Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:00
George Matthew is a managing director
0:02
at Inside Partners, where he invested
0:04
in Weights & Biases, Jasper, and
0:06
others. He has over
0:09
20 years of experience developing
0:11
high-growth technology startups, including most
0:13
recently being CEO at Kespry.
0:16
George joins the podcast to talk
0:18
about his path to becoming an
0:20
investor, his data-first thesis about investment,
0:22
the AI business landscape, his book
0:24
recommendations, and more. This
0:26
episode of Software Engineering Daily is
0:28
hosted by Jocelyn Bern-Hul. Check
0:31
the show notes for more information on Jocelyn's work
0:33
and where to find her. George
0:48
Matthew, welcome to Software Engineering
0:50
Daily. Great Jocelyn, great to
0:52
meet you as well. Pleasure to be here. We're
0:55
really excited to spend some time talking with
0:57
you personally, as well as for our show's
0:59
audience, our technical audience. There's
1:01
nothing hotter to talk about than AI
1:03
or data, I think. You have been
1:06
a storied investor in this space, right?
1:09
So many investments so early on in Weights
1:11
& Biases, Excel Data, which is one of
1:13
my favorites, Jasper. We're
1:15
really excited to have you come and talk with us. But
1:17
before we kind of get into the technologies that you're interested
1:20
in, I want to learn a little bit about you. Are
1:22
you just a young little
1:24
boy out in the playground thinking, like, one
1:26
day I'll be an investor in this space? Or
1:28
how did you come to realize this is what
1:30
you wanted to do? Yeah, I
1:33
would say it certainly wasn't a portion of
1:35
my journey that I thought I was going to
1:38
be an investor when I was a young
1:40
kid in the playground. I think originally my intent
1:42
when I was that young was probably to
1:44
be a pilot. When I
1:46
got a little bit more older, I realized that
1:48
that wasn't going to be the right calling for
1:50
me. And then I really
1:52
wanted to be a medical doctor. And
1:55
as I was studying actually for my MCATs
1:58
my junior year, I came across across
2:00
the NCSA mosaic browser, which a
2:02
certain individual named Mark Andreessen had
2:04
shipped out of the
2:06
supercomputing facility at the University of Illinois.
2:09
And I would say my heart's kept a beat
2:11
back in 1996 and really
2:13
wanted to understand what was going on with
2:16
the emergence of a tool in
2:18
technology like the World Wide Web
2:20
and the modern date web browser
2:22
in those days was pretty early
2:24
in the overall journey. And
2:27
I just got really excited by that and followed a
2:29
lot of where my passion was there to
2:31
go find a job in California to
2:33
work in an early stage company
2:35
that was doing the first generation
2:38
of e-commerce applications
2:41
in the mid 90s working over
2:43
emerging technologies around the World Wide
2:45
Web. You know, it's such an interesting
2:47
moment that we're experiencing right now because I
2:49
do think, you know, my two big sea
2:51
changes that I've experienced, right, was the ad
2:53
to the internet came in, replaced client server
2:55
and then of course open source. These were
2:58
like huge moments, right, that felt like second
3:00
industrial revolutions and they're paling in comparison to
3:02
some of the excitement around generative AI at
3:04
the moment. Help me understand a
3:06
little bit about what is your perspective on investing
3:08
in generative AI right now because there's a lot
3:10
of interest on the consumer side, but
3:13
it's not clear which startups and what their
3:15
strategy should be. A
3:18
lot of my perspective from investing in
3:20
generative AI has really come from being
3:22
originally a builder and someone who's been
3:24
in around data and AI systems for
3:27
a good part of a decade and
3:29
a half. And
3:31
as I started to see this current
3:33
generation of systems continue to evolve, it
3:36
was just very clear in the enterprise
3:38
that the modern data stack itself was
3:40
just an important underpinning
3:42
for how this next generation
3:44
of AI-based applications and systems would
3:46
emerge. And so I really started to
3:49
look at where the modern data stack
3:51
was really headed when it comes
3:53
to all the tools that were required,
3:56
not only in data management, but data
3:58
catalogs, data orchestration, data development. observability
4:01
and really built my thesis on
4:03
just a data first view of
4:05
how this next generation of AI
4:07
and machine learning based systems would
4:10
emerge. And so naturally from that
4:12
layer of the modern data stack,
4:15
started to look at machine learning operations
4:17
and MLOps, right as a necessary
4:19
set of tools that would be required
4:21
for a machine learning practitioner to be
4:23
able to build a model
4:25
and be able to bring those models
4:28
into production. And so one
4:30
of my first investments joining insight about
4:32
three years ago was in a company
4:34
called points and biases, which became, of
4:36
course, one of the de facto tools
4:39
for all the experiment tracking, hyperparameter tuning
4:41
version controls that were necessary for a
4:43
machine learning practitioner, to be able to
4:45
get their job done and effectively build
4:47
models and bring them into production. I
4:50
think as you started to see the
4:52
evolution of MLOps, then became clear that
4:54
there was this opportunity to take the
4:57
data that was coming
4:59
from these modern data stacks and
5:02
merge them properly with the AI
5:04
systems, which are of course now
5:06
L ones and transformer based AI
5:08
systems and build these next generation
5:10
of generative AI applications. And of
5:13
course, one of our investments about
5:15
a year and a half ago
5:17
was a company called Jasper, which
5:19
is building a generative AI application
5:21
for how content writing could be
5:24
more naturally be done with a
5:26
copilot to support how contract writers
5:28
work on a day in day out
5:30
basis. So the thesis for me was
5:32
always starting with a data substrate, which
5:34
at least in my case, in the
5:36
enterprise was very much targeted towards the
5:39
modern data stack itself, and building upon
5:41
itself to go into MLOps and more
5:43
recently into generative AI applications. There
5:45
is a blog post on the website about the
5:47
difference between MLOps and LLMOps. And you know, I
5:50
interviewed Chris Nagati from Fiddler on the show as
5:52
well. So I spent a lot of time in
5:54
that MLOps space. Help me understand what
5:56
the well, maybe we should sort of say like
5:58
what we say this MLOps or LLM ops. Let's
6:00
just set a business case. Why do people care
6:02
about that? And then let's kind of compare and
6:04
contrast the two. One
6:06
of the things that we started to
6:09
notice with emergence of this category of
6:11
machine learning that was
6:13
really focused on transformers and large
6:15
language models was that you were
6:18
shifting from a model-centric
6:20
experience to getting
6:22
a model into production and getting success
6:25
out of the outcomes of model
6:27
production to almost a data-centric world
6:29
of getting models into production. And
6:32
the reason I mentioned that is
6:34
that the models weren't tremendously changing
6:36
in terms of what was going into the
6:38
algorithm, the models per se. You
6:40
might see some shifts in the weights
6:42
and biases surrounding a model, but less
6:45
so around the underpinnings of the
6:47
model itself. What was changing pretty dramatically was
6:50
the data that was coming into the models.
