It's Current and consumers like.
19:46
And so I don't think Foundation models change.
19:50
Really anything about the strategy of an ml Ops company?
19:54
Like if I were animal Ops company, I would remain
19:58
maniacally focused on finding end, customers who have massive
20:03
amounts of data and to have specific use cases.
20:07
And I would, I would find like Industries.
20:09
So I would go after Healthcare and fin serve and Retail and I
20:13
would have No ties to use cases in each of those categories.
20:16
And I would go after the big Enterprise customers in each of
20:18
those categories. And I would make sure I
20:21
understand how I dove tail with a foundation model.
20:24
If those customers are wanting to use something like G PT 3,
20:28
but I think like it doesn't and I think maybe it's just because
20:31
I just came from this conference where everyone sort of like
20:34
excited about you know Foundation models but still kind
20:37
of focused on like hey we have this AI Center of Excellence,
20:42
that is deciding what use cases we need.
20:44
And we're looking at all of our bread, your data, and we're
20:46
figuring out how to get from this data to that you say that
20:49
like Roi of ML, and that's just not something that Foundation
20:53
models are going to help with. And so, I worry a little that,
20:57
like, everybody. I mean investors classically
21:00
because like, we can't focus on one thing for more than 20
21:02
seconds. Are just all over the place,
21:04
like now everything's Foundation.
21:05
But like, I do, I wouldn't want CEOs of these MLS companies to
21:11
and I'm sure they're not, but, like, over rotate towards
21:13
reactants. Not because it's a completely
21:15
different model, it's a completely different and market
21:18
and I think there's room to build big businesses in both
21:21
categories. He is the queen of Emma of the
21:27
torque and this is your daily dose of Emma logs.
21:31
Well if you are dead serious about the ml Ops, you have to
21:36
immerse yourself in the em ups content.
21:40
The best way to do it is to subscribe to the in Iraq.
21:44
The community podcast, all great people are here on this podcast,
21:49
so good luck and keep learning. I love that.
21:57
Yeah, so good. So now the interesting piece for
22:02
me is when you all are looking at companies right now.
22:08
What are you seeing and what is interesting?
22:11
Because I've heard A little bit from other investors that I
22:16
talked to, and they've kind of had the same sentiment around.
22:21
I'm not really seeing super Innovative stuff.
22:25
Come out on the front's that our people are trying to raise money
22:31
on 4ml Ops companies. So it's like another data
22:36
labeling company but it's not super different than the V7 or
22:42
snorkel or scale or what's your Differentiating Factor there.
22:45
Like, why would you as a VC decide to give that money and
22:50
especially if you are a company that is just starting now.
22:54
It's almost like, have you either of you seen any really
22:58
Innovative companies recently that are just getting started?
23:04
I know you're both kind of in the growth stages.
23:07
So maybe it's like, yeah, well, it's not really where we focus
23:11
our time and energy but maybe you Have seen in in growth
23:16
stages, I don't know. Like, has there been anything
23:19
Innovative in your eyes in the envelopes fear that you want to
23:24
call out right now and feel free to like mention her by name.
23:27
If you thinking about danilov space over the past, maybe now
23:31
five years at this point where that sort of like initial cohort
23:35
of some of the companies that got that that name started to
23:38
pop out and and to your point that seems like, you know, that
23:42
tool chain that kind of classic die.
23:44
A gram people show of like the steps of ml development has a
23:47
really changed, but there are new companies that pop up and
23:49
you point Solutions and you open source libraries and I would
23:52
say, I guess just reflecting on the first party question there,
23:55
but I think a lot of that is due to a couple things.
23:59
I think the first is there actually isn't a really well
24:03
accepted or widely accepted ml development.
24:07
Philosophy, like ml Ops refers to this really disparate tool
24:11
chain of a number of different kind of Bees and point
24:14
Solutions. And I would say the vast
24:16
majority of the questions people have about the space and you
24:19
can, you can attest this when you see the greedy is that, it's
24:22
the Tidy sum together. And what do I do with them?
