Latest / Agentic Conversations (formally mlops.community) / LLMs to agents: The Beauty & Perils of Investing in GenAI // VC Panel // Agents in Production

Transcript

LLMs to agents: The Beauty & Perils of Investing in GenAI // VC Panel // Agents in Production · Agentic Conversations (formally mlops.community)

0:19
All right. Welcome everybody to stage one,
0:22
track 1 of this exciting event. Agent in Production.
0:25
I think you just witnessed an agent in production walking by
0:30
the latest model and addition. My name is Paul.
0:33
I am VP of AI at Pross, and it is my pleasure to host the next
0:39
fireside chat debate conversation with our friends,
0:44
which should be appearing behind me in a second.
0:47
We've got Sandeep, Mira and George joining us from Pross
0:51
Ventures, Red Point and Sequoia. I'm very excited to to host you
0:56
guys. Can you hear me?
0:57
Welcome to this panel. Yes, I can hear you and see you.
1:02
Fantastic. So I'm very excited to have this
1:06
conversation because I think collectively you probably have
1:10
seen all the companies, AI companies out there doing
1:13
something interesting on AI and agents.
1:15
So we're going to have a lot to talk about.
1:19
But before we do that, maybe I ask you to do a quick
1:21
introduction of yourself, what you're working on, maybe
1:24
companies you've invested in that you're excited about, and
1:27
then we can dive into the questions.
1:30
Mira, since you're the 1st on my screen, you want to kick it off.
1:35
Abuja and Paul, thanks so much for having me.
1:37
Hi everyone, I'm Mira Clark. I am a Bay Area, the semester at
1:40
Red Point. Red Point is a fund out here
1:43
that's been around for 25 years, so fairly long track record of
1:46
investing across consumer enterprise infrastructure.
1:48
And as we all know, all of these today touch a high.
1:50
So think anyone from Mistral or modal to a poolside or a layout
1:54
and everything in between. I sit within our early stage
1:57
practise, so I look at seed and Series A companies primarily
2:00
across software but also touching a little bit of
2:02
consumer land. It is hard to choose a favourite
2:06
child or at least too many children.
2:08
And so maybe mentioned the most recent deal I announced which is
2:11
in a series and a company called Trump Pulls.
2:14
Trump Pulls is building an army of agents that streamline the
2:18
administrative tasks across the construction industry.
2:21
And so they are selling into many of the largest general
2:23
contractors across the US and really saving them a lot of time
2:27
and money across their day-to-day operations.
2:30
That's fantastic. I mean a lot of those names
2:32
everybody I think watching will recognise and some of us will be
2:35
using a lot of the companies you mentioned.
2:37
So we're going to have lots to talk about.
2:39
Sandeep, since you're next on the screen, you want to follow
2:42
with a quick intro. Yeah, sure.
2:44
Happy to. Thanks Paul.
2:45
So I'm Sandeep, I look after European ventures and spend a
2:49
little bit of my time in the US for for process.
2:51
So given you're, you're moderating, I'm sure the, the
2:54
audience will know who process is based on this event.
2:57
We are global investors. We invest across categories,
3:02
across sectors and between early and late stage.
3:04
I spend most of my time focused on our early stage investment
3:08
practise. And I some of the some of the
3:11
more recent investments that we've done in the, the space
3:14
that would be relevant for the audience or companies like Cordy
3:17
in the healthcare space, Martian in the infrastructure space and
3:20
also Emma in the, in the agent space.
3:23
But yeah, very excited to be here.
3:25
Great. Thanks for joining, George.
3:28
Yeah, hi everybody. I'm George.
3:29
You can probably tell my accent. I'm British based in London, a
3:32
Sequoia that's kind of novel because the open office in
3:34
London about four years ago. So we do everything from precede
3:38
to pre IPO investing. You know, we were only investors
3:40
in companies like NVIDIA, Apple, Google, organisations that have
3:43
gone on to be pretty Seminole and obviously the AI space.
3:45
I mean, out of Europe, we're investing as one team from
3:48
Israel to to San Francisco companies we've invested in
3:52
recently, a business like Anterior, the focus is on
3:54
automating claims management and prior authorization in the US
3:58
healthcare market. I like it because they focus on
4:00
the big problem, how to cut down these hundred page documents to
4:03
something you can process in minutes, but also the small
4:05
details of automating, calling a hospital to check on, you know,
4:08
a missing detail in a form, kind of things like that.