6:52
And this is where we
6:54
started to see this current
6:57
generation of transformers and LLMs
7:00
in particular emerged that the more
7:02
data that you brought into a
7:04
large language model, the more
7:06
human-like and reasoning you started to
7:08
see the models perform in terms
7:10
of when they were put into
7:12
a general purpose scenario
7:15
like what you saw with
7:17
OpenAI's chat GPT or even
7:19
domain-specific models that emerged, for
7:21
instance, like Bloomberg GPT. And
7:24
in all these cases, what became
7:26
clear was the techniques that you
7:29
were using to go build a
7:31
LLM was going to be similar,
7:33
but in some ways quite different
7:36
from the techniques that you would
7:38
be using to build a
7:40
model for the purposes of
7:43
what would be a traditional computer
7:47
vision, for instance, or around a
7:49
predictive statistical analysis that you might
7:51
be doing. So we started to
7:53
think about this at InSight. Where
7:56
would be the delineation
7:58
of where to and
8:00
platforms would evolve for the
8:02
systems of record versus
8:05
the systems of prediction
8:08
versus the systems of creation. And so
8:10
we've had a lot of history around
8:12
the systems of record and the systems
8:14
of analysis or prediction, but we haven't
8:17
had a lot of history in terms
8:19
of really understanding what these sort of
8:21
new systems of creation were. And
8:23
I think that's where we put a
8:25
lot of our attention more recently in
8:28
what's basically known as LM
8:30
ops versus ML ops. And
8:33
in that regard, as we try to
8:35
navigate what those differences were
8:37
and those similarities were, there were
8:39
some things that really struck us
8:41
as clear delineations. First and foremost,
8:43
if you look at a generative
8:45
model and compare it to historically
8:49
predictive-oriented applications, there was
8:51
this tremendous benefit that
8:54
came out of transfer learning.
8:56
And the models themselves had
8:58
this ability to, with very
9:00
little bit of data, few-shot learning, single-shot
9:03
learning, be able to transfer learn quite
9:05
a bit from what was previously taught
9:07
to the model itself. And so that
9:09
was really different from what was historically
9:11
the case with a predictive model that
9:14
was in market prior, whether you were
9:16
using a neural network or
9:18
any other algorithmic method. In
9:20
a similar way, you started to see
9:22
a difference in terms of how the compute had
9:24
to be managed, right? Because there was
9:26
a tremendous amount of just
9:29
compute required, particularly GPU-based compute that
9:31
was required to be able to
9:33
not only do the heavy training
9:35
that was required to build a
9:37
model with the number of parameters
9:40
that we're now seeing in a model reaching as
9:42
much as a trillion parameters in
9:44
the GPD-4 style of model, but
9:46
also in the inference itself. It
9:48
was a compute-intensive experience to be
9:50
able to handle the inference even in
9:52
a model like what you're seeing with OpenAI's
9:55
GPD-4 or Entropic or Cohere. And the last thing
9:57
that we saw is a difference in the data.
10:00
difference is just the feedback loops.
10:02
If you think about the use
10:04
of RLHF, a reinforcement learning through
10:06
human feedback or an RLAIF, in
10:09
all of these situations,
10:11
the feedback loops were more important
10:13
than ever just to be able
10:15
to improve models, particularly as they
10:17
were in production. And you're starting
10:19
to see this notion that the
10:21
model is almost a living,
10:25
you know, thing that continues to
10:27
improve upon itself using reinforcement learning
10:29
beyond the initial training runs themselves.
10:31
And so those were some of
10:33
the key things that really pushed
10:35
us to push out our
10:38
perspective on what
10:40
was really importantly delineated
10:42
around building for LLM
10:44
ops versus ML
10:47
ops. And of course, there were many
10:49
things that were similar in nature. David,
10:51
privacy was very similar in terms of
10:54
how you worked with ML ops versus
10:56
LLM ops, model governance, model
10:58
security, which I'm sure you spend a
11:00
bit of time with Christian and
11:02
one of our portfolio companies in that regard. But
11:04
in that regard, we wanted to
11:07
just really call out for anyone who
11:09
was interested in the next generation of
11:11
models as they emerged, the LLM models in particular,
11:13
what would it take to build it and how
11:16
is it different from this last generation of machine
11:18
learning. And that's really what we try to encapsulate
11:20
in that article. Okay, and so
11:22
in a couple of ways, just to summarize a
11:24
little bit what you're saying is like on the
11:26
left hand side of the diagram, it's a lot
11:28
of the same problems of data preparation, data privacy,
11:31
putting the data that from an
11:33
enterprise perspective, you feel comfortable is ready to
11:35
go in. But on the right hand side, it's
11:37
quite different. Because unlike ML, that's telling you
11:39
about the data you've already got, it's generating
11:41
new data. And there's a much more of a
11:43
role for the human in the loop on
11:45
that right hand side of iterating
11:47
and using the model that a fair assessment.
11:50
Yeah, I think it's fair. And I think
11:52
there's some things that go into the compute
11:54
management, which we've never had to really think
11:56
about at the scale that we have to
11:58
think about, particularly in in both the model
12:00
training as well as the model interests. And
12:03
we learned our lesson in cloud data. So
12:05
we're going to think about it early now.
12:07
Think about our expenses early. I will just
12:09
share, put in the notes as well. You
12:11
have a really amazing landscape of LLM. People
12:14
are sending it to me all over, you know,
12:16
all over the LinkedIn and my friends are like,
12:18
hey, have you seen this? So I think it's
12:20
terrific. And I would definitely encourage the audience to
12:22
take a quick look at it because I think
12:24
it has this notion of end user management that
12:26
I wanted to really double click on it. I've
12:28
heard you do some other interviews where you talk a
12:30
little bit about how the ability
12:33
to integrate human feedback is an
12:35
asset in your mind. Is that what you mean
12:37
when you talk about end user management as part
12:39
of this? Yes, it is
12:41
one of the pieces, right? For sure. The
12:43
feedback loop is pretty important
12:46
because if you start to think
12:48
about how model alignment occurs, right?
12:50
Building models that have values
12:52
that are aligned with human beings. The only way to
12:55
be able to do that is to have humans in
12:57
the loop to be able to provide the feedback to
12:59
models as they continue to be aligned for our needs.