24:24
And where do I start? And so, you know, I think that's
24:28
sort of led to a lot of like, the maybe the actual to leg or
24:33
category of companies, is it Innovative?
24:35
But where they do innovate, is offering sort of an opinionated
24:38
workflow. A better kind of developer
24:40
experience and something gel and I get really excited about to is
24:44
Is sort of a really good on-ramp for folks, where, you know, you
24:49
can look at a lot of these companies that have popped up
24:51
maybe at the end of the ml development tool Cheney or in
24:54
the middle. And what they struggle with is a
24:56
little bit of a this kind of chicken and egg problem where a
24:59
lot of the customers that know who you are and what you do are
25:02
either super, super small, or they're so big and so developed
25:05
on the ml side to the point where they already have like a
25:07
sprawling infrastructure and you're kind of coming in and
25:11
offering a lot of services to move things around.
25:13
And so, where we go? Studies I can't wear.
25:15
Is there a weed in the ml development?
25:17
Kind of life cycle where it's easy for developers to get on
25:21
board and it's easy for them to get started.
25:23
It's easy for them to count Plug and Play into their developer
25:26
workflow and to kind of get to the end deployment State,
25:29
without sort of all all the, all the kind of massive, the MLS
25:33
infrastructure in between it. So obviously, you know, Joe
25:36
mentioned snort bolstered by think, is it company?
25:38
We really admire that then sort of taken that approach starting
25:42
with data labeling but being Really taking their kind of
25:44
unique, programmatic approach and building deployment and, and
25:47
monitoring on top of that as a sort of full platform, you know,
25:50
would be remiss to not mention, how do you face given all the
25:53
kind of velocity and acceleration?
25:55
They've seen as just like the the number one place to start.
25:58
If you want to use sort of the next generation of models and in
26:01
any of your applications or S developer just want to Tinker
26:04
with them. And so, we, I think we get
26:06
really excited about those companies that are building
26:08
really strong communities, really strong opinion of the
26:10
workflows at the beginning of the life cycle, and kind of
26:13
expanding from there, So if I understood this correctly,
26:17
you're mentioning these two different prototypes of
26:22
companies. What is with the loan data
26:24
scientist? And he or she is having to
26:28
figure things out on their own and that person needs some kind
26:34
of easy way to jump onto a platform and have the end-to-end
26:41
solution almost done for them. And then you've got the other
26:43
case where it's like, I think about Banks because you have the
26:47
gigantic financial institution that has a 20 different types of
26:53
streaming infrastructure. And so a snorkel is useful there
26:59
because you can basically go around all of these obstacles of
27:05
the heavy Legacy that is in place.
27:08
Totally totally. Do, you know, it's huge.
27:10
Huge banks are huge insurance companies and obviously, you
27:13
know, you do your best to filter them and be thoughtful about
27:16
your sales funnel, and what qualified actually means.
27:19
But it's a good indication of sort of where this base is.
27:22
And I think you said it, well, like with the on-ramp companies
27:27
or wedging into sort of the beginning of the workflow, you
27:30
have this Advantage, we're on the early side, or maybe it's
27:34
bottom up depending on your product is companies that are
27:37
taker in just getting And, you know, one day of scientist, they
27:40
can start there, and they can build their, it can be very
27:43
user-friendly. You can sort of be the sort of a
27:46
right of first refusal almost. I'm sort of that workload and
27:49
what they're building. And then for the Enterprise's, I
27:52
think what's interesting is like, yeah, it's easier if
27:55
you're bullish on a lot. So you should have bullish again
27:58
on this concept of like more and more ML workloads.
28:01
And so more and more kind of ml development from from end, you
28:04
know, and businesses and Enterprises.
28:06
And in that case, wedging at the beginning even Today means are
28:09
going to capture a huge kind of portion of that pie when you
28:12
when you fast forward five ten years from now.