4:10
So big and small to unlock an opportunity.
4:13
Cool. Well, that I mean, as I
4:15
mentioned, I think one of the things that you collectively
4:18
have seen is all of the AI companies that are doing
4:21
anything interesting will have come by any of you, all of you
4:24
potentially. The theme of the conference is
4:27
agents in production. So a lot of folks who are
4:29
watching are including, you know, our teams of pros S are
4:32
working on, you know, finding ways to put these agents into
4:35
real systems and to create businesses and experiences
4:37
around those. Could you maybe share what
4:40
excites you about AI agents? What are the things that you
4:43
know are ready today for production that you're already
4:46
seeing being part of products of businesses you've invested in
4:51
and things that are generally exciting for you for the next 12
4:53
months? And I'll leave it to any of you
4:56
to raise your hand. Yeah, go ahead, Mia, not.
4:59
The way you start up, I think you know, historically, right,
5:03
you think about Main Street and these have been nine to five
5:06
businesses, 96 businesses, but a lot of life happens kind of
5:09
outside of that window. I think as we've moved online,
5:12
expectations from consumers have really shifted to this always on
5:16
mentality of where is the web and shouldn't it be answering my
5:19
questions. What I think is really
5:20
interesting with these agents is this 24 by 7 support that
5:24
they're able to offer. We have companies in the
5:26
portfolio that are answering phone calls, scheduling
5:29
appointments, kind of across healthcare insurance, kind of
5:33
PNC insurance, you name it. And to me that it's really
5:36
compelling. It's not this whole, you know,
5:38
scary, daunting labour replacement story that I think
5:41
everyone is really excited to tell, but really this support
5:44
and that I don't think there's any, you know, customer success
5:46
agent that wants to be online at 1:00 AM in the morning answering
5:50
this new OPS question. And so I think the ability for
5:52
these agents to come in and answer a lot of these more or
5:56
perform a lot of more, these more rote tasks to us is really
5:59
exciting because it leaves, I would say the customer facing
6:01
roles to really do what they do that which is deeply understand
6:05
consumer pain points, build trust with them, help them more
6:08
from the navigation perspective, it's really bringing the
6:11
hospitality aspect back to a lot of these industries.
6:14
And so as we think about really enhancing the consumer
6:16
experiences, whether that is in their dental office, down the
6:20
street at the hardware store or what have you, to me, that is
6:23
really exciting because I think we talk a lot about tech making
6:26
consumers lives better, making people's lives better, both
6:29
personally and professionally. But I think Asians are really
6:31
going a long way to streamline that as you think about kind of
6:34
taking some of the work off of people's plates that they
6:37
candidly don't want to be doing. Just to build on some of the
6:42
things that Mira said that one of the areas that I found quite
6:45
interesting and Paul, you know, here at Process, we spend a lot
6:47
of time looking at some of the customer support type agents
6:51
that exist. One of the things that I've been
6:53
spending a little bit of time looking into is as it pertains
6:58
to customer support, the concept of agent is not new, generally
7:02
speaking. I mean, the concept of dealing
7:03
of typing into somebody and trying to, to, to, to interact
7:06
with somebody on the other side is not new.
7:09
And for me as a customer, if I take an example of a travel
7:12
booking and I'm dealing with an online agent, I know that I'm
7:16
dealing with an online agent. So all I try and do is prompt
7:18
them in ways to try and get to a human, even if it may be 1:00 AM
7:22
or 2:00 AM at some place elsewhere.
7:25
And what I'm actually interested in spending time on is how to
7:28
reshift the customer's perception.
7:30
So in that case, me to for me to understand that actually, if I
7:34
prompt the agent better, I will actually get more help versus I
7:39
don't need to try and find the quickest hack to redirect myself
7:43
to a a human. And that's actually an area
7:46
where I've been spending a little bit of time where
7:49
customer, because changing consumer behaviours is
7:52
incredibly difficult and what tools exist out there, people
7:57
are building, especially as these models get more powerful,
7:59
that can change customer behaviour where we're used to
8:02
trying to just get a human on the phone as quick as possible.
8:06
Maybe just pick up a final piece here.
8:08
And I think you know what ties all that together is this idea
8:10
that you reduce the marginal cost to deliver some service.