13:02
I think there's also another piece of the puzzle, which
13:04
is when you think about beyond
13:06
the model itself and how you
13:08
instill that into say, for instance,
13:11
an enterprise experience, you need more
13:13
than just a powerful model with
13:15
good feedback loops. They also happen
13:17
to need to get good private
13:19
data sets, right? It turns
13:21
out that those private data sets
13:23
are the ones that really enable
13:26
models to further target
13:28
and focus the task
13:30
at hand and hallucinate less,
13:32
right? So the more complimentary
13:35
private data that you have to
13:37
the task at hand, the less
13:39
likely we will hallucinate on things
13:41
that were basically not
13:43
having enough data that was being folded
13:45
into the model in terms of the
13:47
model training itself. The third
13:50
thing I would also mention on top of
13:52
that from a user experience standpoint is
13:54
that you still need to think about this as
13:57
an enterprise application. And so great enterprise applications have
13:59
great UX. some workflow associated with it. And
14:01
if you're building an
14:04
enterprise grade application, it's
14:06
not just about the human
14:08
feedback loops in the models themselves, but
14:10
it's a great user experience for the
14:12
application front end. And it could be
14:14
as simple as the chat in the
14:16
face. It could be a more complex
14:18
workflow, but nevertheless, a great user experience
14:21
surrounding that is absolutely essential for any
14:23
AI based application to prosper in the
14:25
enterprise. Have
14:31
you tried Firebase or Superbase, but
14:33
wanted a better developer experience? Convex
14:36
is a backend as a service product
14:38
that's changing the way founders think about
14:40
designing their infrastructure layer by letting anyone
14:42
write their backend like they would write
14:44
their front end. Convex is designed for
14:47
devs to write their backend in pure
14:49
TypeScript, and it features a fast database
14:51
written in REST. Also, the self to
14:53
manage version of the platform is now
14:55
open source, and the team has been
14:57
excited to share their thoughts behind the
15:00
decision. Convex's CTO and co-founder, James Cowling,
15:02
recently joined the Software Engineering Daily podcast
15:04
to talk about the decision landscape for
15:06
teams considering the shift to open source. Check
15:08
out the conversation to hear all about it.
15:10
Be sure to visit convex.dev to see why
15:13
Convex is the best backend as a service
15:15
for developers who want to shift quickly. Yeah,
15:24
I wanna get back to that, but first
15:26
I wanna talk a little bit about this,
15:28
like high value private data, high value data
15:30
sets for train on to help give direction
15:32
to the generative AI model, LLM model. My
15:35
experience has been the people who
15:37
have the tastiest, most desirable private
15:39
data are typically the ones who
15:42
have the least capabilities often in
15:44
building their own software, creating their
15:46
own tools. Do you think
15:48
that has implications? Because I agree, you have
15:50
to have this private rich data, but do
15:52
you think that has implications for this whole
15:54
debate around will it be proprietary open
15:56
source models that win
15:58
overall? to form an opinion
16:00
there when you know that the organizations with the best
16:03
data often are going to be
16:05
reaching out to maybe a more proprietary organization rather
16:07
than building their own, with say a bundle of
16:09
open source models. Have you given some thoughts
16:11
to like, what is that sort of adoption route going to
16:13
look like? Let me kind of call out
16:15
a few things in terms of what we've seen up
16:17
to this point at the sort of tail end of
16:19
2023. And
16:22
then where things continued or will continue
16:24
to evolve in 2024. First
16:27
and foremost, like anyone who's gotten
16:29
a model into production right now, it
16:31
does seem like it's pretty
16:34
much open AI, right? And there's probably
16:36
a few other things that are coming down
16:38
the pike, including work that Anthropic and Cohere
16:40
are doing as well as a number of
16:42
the open source providers and
16:44
particularly longitude is a pretty exciting option
16:46
when you look at the compactness
16:48
of the model itself and just
16:50
how it's commercially available to be
16:52
licensed from an open source underpinning.
16:55
I think for the folks
16:57
who are building, and I
17:00
mentioned that this need to have proprietary
17:02
or private data sets, I
17:04
don't know if it matters
17:06
as much whether you're kind of
17:09
working over time with an
17:11
open source or a closed source model.
17:13
I think it matters a
17:15
lot in terms of what the model
17:17
performance is. And it matters a
17:19
lot in terms of what the likelihood
17:22
that the model hallucinates as
17:24
it is out of the box, how much you
17:26
can either fine tune it or you can use
17:29
retrieval augmented generation
17:31
surrounding the model and being in
17:33
production so that you can sort
17:35
of in a constant deliver the
17:37
results that you wanted to. But
17:39
in almost all those cases, it's
17:41
really going to be about model
17:43
performance and feedback loops and
17:46
whether it achieves
17:48
the objective that you would want in
17:51
terms of your model being in production.
17:53
And right now it does seem that
17:55
the closed source and particularly OpenAI is
17:57
a closed source model today. is
18:00
the highest performing model. And that's
18:02
where most of the fine tuning
18:04
is happening as we speak. I
18:07
think over time, we're gonna see some
18:10
amount of diversity. It's not going
18:12
to just be open AI, closed
18:15
source models only. When you think
18:17
about the type of narrow
18:19
AI use cases where you
18:22
look at like trade settlement and clearing
18:24
and a back office for a financial
18:26
services organization, well, some really massive opportunities
18:29
there in terms of being able to
18:31
introduce a generative model to help compliment the
18:33
work that humans are doing in that regard.
18:36
But when you look at what kind of
18:38
model would you use, do you need to
18:40
have a model that understands 14th
18:43
century European history to be able to do
18:45
trade settlement and clearance? Probably not, right? And
18:48
so I think this is exactly where a
18:50
smaller form factor model, whether it be open
18:52
source or not, it seems like some of
18:54
the more capable, smaller form
18:56
factor models are coming from the open source world. Those
18:59
models could be just as relevant
19:02
for tuning and training a
19:05
very specific private data set to be
19:07
able to accomplish a very specific task
19:09
at hand from a narrow AI standpoint.
19:12
So I think some of that's gonna be coming
19:14
to a theater in Arizona in 2024. We
19:17
just haven't seen it yet mainly because it
19:19
seems like most of the models in production,
19:21
at least the generative AI transformer based models
19:23
that are in production today are
19:25
very much leaning towards a fine tuned version
19:27
of open AI. That's
19:30
really helpful perspective. Just covering the media, it
19:32
seemed like that was a more mature discussion,
19:34
but I think what I'm hearing is we're
19:36
still early days there in terms of which
19:39
direction most enterprises are gonna go for
19:41
their enterprise adoption of OLMs. Yeah,
19:44
I mean, if we kind of cast this at the
19:46
end of the first inning, I would
19:48
agree that there's quite a bit more
19:50
of the game to be played out.
19:53
And I think we're gonna see more
19:55
variety and diversity in models as
19:57
they're in production. It just so happens. first
20:00
hitting of the ball game at the end of it,
20:03
there seems to be only one that has moved
20:05
these reach to production level value for most
20:07
enterprises. This
20:14
episode of Software Engineering Daily is brought
20:16
to you by Vantage. Do you
20:19
know what your cloud bill will be for
20:21
this month? For many companies, cloud costs are
20:23
the number two line item in their budget
20:25
and the number one fastest growing category of
20:27
spend. Vantage helps you get
20:29
a handle on your cloud bills
20:31
with self-serve reports and dashboards built
20:33
for engineers, finance, and operation teams.
20:36
With Vantage, you can put costs
20:38
in the hands of the service
20:40
owners and managers who generate them,
20:42
giving them budgets, alerts, anomaly detection,
20:44
and granular visibility into every dollar.