28:15
And so even then if they have to your point, like if you're a GP
28:17
Morgan or Goldman, that has grown deeper structure, POC with
28:22
every Emma, loves company you've ever heard of, and it's still
28:25
really useful to be where developers start and where they
28:28
kind of choose to begin their workflow.
28:31
I love that. So me as the mL of skeptic that
28:36
I am because I can Can't be in the middle of this sand and with
28:43
a clear conscience say that it's all hunky-dory in ml Ops land.
28:49
How would you respond to the idea that I present that says,
29:00
the majority of ml Ops companies that I know have not found
29:04
product Market, fit Island. Largely agree with that
29:07
statement. And I think it gets to I mean.
29:16
Certainly a controversial one. I think it's a really good
29:21
point, and I think it gets to, it's a sharper way of saying,
29:24
what I was saying earlier around the Consulting style
29:28
go-to-market of, and the sort of like, ease of getting 21, 22, 25
29:33
million and are are for some of these businesses.
29:36
And then the like challenge associated with, scaling from
29:39
there. Because what that indicates is
29:44
that it's it's not quite Right product for you to fit just yet.
29:49
Because I think the reality is like, Due to the nature of the
29:54
organizations that ml Ops companies are selling into.
29:59
You end up having to do quite a bit of Consulting style work to
30:04
match, whatever. You've built to the end to make
30:08
it useful and drive Roi for these large Enterprises.
30:12
And as soon as you start doing that, it takes it farther away
30:15
from being a productized version of like, hey we can we have this
30:19
we're not changing it, we're selling it.
30:22
And the go-to-market efficiency metrics are going to look great.
30:25
And it's just true and I think that I would say that is a
30:29
Vestige of the fact that machine learning in large, Enterprises
30:34
in the machine learning, Market is very early in its Evolution.
30:37
I don't think that these companies will never find
30:40
product market, and I think some of them are on their way towards
30:43
it and some of them have it, but I do think there is an element
30:46
of like, it's still probably not product Market fit if you have
30:50
two or three or even seven contracts in a couple, big
30:54
Enterprises that are large. A CB you do have to Until you're
30:58
seeing or Chi double sales motion and Market, Poll for
31:01
exactly the products that you've built with, not a lot of
31:04
customization on top of that to really clean.
31:07
Like, hey, we've got product Market fit, we're ready to scale
31:10
aggressively, and I do think that will come as there are more
31:13
productized, use cases uncovered in some of these big categories,
31:16
like retail Healthcare Etc. But we're not there yet from a
31:20
market Evolution perspective. Mommy.
31:24
And she'll cover that really well.
31:26
And it kind of pulls the strings together on a lot of the
31:30
commentary we've had on the different angles of ml Ops or
31:32
like, there isn't a broadly accepted workflow.
31:35
It has been early Innings on like the end use cases that so
31:38
totally agree. I think it to your point gel.
31:42
He's just a sharper way of basically saying.
31:45
There's not that much product working for ya.
31:48
So all the sponsors that we once had we lost them laughs.
31:53
That was they all I said now we're out of here now.
31:57
Well, I really just the only thing, I'd guess I'd really
32:00
double-booked on there is like, I do think it's, it is an
32:03
indicator of the market Evolution, which I truly believe
32:08
over time, there will be more and more of these consistent use
32:12
cases. Identified in part of it, too is
32:14
like these Enterprises have to get their data in a place where
32:16
it's able to be used and like it's all in the data warehouse
32:20
or anything a lake or however you want to structure it in,
32:22
like you can get it to a point. Word.
32:24
Then you can apply machine working on top of it.
32:26
And so, I just think we're early days of it but it's not that
32:29
these are bad ideas or that these aren't going to be big
32:32
businesses. It's just that right now is a
32:33
hard time in the evolution of the market.