8:13
I think there are plenty of examples of industries which
8:15
have been very resistant to digitization, places where I
8:18
think we would have assumed maybe in the marketplace era,
8:20
marketplaces might emerge in both consumer action and in B2B
8:23
context. I think we're seeing this
8:24
phenomenon of like 0 cost bids a little bit to mirrors point of
8:27
being able to go out 24/7. There are so many industries
8:30
that are heavily brokered that just intermediate by a bunch of
8:33
these different partners. In today's world, you can
8:35
actually use agents to go out and get quotes from millions of
8:37
these different people 24/7 in many languages to build together
8:41
a lot of the component parts that look like a marketplace
8:43
experience to digitise a bunch of those supply chains for the
8:46
first time. And that's the first opportunity
8:48
we've seen really to unlock that from the heydays of B2B
8:50
marketplaces to now. Yeah, that's, that's exciting.
8:55
I think one of the things that caught my attention that, you
9:00
know, we'll have other speakers discuss today is multiple agents
9:04
working together. I think Mira, you alluded to it
9:06
as a swarm of agents. And can you sort of share, you
9:12
know, what you've seen in in, in, in the companies you've
9:15
spoken to or you've invested in the one that you mentioned that
9:18
actually uses this in production?
9:19
How does that work? What's the, you know, experience
9:22
to the users of that particular product?
9:26
Yeah. I think first you have to
9:27
address the why there are multiple versus kind of 1.
9:31
And I think that comes down to two primary dynamics. 1 is agent
9:35
capabilities, which are good today.
9:37
They're not perfect. And so it's kind of thinking
9:39
about can you build a better model if it is more specialised,
9:44
we're with greater guardrails around it in terms of the tasks
9:46
it's being asked. I think equally as importantly
9:48
though is that human psychology perspective I think to some
9:51
deeps comments on kind of our nature to prompt and prompt and
9:55
prompt till we get a human. Similarly, as you think about
9:58
agents becoming these new employees, if you will.
10:01
And it's interesting when we talk to a lot of the customers,
10:05
they have named these agents and it's they are talking about
10:07
Sally, they are talking about Annie and Sally does a different
10:10
job with Annie in the office. But in their mind, they don't
10:14
have employees, right? That both, you know, check
10:16
someone's eyes and file the bills and then sell you the
10:19
glasses or what have you. These are independent workers.
10:21
And so I think some of what we're seeing is candidly just a
10:24
reaction to how humans work and agents, you know, taking this
10:28
gradual progression and inserting themselves into the
10:30
way that work gets done in a way that isn't overly disruptive for
10:35
the consumers. In terms of what that means with
10:38
these companies, I think a lot of it also comes down to this
10:43
almost element of AI overwhelming them and this idea
10:48
of, OK, if I open a ChatGPT window or a Gemini window, I can
10:52
ask to think anything and everything.
10:54
And I don't know where to start by starting with these kind of
10:57
individual agents. Companies can say, let's take a
11:01
general contractor for example. I just want an agent to answer
11:04
my questions surrounding the specs in this room.
11:06
I just want an agent that's going to update scheduling.
11:09
And so you can roll these out kind of sequentially versus
11:12
overwhelming someone with the you have every tool in the world
11:15
at your fingertips. Good luck using it.
11:17
Because again, I think where we're seeing AI penetration
11:20
today in a really impressive and meaningful way is in a lot of
11:23
these industries that have been fairly under penetrated by tech
11:25
historically. And so the willingness to change
11:29
or be like eagerness to adopt is just quite different than if
11:32
you're selling into a developer sitting in San Francisco or
11:35
London or New York. But George Sandeep, I don't know
11:38
if you guys are seeing things that are dramatically different.
11:44
I mean, I think you put it beautifully.
11:45
I think this idea of finding the right form factor for it to
11:48
start to be activated in companies I think is critical.
11:50
I think the nice thing is for a lot of those processes,
11:52
playbooks exist that as you said, you know, humans already
11:55
follow day-to-day. And I think you can start with
11:57
something that is relatively granular, some like accessible
11:59
atomic unit. But I think what's really
12:01
interesting is you're starting to see agents having quite
12:03
sophisticated fall backs. Like we have instances now where
12:06
a company like Expo and our portfolio that focuses on
12:08
automating pen testing, right? Pen testing is in some ways kind
12:11
of an iterative exercise. You do it typically, you know,
12:14
pretty infrequently as a company because it's expensive, it's
12:16
quite services driven. Now you can start to change the
12:18
relationship you have with that kind of service provider if
12:20
you're using something like an agent.