20:47
With native billing integrations with
20:49
dozens of cloud services, including
20:51
AWS, Azure, GCP, Datadog, Snowflake,
20:54
and Kubernetes, Vantage
20:56
is the one Synops platform to
20:58
monitor and reduce all your cloud
21:00
bills. To get started,
21:02
head to Vantage.sh, connect your accounts,
21:04
and get a free savings estimate
21:06
as part of a 14-day free
21:08
trial. Just
21:16
from a product design perspective, if you're
21:18
designing a product today and you want to
21:20
incorporate an element of LLM, do
21:23
you think there's going to be impacts on the
21:25
way that we actually even do product design?
21:28
I think so. As a former
21:30
product manager myself and someone who's
21:32
led product management teams prior to
21:34
becoming an investor, I think
21:37
for product leaders, you have
21:39
a few things that you have to think
21:42
about. One
21:44
is when you design software, historically,
21:46
most software has had an underpinning
21:48
of a rules engine associated with
21:50
it. These are the sequences of
21:52
things that we encapsulate in our software, and as
21:54
long as it does those things, it comes to
21:57
its terminus scalcum, and here we have a beautiful
21:59
product. piece of potentially workflow
22:01
based software that follows
22:03
those rules. That's how it works every time.
22:05
Seems to have worked in software for 40
22:07
plus years. Now, I think
22:09
what we are on the
22:12
precipice of is systems
22:14
that are less about these
22:16
deterministic outcomes, and more about
22:19
having a probabilistic set
22:21
of reasoning associated with it,
22:23
like human like smarts and
22:25
reasoning associated with it that is
22:28
embedded into the software itself. So
22:30
imagine the rules engines of yesteryear
22:33
now being replaced by a
22:35
generative model as the underpinning engine model
22:37
that continues to improve upon itself, a
22:39
model that learns, a model that has
22:41
reinforcement learning surrounding it. And so if
22:44
you're a product manager thinking about the
22:46
new products and services you're coming to
22:49
the market with, you have
22:51
to almost now introduce this mindset
22:53
of what a probabilistic
22:55
reasoning system could look like
22:58
in software, either alongside of
23:00
your rules engine or completely
23:04
reimagining your rules engine one or
23:06
the other. And so people
23:09
in a funny way sometimes ask me,
23:11
what's the scale and impact of AI
23:13
in the market? Well, at the very
23:16
least, it's going to be the total
23:18
addressable market of all software. And it
23:21
could be the total addressable market
23:23
of all humanity. But at the very least
23:25
on the software end of the spectrum, you
23:28
have this really interesting moment
23:30
as a product leader, to be
23:32
able to take everything that you
23:34
have historically known in terms
23:37
of building rules based software and replacing
23:39
it with a probabilistic model that has
23:41
human like reasoning associated with it. And
23:43
I think there's some powerful things that
23:46
are coming about, particularly in enterprise software,
23:48
if you take that first principle mindset
23:50
to how you're designing the software of
23:52
the future. Yeah, I agree with that,
23:55
right? It's a different mindset, different set of
23:57
tools. And I also think subject
23:59
matter, I experts are going to be invited back into
24:01
the design sessions more. We kind of over indexed on
24:04
like, let's ask the users, but you know, kind of
24:06
back to what you were saying about Mosaic coming out
24:08
and when we, when the internet happens, you know, apps
24:10
kind of were the pointy end of the sphere. And
24:13
to build those apps, you need the subject matter
24:15
experts who deeply understood what are the
24:17
expected outcomes? What is that workflow?
24:20
And similarly, I think for product
24:22
design, there's going to be a requirement,
24:24
right? To have these Sherpas help guide
24:27
everyone through complex, very
24:30
specific enterprise use cases. Subject
24:32
matter experts, process experts, folks
24:35
who understand how the backend
24:37
workflows between humans and machines
24:40
have historically worked. Those
24:42
are all opportunities to reimagine it,
24:44
what generative AI based software is
24:46
it's underpinning. Absolutely. Absolutely. So
24:49
I think that's kind of exciting in a way, because
24:51
a lot of that's still broken. As
24:58
a listener of software engineering daily, you
25:00
understand the impact of generative AI
25:03
on the podcast. We've covered many exciting
25:05
aspects of gen AI technologies, as well
25:08
as the new vulnerabilities and risks they
25:10
bring. Halfer1's AI Red
25:12
teaming addresses the novel challenges of
25:14
AI safety and security for businesses
25:16
launching new AI deployments. Their
25:19
approach involves stress testing AI models and
25:21
deployments to make sure they can't be
25:23
tricked into providing information beyond their intended
25:26
use. And that security flaws can't be
25:28
exploited to access confidential data or systems.
25:31
Within the HackerOne community, over
25:33
750 active hackers specialize
25:35
in prompt hacking and other AI
25:37
security and safety testing. In
25:39
a single recent engagement, a team of 18
25:42
HackerOne hackers quickly identified 26
25:45
valid findings within the initial 24 hours
25:47
and accumulated over 100 valid
25:50
findings in the two week
25:52
engagement. HackerOne offers strategic flexibility,
25:54
rapid deployment, and a hybrid
25:56
talent strategy. Learn more at
25:59
hackerone.com. That's
26:03
hackerone.com/AI.
26:12
One thing I learned as I was looking
26:14
at the MLOps space, I'm going to ask
26:17
you some other questions, but in the MLOps
26:19
space, one thing I thought was a funny
26:21
quote is like, no model runs the same
26:23
way twice, which is sort of the same
26:25
thing that's happening in LLMs, right? There's a
26:27
proliferation of training cycles in the ML world.
26:29
In the generative AI, LLM world, there's like
26:31
a proliferation of models, right? Let's say you've
26:33
got your controls around your data, great. Like
26:36
proliferation of models, you've got bundles of models
26:38
running, not just one. How are
26:40
organizations going to adopt that in a safe way? Is
26:42
it going to be like a feature registry type situation?
26:44
Is it going to be an after the fact audit?
26:46
What do you think is going to happen there? I
26:49
think there's a few things that are happening today
26:51
and there's a few things that will emerge, particularly
26:53
with some of the regulatory frameworks that are also
26:55
coming about. A few things that are happening today
26:57
is just as you called
26:59
out multiple models running
27:02
in tandem with each other, it
27:04
really calls for more orchestration capabilities,
27:07
right? I think that's a big
27:10
need in the market and you're seeing
27:12
existing orchestration providers like Airflow now starting
27:14
to jump in. We were
27:16
investors in a company called Astronomer, which is
27:18
one of the leading purveyors of Airflow. Entering
27:23
into the space as well as
27:25
you have startups that are capable
27:27
of delivering value from a model
27:29
orchestration standpoint. There's some great examples
27:31
there include both Llama
27:33
Index as well as Lion Chain. In
27:36
all of those cases, it's the
27:39
orchestration that's happening across the
27:41
models and the data fabric
27:43
that's underlying these models coming
27:45
into production that's driving the
27:48
need for orchestration tools. In
27:50
a similar way, you're seeing
27:52
the observability market continue to
27:54
evolve and grow quite significantly.