32:35
It's like a still a little early for an ellipse considering that
32:40
you mentioned the ml Ops market and considering everything else
32:45
that we've just been talking about especially when it comes
32:48
to ml Ops tools and product Market fit and where you play in
32:53
the growth. Action.
32:56
And now thinking about the current market conditions
33:00
because I'm not so sure I'm not a investor but I've heard it's
33:04
hard out there for an investor and it's hard out there for
33:07
Founders that are in the growth stages.
33:10
How are you adapting to our current reality?
33:16
Yeah. For sure.
33:17
Let me talk about a little bit of that, but first, I just, I
33:19
love when we say like, it's hard out there for, you know,
33:21
investors or like, we are so fortunate to be in a time where
33:25
it's like a IML is exploding and fundamentally changing the way
33:29
the world works. And so while it is you know,
33:31
harder maybe that at was in The Glory Days of 2021, it's still
33:36
an awesome time to be in this space but the funding Dynamics
33:39
on the investing side Jokes Aside like are real.
33:42
And I think, let me just explain sort of what we're seeing from
33:44
an investment perspective and how Trying to play in the market
33:48
like you basically as we were talking about before where you
33:50
have sort of like that a lopsided and then you have the
33:53
like Foundation models machine learning as a service side on
33:56
the envelope side, what you saw is in 2020 and 2021, a lot of
34:01
these ml Ops companies sort of scaled to 1 million 2, million
34:05
3, million and ARR. And because of the Dynamics of
34:08
the of the fundraising environment.
34:10
In that time period, we're able to raise up quite high
34:13
valuations. So you know, in even upwards Of
34:16
a billion dollars off of, you know, a million or two million
34:19
in error are. And now, of course, with the
34:21
resetting of the market, you know, that risk reward is, is
34:25
quite challenging and we're it when you pair that, with some of
34:29
those companies having a harder time, sort of, like, hitting
34:32
that go-to-market velocity, like we're talking about earlier and
34:34
continuing to scale into that valuation.
34:38
It makes it very hard to find, really sort of interesting
34:42
rounds. In the ml offside at the growth
34:45
stage of the market. Because Just kind of waiting for
34:47
some of those companies, all of which are very high quality
34:50
companies, but you have to wait for them to grow into their
34:52
valuations given the massive reset in the market and given
34:55
just the like absolute number of the valuation is high.
34:59
When you look at sort of compressing public market and so
35:02
that category has been a little bit more of a wait-and-see and
35:05
let's try to get sort of ml, Ops moving more and more into their
35:09
valuations before we go in and see what the next rounds might
35:12
look like. We've seen almost like the
35:13
entirely opposite thing on the foundation.
35:16
Model like and then allows the service side of the market where
35:20
because those many of those organizations are many of those
35:23
companies, tend to be compute-intensive.
35:25
If they're building their own models, what ends up happening
35:28
is the first round of these businesses fundraising around is
35:32
sort of like Hey we're going to raise 20 or 50 million at, you
35:35
know, 200 or 250 posts with just a team and an idea.
35:40
And, you know, that's that's different than the other seed
35:43
and series a rounds that exist in the market.
35:46
And then you fast forward, six months and these companies,
35:49
maybe have a demo that they've used that sort of first round of
35:52
funding to build out. But then they need a lot more
35:55
Capital to fuel a lot more compute requirements.
35:58
And so then it becomes like, hey now we're raising 150 million
36:01
dollars at, you know, a billion plus.
36:03
And that's a fard risk-reward to because again, you still don't
36:07
have a product and Market, Isola net revenue.
36:09
And so it's your kind of even though we're a year later in the
36:13
market has reset. I think there's so much
36:15
excitement. I meant about the potential of
36:17
this category and the unique dynamics of the category,
36:20
require a lot of capital, particularly if you're building
36:24
sort of any of your own Foundation models and so you end
36:27
up with a challenging fundraising environment in the
36:29
opposite direction for those types of businesses.