12:22
And to run that process, you obviously need to continue to
12:24
try different approaches on the same infrastructure over time as
12:27
different threat factors evolve and it just changes the way you
12:30
think about consuming that in service.
12:34
I I think just the only thing that I would add again, building
12:38
on something Mira said, is, is that what exists today isn't
12:41
perfect. And the question of what level
12:44
of perfection or how accurate do we need to have as you have, you
12:49
know, maybe one that's 90% and the second that's 70% and the
12:52
3rd that's 60%, what level of accuracy?
12:55
And I think it's predicated upon what industry you're servicing
12:58
and where you have the ability to accept a lower level of
13:01
accuracy. So I think that's one thing that
13:03
we're spending a lot of time trying to think about the the
13:08
the cost of inaccuracy and inaccuracy relative to what
13:11
already exists in an industry. When you speak to founders that
13:18
are building companies around this next generation of, of
13:22
generative AI and, and in particular agentic systems,
13:25
there's a lot, I mean, like, you know, when we build agent
13:28
systems at pros, there's a lot of that is, you know, at this
13:31
moment, like you said, it's early stages trying to sort of
13:34
get it to work a first time. Then measure if you're at the
13:36
60% mark or the 99th percent mark, because we do see that
13:41
there's a path to getting there. In some cases, it'll be an
13:44
expensive and long road. In other cases, just a matter
13:47
of, you know, months and we'll get there and, and integrations
13:49
and, and making engineering work.
13:51
When you speak to founders and teams that that that you know
13:54
are working on these kinds of promises, what are the things
13:57
you look out for to assess whether you think this, you
14:01
know, this is the right team you want to be partnering with from
14:05
an investment point of view? I'm happy to start on that.
14:09
Well, I think, I mean, there are some warnings maybe from some
14:13
previous, you know, incredible technologies like autonomous
14:15
vehicles, right? I think there was an assumption
14:17
that the performance criteria was as good as humans And like
14:20
it's broadly been at that level of performance for a while.
14:22
But it turns out it's kind of humans plus risk appetite, plus
14:25
regulation, plus a bunch of different barriers.
14:27
And we've seen over the last 15 years incredible work, but huge
14:29
amounts of investment. And obviously, we're still yet
14:32
to see, you know, autonomous vehicles outside the San
14:33
Francisco, you know, on our roads.
14:36
I think when we spend time assessing that opportunity, I
14:38
mean, speaking to customers is critical, right?
14:40
In some ways that matters the same As for any founder in any
14:43
company, you know, understanding the voice of your customer.
14:45
But here it matters because you're dealing with, you know,
14:47
an existing process typically in many cases that has some kind of
14:51
performance expectation that might be accuracy, precision, it
14:53
might also be cost, right? So the benchmarks are typically
14:56
quite easy to understand. There is some level of
14:58
innovation budget that will come from a lot of companies in
15:00
today's world towards exploring opportunities around that that
15:04
will take you so far. And I think in some ways, you
15:06
know a lot of companies that have been successful have found
15:08
a clever way to access those budgets very, very quickly to be
15:11
able to get early validation in a particular deployment.
15:14
I think the really crucial thing is actually to speak to people
15:16
that are building the foundational infrastructure
15:18
because for some of these iterations of these models, the
15:21
performance might not be sufficient in some ways.
15:23
History has shown us it takes exponentially longer, it is
15:25
exponentially more expensive for every incremental percentage
15:29
point of performance you eke out.
15:31
And in some cases, you know, that means we might be able to
15:33
get there and we're six months out, 18 months out, and in some
15:36
cases it might be years, right? And the timing you have is an
15:38
early stage company that really matters.
15:43
I think building on one thing that George said, it's one of
15:48
the things that we've seen, at least as it pertains to
15:50
companies that are servicing the enterprise, is there are
15:55
organisations which were so, so early stage companies, early
15:58
stage companies that will work with enterprises as design
16:00
partners and kind of tapping into that innovation budget.
16:02
And then there will be companies that sign up customers and they
16:05
will go through their entire procurement process, which will
16:08
typically be around, you know, varies for every organisation,
16:12
let's just call it 100K type ACV.
16:14
And once you go through those processes, they've really been
16:18
tried and tested to ensure that it fits, you know, regulation
16:22
compliance is, is there a clear need?