27:56
Again, you've had Krishna on
27:59
the podcast already. If you look at
28:01
Fennler's business, if you look at some
28:03
of the other companies in that category
28:05
of model and observability, their
28:07
focus is really to be able
28:09
to do the necessary model
28:12
observability that lets
28:15
you understand the model performance,
28:17
understands model performance from an
28:20
overall monitoring standpoint, and then
28:22
introduces things like traceability and
28:25
bias detection and fairness into how
28:27
the model monitoring is occurring. And
28:29
so that's a very, call it
28:31
internal view of how models
28:33
are either observed properly or orchestrated
28:35
as two great examples. But there's
28:37
also this external factor that's coming
28:39
to the theater nearest very quickly,
28:41
which is the regulations themselves. And
28:44
what I'm going to kind of see,
28:47
you know, my kind of prediction going into 2024 is
28:49
that there
28:51
is a cast of regulations that
28:53
are coming on a per industry
28:55
basis where regulators are going to
28:58
have updates in terms of what
29:00
they're expecting models to do. And
29:03
that's going to have to be managed from a
29:05
software standpoint, from a governance or compliance
29:07
perspective, a GRC software standpoint.
29:10
And so that's something that we haven't seen
29:12
yet, particularly in the world of AI. But
29:14
I think that's actually something to be really
29:17
kind of paying attention to going into 2024,
29:20
particularly the executive order and additional
29:22
other layers of regulatory frameworks coming
29:24
into the fold. Yeah, I work
29:26
with financials who are already pretty
29:28
highly regulated. And this new wave
29:30
of regulation is not only impressive, where
29:33
you see GDPR, EU suggesting expanding, you
29:35
know, you can elect not to be
29:37
part of a AI driven
29:39
decision, right? As part of as your
29:42
customer experience, like as a citizen, I think that's
29:44
great as a technical person, it makes me like
29:46
a little sick inside, because how do you do
29:48
it? Yeah, how do you do
29:50
it? It's like, okay, we know how
29:53
to put humans and machines together and
29:55
build systems and processes in the financial
29:57
services world effectively. But now if
29:59
you're being asked after you've done that and
30:01
after you've done it well, how do you
30:03
take the machine out of the loop and
30:06
only have a human-based process? Tough. Right.
30:09
We're kind of at the very beginning of that. We don't know
30:11
who's the regulator going to be. There's competing sets of proposed regulation.
30:13
One thing I do advise companies we work with is
30:15
to get started with what you know. There's some no
30:17
regrets work that should get started now
30:19
because waiting for regulators is a losing business,
30:22
right? That can be tough. So
30:24
yeah, we talked about regulatory risk. We talked a little
30:26
bit about data risk, right? And the need to orchestrate
30:28
data and make sure you have a private reliable data
30:30
source if you're doing an enterprise use case. What
30:33
about, we'll just briefly touch on these just because
30:35
if you have an expertise in this area, but
30:37
there's the other two kind of concerns that slow
30:39
down adoption in companies are bias,
30:42
right? Just model bias and then some concerns
30:44
around like just bad actors, prompt injection, that
30:46
type of thing. What are you hearing and
30:48
what are you telling people about those two
30:50
areas of risk? Yeah, by the
30:53
way, I want to hit that in a second, but I
30:55
do want to just elaborate on the data question a little
30:57
bit more and then we'll go into those
30:59
situations and detail those out.
31:02
But in the data side of the world, and
31:04
particularly the privacy preservation around data and how do
31:06
you handle it? I think there's two really compelling
31:08
things that I've seen so far. One
31:11
is where you're using synthetic data
31:13
to generate like copies that are
31:16
statistically significant for the value of
31:18
a training run, for instance, or
31:20
being able to just handle a
31:23
privacy preserved version of the data,
31:25
particularly as you're doing testing before
31:28
my model goes into production or an application
31:30
goes into production. So there's something
31:32
that's iconic that are doing a tremendous
31:34
job in terms of just handling synthetic
31:36
data generation. In a similar way, what
31:39
we're starting to see is this notion
31:41
of like, where do you keep all
31:43
of your private data properly stored and
31:45
managed? And so there's this idea of
31:47
a privacy vault, right? And this is
31:49
where Skyflow was probably one of the
31:52
more unique players right now in this
31:54
space, really focused on just a privacy
31:56
preservation mode around keeping all the PII
31:58
data in one place. going to secure
32:00
API that has a one way in
32:03
and out door to get through to
32:05
get your information that's privacy preserving and everything else
32:07
is kind of left
32:10
in the hands of the application
32:12
developer. But the privacy aspect
32:15
of what you need to handle from
32:17
a PII compliance standpoint is delivered by
32:19
a service like Skyflip.
32:22
So knowing that is in
32:24
place, then let's kind of talk
32:26
a little bit about. I want to hear what
32:28
you have to say about that, but I
32:30
cannot overstate what a huge sea change it
32:32
has been. The embrace of synthetic data, the
32:35
embrace of things like privacy vault and organizations.
32:37
I mean, it is a major step
32:39
forward for organizations that were pretty immature
32:41
on that. So I'm a big tonic
32:43
fan as well. So we'll do
32:45
another show on synthetics another day. That's great. But
32:48
I think it actually leads nicely to the
32:50
question that you were originally asking, which is,
32:52
okay, how do you really think about just
32:54
the security concerns and what are some of
32:57
the underlying risks? I think one of the
32:59
biggest things that we continue to
33:01
see in the space is
33:03
that this is not
33:05
only an opportunity for great things
33:07
to happen, good things to happen,
33:09
but also a veritable treasure
33:12
chest for nefarious actors to
33:14
do unethical things. And
33:17
I think some of that is just in
33:19
the fact that you have these
33:21
models that are capable of doing many
33:23
things and they can be applied in
33:26
very unethical ways. Some of it is
33:28
just you're just less careful with the
33:30
security and the systems surrounding how you're
33:32
building and bringing things into production. I
33:34
think when it comes to the point
33:36
of cyber security that I have been
33:38
paying a lot more attention to more
33:41
recently is the availability of unethical models.
33:43
So an example of this is like
33:45
form GPT. Turns out
33:47
that a number of black hat
33:49
hackers that had some very capable
33:51
skills in terms of hacking put that
33:54
all into a generative model that they
33:56
unethically trained and release that into the
33:58
wild. level one
34:00
flathead hacker can have the skill
34:03
sets that were now available to
34:05
someone who is way beyond that
34:07
capability and skill all encapsulated in
34:10
the co-pilot that does you know
34:12
the work that a level 506
34:14
flathead hacker could do with the
34:16
tools built into. Demonstrating bad actors.