36:31
So we've made one investment that still in stealth and other
36:36
than that sort of work. Kind of waiting and seeing the
36:39
mL of stewing companies are going to have to grow into their
36:41
valuations, considering they were able to raise at monster
36:47
valuations a year ago or two years ago, Has it not hit the
36:52
point yet where these companies are needing to?
36:58
Get cash and just say, give me money, I don't care what my
37:02
valuation is again and they're willing to do down round.
37:05
Yeah, I don't think that's happened quite and are at least
37:09
as far as we've seen. I think it's possible.
37:12
It will happen in the first half of next year but the reality is
37:16
so many of these companies raised so much money.
37:19
Last, sure that they have decent Runway and I think our other
37:23
either by very thoughtful people.
37:25
And the other thing, I'll say these companies Have real
37:27
rugged, you know, they can sign contracts to get cash from large
37:33
Enterprises and so it's not like the cash crunch hasn't quite
37:36
happened yet and you've seen some flat rounds.
37:38
But we have, I haven't seen any down rounds in that category
37:41
yet. Excellent.
37:43
So, one thing that I want to ask before we have to jump is what
37:49
are some companies that you all looked at and you passed on and
37:53
now you wish you wouldn't have The ones that got away it's
37:59
under its I mean it's directly related to actually what you
38:01
were just talking about which is its rounds of companies that I
38:05
wish we had been in. And so like particularly before
38:09
they got marked up where the risk-reward was, I think a
38:12
little bit harder and so snorkel is one the like 200 million
38:17
dollar round. I really wish we were in that
38:20
one. Alex knows, it's an amazing
38:25
company. As I've said, the hugging face
38:27
That free to billion dollar round.
38:30
My God. It's so funny.
38:31
Ever. So we've had, I don't know how
38:33
many investors on here, every single one of them has said
38:36
hugging face, but not at the last round.
38:39
The one before that. Well, interestingly to your app
38:43
to give credit to addition. You have to because it hugging
38:47
face at that round was a really different business.
38:49
Like you had. I give a ton of credit to
38:52
addition I've because you had a substantial amount was de-risked
38:56
between. That round and the two billion
38:58
dollar round. And so good on, good on hugging
39:01
face, amazing job and amazing. For addition to have see the
39:03
potential Yeah. Actually, it was so George
39:07
Matthews came on and then set Cube.
39:09
And both of them said I just couldn't see, Transformers
39:12
becoming what they have become today.
39:14
I did not have that Vision, I didn't have that foresight.
39:17
And now that it happened. I'm like, how did I not see
39:20
this? Yak?
39:21
Yep. I think a lot of meat.
39:23
You got one that got away now. I mean I would I would
39:26
definitely do our events and snorkel from our work and on the
39:30
large on the large model side. Some of these ml as a service
39:33
companies, I mean, they're just shouldn't A number of eat
39:36
credible, incredible teams. So even on that side, not just
39:39
the envelope side. Even in the early Innings there,
39:42
still something that we feel like I've gotten away, full
39:44
incredible. Well, thank you both met me and
39:47
Joe, this is so awesome talking to you.
39:49
It's so cool to see how you're viewing the market considering
39:53
there are so many questions that are happening right now.
39:58
In this space for me, I am constantly thinking deeply about
40:03
this and wondering What is going on with the large language
40:08
models? And how does that pertain to
40:11
what is going on with ML Ops? And I really like the theories
40:16
that you put forth. It kind of makes me rest a
40:20
little easier that the ml Ops Community isn't going anywhere.
40:23
Anytime soon, the large language models are going to continue to
40:27
be a thing but that doesn't mean that it's going to turn ml Ops
40:32
into what Hadoop is today. So thank you both.
40:37
This was awesome. Awesome!
40:40
Thanks for having us. This is Skyler.
40:43
I lead machine learning and health rhythms.
40:46
If you want to stay on top of everything happening in ml.
40:49
Ops, subscribe to this podcast. Now.
40:53
Now, now, now now now Now.