16:25
Because there's been there large enterprises, it's, it's very
16:28
complex for a company to go to through, through a large
16:32
organisation procurement process.
16:33
So we will take for for earlier stage companies, we'll take some
16:37
validation through through companies that have gone through
16:39
that. You, you mentioned George, that,
16:44
you know, getting up that sort of accuracy curve and sort of
16:47
getting to whatever that, that, you know, they're required
16:50
threshold is if it's 90% or 9095% accuracy.
16:55
I don't know if, you know, there's been some noise recently
16:59
on, on, on the fact that we may be plateauing in terms of
17:03
performance of some of these models where some of the big
17:06
labs are, you know, allegedly seeing a sort of diminishing
17:10
returns on investment in getting these models to the required
17:14
accuracy levels to get them to execute the tasks with a certain
17:19
amount of improvement, right reasoning capabilities, planning
17:22
and so on. Is that something that you are
17:26
in any way seeing or noticing or paying attention to?
17:29
Or is there sufficient runway, let's say on the existing
17:33
capabilities that we already have out there today with
17:36
existing models to build, to build on top of and you're not
17:40
not necessarily dependent on on on the success of let's say the
17:45
next generation of models to get those businesses to to become
17:49
viable. I.
17:51
Think, I think it's a wonderful question.
17:52
There's so many moving pieces as part of that ecosystem though,
17:55
right? You have obviously the
17:56
performance of the base models and how they've been evolving.
17:58
You have this incredible proliferation of open source
18:01
models, you have multi modality showing up in ways that are
18:04
quite unexpected. I mean, who would have thought
18:07
that actually some of the biggest improvements you see in
18:08
computer vision deployed in industrial manufacturing
18:11
actually come from prompt engineering right in LMS being
18:14
overlaid on top of, you know, traditional vision models.
18:17
That's something that I definitely didn't foresee and
18:18
it's something we hear from a lot of our founders being a kind
18:21
of surprising expected performance for Uplift.
18:23
Second thing though is all the tooling, like the data pipelines
18:25
around it just continue to improve immeasurably.
18:28
So the ability for a wider pool of people who are potentially
18:31
less specialised in one particular domain to have more
18:33
strength to wrangle the performance of some underlying
18:36
model has just improved exponentially over the last 18
18:38
months. So I think if you look at it and
18:40
the base model performance is one thing, but all of that
18:42
ecosystem of tooling around it means we've got a long way to
18:44
go. Yeah, maybe to double click on
18:47
that. I think the data point is a
18:48
particularly important one. I think it's easy to look at
18:52
certain sectors where you say, OK, open AI as about as good of
18:58
performance as anyone else because they have access to all
19:00
the data. There are many industries out
19:02
there, right, where this data is not publicly available.
19:05
And so kind of the larger model providers haven't been able to
19:08
train on it. They think Healthcare is a very
19:09
obvious one, construction is a very obvious one, but there are
19:12
quite a few others. And so as we sort of think about
19:14
the industry that we're going after, those are particularly
19:17
interesting for us because I think that's where you still can
19:19
drive a pretty meaningful delta by getting more data kind of
19:22
into the system some under train on.
19:26
So if, if, if maybe taking on a little bit of an experiment
19:29
based on the previous conversation, if you, if you
19:31
were listening to some of the, the, the, the, the, the fireside
19:34
tie between Euro and and Tom from Hugging Face where you
19:39
know, the same was made. Well, if you just stop all
19:41
technology development today, right?
19:43
We still have 15 years of technological dividend to build
19:47
viable businesses on top of, you know, would you agree with that?
19:52
And, and the businesses you're, you're talking to the founders,
19:55
the things that they're proposing, to what extent are
19:57
they dependent on future breakthroughs versus they're
20:01
actually taking what we already have and just building
20:04
businesses on top of that? Because the, the, the truth is
20:06
that the, the rate of acceleration in the last, you
20:08
know, couple of years has been so high that it also takes time
20:11
to business, build a business proposition and a product on top
20:14
of that. So you could take the physician.
20:15
Well, we don't need more technology breakthroughs to
20:18
continue for at least, I mean 15 years.
20:20
For a long time, but let's say 5 to 10 years and build entirely
20:23
new propositions around that. Would you agree?
20:27
Yeah, I wish I'd been listening to more of that conversation.