34:19
Yeah democratize into a tool that's generally
34:21
available and the dark web for anyone
34:23
to take advantage of that wants to
34:25
use it for nefarious purposes. So I
34:28
think that's one example of where
34:31
the things that we are really excited
34:33
by in terms of all this opportunity
34:35
that's emerging scale of human productivity that's
34:37
coming out of the emergence of generative
34:39
AI there will be these dark corners
34:41
and these dark corners are going to
34:44
come to a theater near us and
34:46
they're going to be pretty pronounced and there's
34:49
going to be some really public incidences that
34:51
will come about in 2024 and beyond that
34:53
we just have to be much more aware
34:55
of and much more
34:58
guarded about than ever before
35:00
because the use of this is
35:02
now you know possible for
35:05
not only some of the black
35:07
hack techniques that I just mentioned
35:10
but also just impersonation. Right. You
35:12
can take three seconds of this
35:14
conversation that we're having and synthetically
35:16
generate you know my voice or
35:19
your voice Jocelyn and now put
35:21
that into an interactive voice response
35:23
system and break through the typical
35:26
IVRs that have been in place
35:28
to secure someone's input
35:31
into an online or
35:33
voice based account update for a bank
35:36
and so this is actually
35:38
a lot trickier than all the rainbows
35:40
and unicorns from all the opportunities that
35:42
are coming there's some really dark corners
35:44
that we're still about to face as
35:46
a society. We're in this
35:48
awkward liminal space between incredible
35:50
excitement about the opportunities and capabilities
35:52
but also we don't have enough
35:55
tools or thought work yet to mitigate some
35:57
of these risks. Yeah I mean the scale
35:59
of deepfakes right. now that are coming to a theater
36:02
in Europe seems just unprecedented and
36:05
that should be troublesome for anyone who's thinking
36:07
about this stuff today. I think I read
36:09
that's one of the things that inspired Joe Biden to
36:11
do this executive order is that he saw a fake
36:13
of himself. Did you see that? It's
36:19
like up to that point where like, you know, we'll
36:21
get around to, you know, an executive order for AI
36:23
and then finally the president sees a deep fake of
36:26
himself. It's like, oh, we should probably do something about
36:28
this. I did not know that. That's
36:30
amazing. It's like such a great
36:32
feeling. It's like sometimes people are like, Oh, you
36:34
sent that email. I'm like, I did. Like I
36:36
already did. There's a million things you say and
36:38
do that you don't remember. So it could easily be fake.
36:41
So let's switch gears a little bit, because
36:43
I wanted to talk a little bit
36:45
about advising on a lot of our
36:47
audience and like technical entrepreneurs, people who
36:49
want to quit their day job and
36:51
start a company. Maybe they're already in
36:53
like a small startup. They're in that seat
36:55
or, you know, a round space. People have
36:58
asked me, what do you do if you're
37:00
an established company? You got started as a
37:02
startup, but maybe you don't have generative
37:04
AI in your pitch. Maybe you don't have it
37:07
in your product. What do
37:09
you do? What do you do to keep
37:11
your investors engaged, to keep customers engaged, because
37:13
it's almost required to talk about
37:15
it now. I'm sure you've gotten a similar
37:17
question. What are your thoughts? Well,
37:19
we've been a little more proactive at
37:22
insight about this, because we generally do
37:24
believe that this is not just happy
37:26
talk. And there's, you know, some genuine
37:28
benefits of reimagining your software business as
37:30
a generative AI based underpinned software business.
37:32
And so in that regard, our inside
37:34
onsite team, which is really
37:37
helping our portfolio companies continue to
37:39
grow and scale in their specific
37:41
journeys, whether it be in sales
37:43
and marketing or talent or popped
37:45
in engineering, or finance or business
37:47
development. In all of those
37:49
functions, you can kind of see two
37:51
big opportunities emerge for generative AI. One
37:53
is just better products and services, and
37:55
then more efficiently running your business. And so in that
37:57
regard, in the last six months, we've been working on a lot of things.
38:00
months, we've spent time with all of
38:02
our portfolio companies to help them think
38:04
through exactly those two levers. Now, what
38:06
are the products and services that
38:08
you can be building with what you have
38:10
today that will enable you
38:12
to continue to expand into the opportunities
38:15
that you have in market because of
38:17
generative AI, as well as how can
38:19
you run your businesses better, mostly sales
38:21
and marketing, it turns out. And now
38:24
it is in our view, a situation
38:26
where I would be actually
38:28
delighted to be a company that's already
38:31
in market that doesn't necessarily have a
38:33
generative AI strategy. Why? It turns out
38:35
that you have a clemency. It turns
38:38
out you have private data sets. It
38:40
turns out you have a
38:42
market where there is distribution because you
38:44
are the incumbent in the space. And
38:48
in all of those
38:50
moments where incumbency distribution
38:52
and data pre-exists, it's
38:55
actually not that hard to build a generative bottle,
38:57
right? I don't mind tremendous amounts of
39:00
data science resources and machine
39:02
learning resources because you're mostly
39:05
requiring data engineering and great software developers,
39:07
data engineering to get the data prepped
39:09
in a way that you can tokenize
39:11
and sort of train a model appropriately
39:13
or tune a model appropriately. And then
39:15
of course, the development resources
39:17
to be able to put that into
39:20
an application in the right context. And
39:22
so what we're seeing is
39:24
a lot of existing folks
39:27
who have had very little
39:29
capabilities expertise from an AI
39:31
standpoint now start to build
39:34
some incredible products and services.
39:37
And if nothing else, look at how the incumbents
39:39
in the space have been so capable
39:41
of building great products. I mean, look at
39:43
what Microsoft has done with
39:45
GitHub Compilot as a starting point.
39:48
Look at what Adobe has done
39:50
with Adobe Firefly as a compliment
39:52
to being able to create a
39:55
product that enables you to
39:57
verify the sources of the
39:59
creation of. generative content and plug
40:01
that into Photoshop. And so in all
40:03
of these examples, what we're seeing is,
40:05
the comments are actually not at a
40:07
disadvantage. In fact, they might be at
40:09
a relative advantage, even if
40:11
you are starting your AI journey very
40:14
late, you can almost sort of catch
40:16
up quickly by adopting some of
40:18
the generative AI capabilities. I love that. I'm
40:20
an optimist at heart and you're absolutely right.
40:22
Where you've got the data, you've got the know-how,
40:24
maybe you have the customer relationships, those are the
40:26
hardest parts. If you're already an
40:28
established, you've got some of that in hand.
40:30
Let's say we're, because you mentioned Microsoft, you
40:33
know, when you talk to entrepreneurs who
40:36
are putting together their ideas, they're very early stage.
40:38
There's a couple of items of common wisdom,
40:41
which I'll share with you and you can
40:43
debunk it or agree. The biggest players, Google,
40:45
Microsoft, Meta, are gonna just sort of mop
40:47
up here. They've already got so much money,
40:50
they've got so much tailwind, you
40:52
know, maybe you have to have a
40:54
different idea process. What
40:56
do you think about that? Well, the
40:58
one incumbent, which just literally emerged out
41:00
of nowhere this past year that no
41:02
one would ever have thought of as
41:04
an incumbent, but now they are, you
41:06
know, within a year inclusive of the
41:08
one incumbents you mentioned, is also OpenAI,
41:10
right? And if you
41:12
notice, developer day, you know,
41:14
not more than a week ago, week and
41:17
a half ago, what you
41:19
saw was just a number
41:21
of startups and startup ideas
41:23
being completely eviscerated. I'm
41:25
digested, I'm digested. By
41:28
just the capabilities that were introduced around
41:30
GPTs and capabilities that
41:32
were introduced around just context windows and
41:34
all the additional features that, you know,
41:36
OpenAI themselves announced. For our audience, if
41:38
you haven't Googled, just go into YouTube
41:41
and just Google, like, did OpenAI just
41:43
kill all the startups? There's
41:45
some really good rundowns on that. I don't believe
41:47
that to be true, but it was a huge
41:49
set of announcements. Yeah, and again,
41:51
can OpenAI do all those things that
41:53
kill all the stars? Probably not, because
41:55
there's so much surface area to cover.
41:57
It's impossible for one company to do.
42:00
all that and that's why a great startup
42:02
ecosystem continues to thrive particularly in the
42:04
JIT or AI space. I think what
42:06
you're going to have to think about
42:08
as a founder though is you can't
42:11
be working towards a
42:13
feature that can be disrupted
42:15
quickly. You have to
42:17
be working towards at the very
42:19
least a strong unique product that
42:22
maintains its value proposition if not
42:24
a broader platform that certainly has
42:26
resistance and resilience in an overall
42:28
market. I think that's harder
42:30
to do. It takes longer. It
42:32
takes more investment to build products
42:35
and platforms over time versus features.
42:37
I think the features are going
42:39
to get really folded
42:41
into the incumbents that are building already
42:44
in the markets that they're in,
42:46
but I think there's ways that you can
42:48
build great products and platforms.
42:50
I mean, just look what Jasper has done.
42:53
They started as a prosumer
42:55
offering in content creation for content
42:57
marketers around generative AI. It
43:00
turns out open AI came in and
43:03
really sort of ate into that original
43:05
core assumption in the most of market,
43:07
but that enabled Jasper to take all
43:10
everything that they learned and go into
43:12
the enterprise and build great worldflows and
43:14
great applications and great embedded experiences for
43:16
how enterprise marketers work today. I
43:19
think that's a lesson to be taken for
43:22
all founders going through this journey that even
43:24
when you feel like you have called a
43:26
product market fit and you have success, you
43:28
will have to make some changes so
43:31
that you can find long-term durability.
43:34
My view is that as a
43:36
founder right now, you can't build
43:39
for features and you can't be
43:41
thinking just about the product
43:43
alone. You have to be thinking both
43:45
about product and go to market simultaneously.
43:48
It's interesting because if you talk to
43:50
any second time founder, that's really how
43:52
their mindset is. First time
43:55
founders always have the sensibility that it's about
43:57
product and yes, it is of course starting
43:59
with product. But second time founders are like,
44:01
yes, it's about product, but then how do I
44:03
go bring it to market? So it's like, George,
44:05
what does that mean? I hear people talk about GTM
44:07
all the time, but I would assume what you're saying
44:10
is like, hey, I've got the next great widget. This
44:12
is a cool product. Just go to market like, hey,
44:14
I have a letter of intent from a big customer,
44:16
and we're going to co-design this. Like, what are those
44:19
practically speaking? What would those be? I
44:21
mean, in the early days, it's good
44:23
design partners. It's good focus for folks
44:25
who are going to co-collaborate in building
44:28
what you're bringing to market. But I
44:30
think soon after that, like when you're
44:32
closer to first generation,
44:34
your MVP call it your first generation
44:36
of the product, then you
44:38
got to quickly move your mind towards
44:40
commercialization and how to get paid what
44:44
you're building. And I think in
44:46
previous incarnations, we've had more time
44:49
and we've had more ability to
44:51
keep building and not have
44:53
to worry as much about commercial
44:55
value and selling and
44:58
being good at finding
45:00
the things that you're building have some
45:02
ability to find. Yeah, yeah. I mean,
45:04
like ruthlessly focused on what's going to
45:06
sell, not just what's like, you
45:08
know, close to your heart. Yeah, and focus
45:10
earlier on that that you want. Yeah.
45:14
And I think that's something that in
45:16
this cycle of iteration and at
45:18
the clock speed that we're moving at right
45:20
now, you're going to have to
45:23
really find your commercial value
45:25
faster. And even
45:27
after you find your commercial value,
45:30
you might have moments where you
45:32
still have to pivot. Yeah, it's
45:34
so easy to say and hard to do like when
45:36
you're in the situation as a product, so you fall
45:38
in love with your product, you've got, you know, certain
45:40
things. And, you know, even though it's a simple statement,
45:42
I actually can't underline this enough for our listeners, how
45:45
important that is that kind of ruthless focus on
45:47
what people are going to buy. I mean,
45:49
for great founders, great founders as product people
45:51
are tortured artists, I get it. I
45:54
was certainly one of those folks myself,
45:56
but I think the best founders in
45:58
time, not a lot. are the
46:00
tortured artists, but are the world-class salespeople.
46:03
And so if you can take the
46:05
combination of tortured artists and world-class salesperson
46:07
and bring that together, those are the
46:10
founders that will go the distance over
46:12
time. Well, speaking of founders that will go
46:14
the distance, right? We talked a little bit about the overall
46:16
markets and things you're looking at in 2024, some like Signal,
46:18
you've got your own
46:21
point of view as an investor. You know,
46:23
you've invested in some, you've got your current
46:25
portfolio companies. Do you want to share some
46:27
stories of how some of your current portfolio
46:30
companies have checked off the boxes you think
46:32
are important in developing in this area? Yeah.
46:35
I mean, I'll give two examples of
46:37
companies that have some tremendous scale.
46:39
I mentioned weights and biases, but
46:41
I think it's important to understand
46:43
that when we had a great
46:45
blog post about Lucas and his
46:47
journey around weights and biases more
46:49
recently, the reason why weights and
46:51
biases has worked so well in
46:53
the market is that Lucas had
46:55
his first startup, which is CrowdFlower,
46:57
and he had all of the
46:59
challenges of building CrowdFlower. It is
47:01
mine as a second-time founder that
47:03
went to market and not
47:05
only built it for that little product,
47:08
but figured out a way that the
47:10
commercial business and scales insight was lucky
47:12
enough to catch weights and biases even
47:14
before the commercialization really started. And now
47:16
we're starting to three years later, we're
47:18
starting to see the tremendous scale of
47:20
how big a business like weights and
47:22
biases, particularly as a leading tool in
47:24
the MLOps, so MLOps World can get.