20:29
It's earlier in San Francisco, so I missed it in some ways.
20:32
I would take the other side of that argument and it's not from
20:34
a performance perspective where I have concerns for some of the
20:37
businesses we're seeing really comes up with the cost.
20:39
And so, OK, performance improvement aside, I think
20:43
there's a question of like when do these business models shift
20:45
to being long term viable? And given the cost structure
20:49
today, many if not most of them are not.
20:55
So that concerns me. But George Sandeep, I don't know
20:59
if you would think about it differently.
21:05
Yeah. It's, it's, it's interesting.
21:06
I think that for us, there are definitely those businesses that
21:12
currently have an incredibly high cost base or the unit
21:16
economics don't quite make sense.
21:18
So they're predicated upon models either becoming cheaper.
21:23
That being said, I think for some of the businesses that
21:26
we've seen or we've invested in, particularly those Tamira's
21:31
earlier point that have been built on top of their own
21:33
proprietary data, those have the ability to continue to
21:38
accelerate, especially because there's so much friction to get
21:41
into some of those industries. So there continues to be an
21:44
opportunity to penetrate into that market.
21:48
But it's, yeah, it's, it's, it's definitely an interesting, it's
21:50
definitely an interesting debate because currently to Tamira's
21:53
point, there's just a lot of things that are prohibitively
21:55
expensive. So I'm glad I'm not a fortune
21:57
teller. And maybe just to pull, I guess,
22:00
a slightly different thread. AI doesn't give you
22:02
distribution, right? It might give you an incredible
22:04
value prop that allows you to have some wage into some large
22:07
market opportunity. But broadly like you still need
22:09
to keep pushing, right, if you want to build a platform and to
22:11
kind of endure in the long run. And I think it really matters
22:14
depending on the dynamics of your market.
22:16
I mean, you come up against in some cases legacy incumbents who
22:18
are incredibly UN automated, in some cases pretty poorly
22:21
digitised and that's just right automation and disruption from a
22:25
company. I think you also need to think
22:26
about, you know, who are you really competing against and all
22:28
that sort of next generation players that are 20/18/2020
22:32
vintage companies with high product velocity and already
22:34
pretty widespread distribution. And thinking about the right
22:37
strategy to take on a business like that is very different from
22:39
like a 1980s, you know, legacy incoming.
22:42
Right now it takes, it takes more than an LM to build a
22:44
business, right? So you'll need to put it in the
22:46
market and make money. It's just to build on the
22:50
economics piece here. I think, you know, we'll talk
22:53
about that I think in some of the sessions later.
22:55
But we recently looked at the cost per token generated over
23:01
the last two years, right? So we've been building in the
23:03
space for for several years now and looked at what the drop had
23:08
been. And so we took the baseline of
23:10
the most powerful model exactly 2 years ago, November 2022,
23:13
which was GPT 3DA Vinci and it cost $20 per million tokens to
23:20
generate tokens. And now that has dropped to the,
23:25
the, the equivalent model in terms of performance.
23:26
The closest GPD 4 O, which costs 0.24 cents for that same amount
23:31
of million tokens. So you're talking about A2 order
23:34
of magnitude drop, 98% drop in about two years.
23:37
At the same time, you know, as we build our own agents,
23:40
obviously you know, we're, we're, we're testing and
23:42
experimenting with a lot of stuff inside pros as we, we've
23:45
been looking at the costs over the last couple of months per
23:48
question that our agents are answering.
23:50
And while you see a tremendous drop in the actual cost per
23:53
token because we're building a genetic systems, they are
23:57
calling many of the models many more times, they're requiring
24:00
many more tokens that per token are more expensive as well.
24:04
And so actually the cost per question that we're answering
24:08
went up right about 25%, right. So it's not a double digit, you
24:13
know, 2 orders of magnitude increase.
24:14
But so it's sort of an interesting and very hard to
24:17
budget for world because on the one hand the cost come down per
24:21
token, but you know, you're requiring more tokens to do
24:24
things, especially in an agentic future.
24:26
And so as we talk about ages in production, you know, getting it
24:28
sort of to become viable is, is, is is hard and definitely hard
24:32
to predict. But maybe to build on that idea.
24:35
So do you see a fundamental shift in the underlying business
24:42
models? Because when they are predicated
24:44
on calling large language models, the marginal cost to
24:50
create value to have an interaction with the user to
24:54
drive engagement is not 0, right.