47:27
I think outside of even just the
47:29
AI space where software businesses are
47:31
becoming AI businesses, what Christine
47:33
Yen is doing with Honeycomb
47:35
is fascinating, right? Because it's not
47:38
just the fact that you're building
47:40
a next-generation observability product that is
47:42
fully differentiated from all the existing
47:44
incumbents in the space, but they're
47:47
doing innovative things like introducing generative
47:49
capabilities into a traditional
47:51
system observability market. And it turns
47:53
out that in the case
47:56
of Honeycomb, what the unique
47:58
value proposition of interest of
48:00
AI was that every observability product
48:02
has somewhat of an esoteric language
48:04
to be able to work in
48:06
the context of finding that observable
48:08
event. If you introduce natural
48:10
language as the front end, how
48:13
much easier and how much more democratized could
48:15
an observability product be? It's exactly
48:17
the thesis that the team went
48:20
in at Honeycomb to go build
48:22
a natural language interface using a
48:24
fine-tuned open AI experience into
48:27
a query builder for Honeycomb. It
48:29
was the fastest adopted feature
48:32
in Honeycomb's history where 40% of
48:34
all of Honeycomb users were
48:36
on that product within the first 24
48:39
hours of it launching to give you an
48:41
idea of how fast the adoption curve occurs.
48:44
What I would generally give advice to the
48:47
founders as they're going through this
48:49
is really continue to find those
48:52
great moments of product differentiation early
48:54
as well as later in your
48:56
product development cycle and then be
48:59
willing to continue to experiment and
49:01
iterate in the first
49:03
commercialization successes that you have.
49:05
Don't rest on those stories.
49:08
Find different ways that you can continue to
49:10
commercialize and scale your business because the market
49:12
continues to evolve very quickly and you want
49:14
to be resilient for all those changes that
49:17
are occurring. Very interesting.
49:19
We're going to wrap up here at the end of
49:21
our discussion and I really appreciate you taking the time.
49:25
Just one last idea here. You have a
49:27
point of view as an investor. In
49:29
the last few years, a lot of investors have different ways
49:31
of approaching the market. Some of them don't have a point
49:34
of view. Some of them kind of like, I'll take your
49:36
point of view and I'll use that too.
49:39
What do you think they're doing well in this space and
49:41
what could they be doing better? I
49:44
think most VCs are just
49:46
very good at sensing when there's a
49:48
shift in the market and really embracing
49:50
that shift. If
49:53
you think about where momentum is occurring,
49:55
VCs have some of the best sense
49:57
of momentum that could be. Now,
50:01
no surprise, kind of, how the
50:03
generative AI market has really evolved
50:06
because VCs have put a lot
50:08
of energy, a lot of investment
50:10
directly into that market. Now, I
50:12
think where we as VCs
50:15
can do a lot better over time
50:17
is just really, truly
50:19
understand how hard the founder journeys
50:21
are and be more
50:23
empathetic to how difficult these journeys
50:26
are because as much as you're
50:28
like, hey, you should move into this portion of the market
50:30
and you should do it. That takes work. That takes
50:33
a lot of energy. It just takes a lot
50:35
of calories. I think the best
50:37
VCs at some point have figured out a
50:39
way to have that empathy in terms
50:42
of really understanding what the founder's
50:44
journey is and being able to
50:47
be a co-pilot to their journeys. I
50:50
think for me personally, that empathy
50:52
came from just having to have
50:54
been one as a founder and
50:56
a builder in the past and that kind of gave me
50:58
a little bit of that
51:00
empathy as I went into venture capital. From
51:04
the venture capitalists that haven't necessarily done
51:06
that, okay, because I think you have
51:08
enough experiences of just being on boards,
51:10
being with founders that you should be
51:13
able to translate that to providing proper
51:15
empathy for them. I think in most
51:17
cases, the more that we
51:19
are empathetic as VCs and the more that
51:21
we understand the momentum that's occurring in the
51:24
market, the two of those things combined together
51:26
generally, you're going to end up with a reasonable
51:28
set of investments made as you're progressing. I
51:30
love that. I
51:33
love the idea too that you don't have to have
51:35
been a founder, but they just be with them. I
51:37
like that phrase, that simple phrase is great. Just
51:40
be with them for a while. Sit with them
51:42
and experience their point of view. I love that.
51:44
Alright, last question. My inbox and
51:46
every bookshelf is just topped up
51:49
with information in this space and reading
51:51
material that I often don't get through. For
51:54
our listeners, what are one or two things
51:56
that you've read or newsletters or articles that
51:59
obviously are not available? Obviously, Insight has some great
52:01
ones, but if there were one or two things
52:03
that somebody wanted to read, what would you recommend
52:06
right now? Yeah. I mean,
52:08
as far as a lot of the thinking
52:10
on responsible AI, certainly the work that we've
52:13
seen in market by observability leaders
52:15
like Hitler has been great. I
52:17
would also add the constitutional AI
52:20
frameworks that are emerging and tropic
52:22
is probably one of the folks
52:24
that have been leading that work
52:26
and some of the work that's
52:28
surrounding constitutional AI and RL, AIF,
52:30
the use of reinforcement learning and
52:32
AI systems to almost guard rail
52:35
how AI is in the wild
52:37
is a very emerging and topical field
52:39
and I would sort of dig into
52:41
that. I think kind of
52:43
more broadly, if you think about responsible AI
52:46
and where it's going to play out in
52:48
the next few years, we just have to
52:50
build better aligned systems to human values and
52:52
I think one of the original thoughts
52:54
on that was just what Brian Christensen had
52:57
wrote around a book called The Alignment Problem.
53:00
I would highly recommend taking a look at
53:02
that book as a way of really
53:04
thinking about how we can
53:06
build better aligned systems. If
53:08
you want to just have a discourse
53:10
on just how we think about AI
53:13
systems and the future of AI systems
53:16
sitting alongside of humanity, one
53:18
of my favorite recent works on this
53:21
topic was a series of
53:23
essays that were compiled into a
53:25
book written by Megan O'Gibley called
53:27
God, Human, Animal, Machine. It
53:30
was a wonderful sort of
53:32
amalgamation of her thoughts
53:34
on what it took to build
53:37
this current generation of AI systems and
53:39
what are the types of people that
53:41
really are finding meaning around
53:44
the AI systems that could
53:46
be part of what I almost think
53:48
of as co-pilots to our lives. Those
53:50
are probably three books and three sources
53:52
that I'd be inspired by. I almost
53:55
didn't ask that question and now I learned
53:57
at least two new things from you. Thank
53:59
you. That's great. Well, thank you
54:01
so much for taking the time. Software Engineering Daily thanks
54:03
you. I know our audience is super excited to hear
54:05
from you. Hope you have a great rest of your
54:07
day. Thanks a lot. Great. Thank you, again. Appreciate it.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More