24:57
So the cost of engagement or let's say the volume of
25:02
engagement is directly correlated with the cost because
25:05
every time you have a user interacting with your product,
25:07
you're obviously paying for, for those tokens in some way.
25:12
Do you see the economics changing of, of, of, of the
25:16
businesses you invest in? Are these variable marginal cost
25:20
indeed now you know, scaling or harder to scale on top of?
25:26
Are the sort of gross margins changing for the for business
25:28
that try and scale? How do you think about that?
25:36
Happy to start. I think I mean, in many cases
25:39
these companies with these very ambitious founders are chasing
25:41
huge prizes. And we touched on a little bit
25:43
earlier, but like if you have an incumbent like Salesforce,
25:46
right, the best place to take on a business like Salesforce isn't
25:48
probably the price like Salesforce deliver a product
25:50
that looks and sounds like Salesforce.
25:52
It's probably to think about, you know, an incumbent like that
25:54
is as a business built on recurring revenue, right,
25:56
independent of usage. In some ways, I mean, yes, see
25:58
base, but independent of usage. The opportunity in a lot of
26:01
these go to market tools now is selling outcomes, right?
26:04
Not necessarily selling a system of record, but really a system
26:06
of engagement that has completely different economics
26:08
for you as a business when you use it.
26:10
Because again, you're getting much more return on investment
26:12
on the human labour because of these very high levels of
26:15
automation. I think if you can get to that
26:17
level of alignment, I mean, you have a different take rate,
26:19
right? If you think about the
26:20
percentage of revenue that you're able to sort of claim, if
26:22
you get to something that is outcome based, we see kind of 15
26:25
to 20% ish or most of the net revenue you're generating, but
26:28
it's much lower in the funnel because you're basically taking
26:30
accountability, being able to do that.
26:33
So I think actually going over an outcome like that, which is
26:35
quite novel and really shakes up the industry in which, you know,
26:39
your incumbent will not be able to follow you because in many
26:41
cases they're a public company that ties those expectations.
26:44
You know, the like, I think that really allows you to create some
26:46
noise in your market. And crucially, you know, if
26:48
you're aligning your cost base with that, it provides some
26:50
scalability as well. I think building on what George
26:54
said and also coupling it with a comment that he made in the for
26:57
the previous question, it also to some extent depends on your
27:02
competition. So we have a portfolio company
27:04
called Superside, which provides creative services to marketing
27:08
teams. And historically speaking, their
27:09
competition were advertising agencies and advertising
27:13
agencies they traditionally operate at, you know, I mean,
27:18
2030% gross margins, right? 40% gross.
27:21
These are high services. That's high services businesses.
27:23
Now you have a lot of these design tools that have allowed
27:26
people to iterate and create thousands of ideas and a short
27:29
amount of time together with creative professionals.
27:32
But instead of having, you know, 4 individuals, you can do it
27:36
with one individual and still have the same output, the same
27:39
outcome to, to, to use George's word.
27:42
And when you look at it, when you look at those comparisons,
27:46
it doesn't matter how many tokens you know you have.
27:49
Like you're, you're thinking about it in terms of the, the,
27:51
the relative benchmark is, is very different because the, the
27:55
competition or the, the, the expectation of how much you
27:59
would pay is, is also very different.
28:00
Now, how does that change over time as your customers become
28:04
more attuned to the fact that, you know, their, their mind
28:07
shift based on what pricing should look like that we'll see.
28:11
But I think at the moment we're still in the first with the
28:14
initial innings of trying, using, testing, paying for some
28:18
of these tools that I think that the the competition is more so
28:22
with the incumbent and therefore higher willingness to, to, to
28:25
try and and to pay for it. Yeah, that's a great point.
28:29
I mean, I think in general, regardless of let's say where
28:34
the take rate comes from and the structure of the business, the
28:37
unit of intelligence is going to get cheaper and will trend to 0.
28:40
So at some point, just maybe maybe just be a matter of time
28:43
until it becomes viable for the businesses that are that are
28:46
building on top of these things. Now I'm I'm going to bring in a
28:49
question from Ping LAN who's sharing a question on gradual
28:55
for VCs. Would you regard start-ups
28:57
building AI agents as SAS companies or as infrastructure
29:00
companies? Like that's a trillion dollar
29:05
question. It depends who you ask.
29:08
And I think the answer is candidly evolving.
29:10
I think maybe just looking at Red Point anecdotally, it's a
29:13
fund we historically had call it majors and minors as investors.
29:17
And some of us would look at consumers, some would look at
29:19
SAS, some would look at infrastructure.
29:21
I would say over the past, call it 12 to 18 months, that model
29:25
has entirely been blown up. And traditionally consumer
29:28
investors are looking at infrastructure because it's
29:30
infrastructure API businesses, power consumer experiences.
29:34
You're looking at infrastructure investors looking at, yeah, a
29:36
lot of these consumer companies because it's they come within
29:40
network as it relates to kind of the researchers that are
29:42
building them. And so I think you were seeing
29:45
at many funds, maybe not all funds, but this reversion to
29:47
more of this like generalist mindset, again, coming with
29:52
opinions or subject matter expertise in specific domains or
29:56
cost specific themes in terms of taking a point of view on where
29:59
the world is headed. But I feel like the business
30:01
model itself is something that investors today are a little bit
30:06
less focused on, at least from my vantage point or from where
30:10
I'm spending time. I, I, I think from, from our
30:15
perspective, it's who is your end customer and how used to
30:21
buying software are they? So, so and, and how like what
30:26
budget are you tapping into? What's their alternative?
30:28
So if you're used to, if you're selling something to developers
30:31
that were typically speaking used to buying software, used to
30:34
buying products like this versus selling to some something to
30:36
somebody in the construction industry, to use the trunk tools
30:38
example, those are very, very different and those were have
30:41
very different sales motions or to use George's commentary, they
30:44
have very different distribution mechanisms.
30:46
So that's kind of how we try and think about this.
30:48
How are you acquiring these customers?
30:50
What budgets are you tapping into versus trying to use a SAS
30:53
or an infrastructure type monitor to it?
30:58
Any anything to add? George, I see you're nodding.
31:02
No, nothing to add. I think that's exactly.
31:04
Right, great. So we're almost out of time, but
31:06
I since we've been talking about predictions and the way the
31:10
world's going to look in a year, maybe we can do a closing round
31:14
or a or, or a wild prediction that you may want to share about
31:18
the future of AI companies and the AI industry for all of the
31:23
folks watching to keep an eye on for the next 12 months.
31:29
I'm happy to start. I mean, I come from a fintech
31:31
background. There's a big problem statement
31:33
about the last 15 years has been designed around 2 factor
31:36
authentication and thinking about protecting people from
31:39
having their car details stolen by people making payments
31:41
without their consent. In the future, it's inevitable
31:44
that we'll have agents paying agents.
31:45
We'll have agents paying humans probably to procure work for
31:48
more complex workflows. And we need to find the right
31:51
structure, the right infrastructure, but also the
31:52
right regulation to enable that. Maybe double clicking on what
31:58
will change, I think we are very excited and intrigued by how the
32:02
atomic unit of interaction will evolve.
32:04
Today. Much of our back and forth,
32:07
these agents are through chat bots.
32:09
I think we're seeing a lot of voice popping up, but whether it
32:12
is avatars, infinite canvases or something else entirely, I think
32:18
we're seeing a lot of experimentation and we are very
32:20
excited to see, yeah, how those interactions will change over
32:23
the coming quarters. I think one thing for us being,
32:28
you know, historic, historically a consumer focused investor, I
32:32
think the, or from an ecommerce perspective is, is really
32:36
looking at the future of, of ecommerce and, and shopping and,
32:39
and looking at the next, the next wave of, of personal
32:41
shopping assistance and personal digital assistance.
32:44
And I, I anticipate that that will come.
32:46
It probably won't be, I think it was Bill Gates that commented
32:49
that, you know, whoever creates the personal assistant will
32:51
build the next trillion dollar company.
32:52
But I, I, I don't know if it's going to be that far, at least
32:55
in the time horizon that I have in my mind, But I think that
32:59
personalised, very, very attentive shopping assistance is
33:02
something that could come in the, in the near future.
33:05
Very exciting. Well, I know that for a fact
33:08
that a lot of the speakers we have for the rest of the day are
33:10
actually working on making some of your predictions come true.
33:14
So hopefully they're successful and that becomes reality and
33:17
we'll be able to check in, in 12 months.
33:20
In the meantime, thank you very much, Mayor George Sandeep for
33:23
joining us today.