WEBVTT

STYLE
::cue {
  font-family: Arial;
  color: white;
  font-size: 40px;
  background-color: rgba(0, 0, 0, 0.65);
  padding: 40px;
  border: 10px;
}

00:00:03.680 --> 00:00:06.000
I’m Winston Weinberg,
I’m the CEO of Harvey

00:00:06.000 --> 00:00:08.040
and we are building a
collaborative deals platform

00:00:08.040 --> 00:00:11.280
with PwC, where we
basically combine the best

00:00:11.280 --> 00:00:15.760
AI talent and product with the best
domain expertise of PwC.

00:00:16.040 --> 00:00:19.040
I'm Gabriel Pereyra,
I'm president at Harvey.

00:00:19.280 --> 00:00:23.640
My background is AI research.
Prior to Harvey I worked

00:00:23.640 --> 00:00:26.880
at some of the labs where
generative AI was invented.

00:00:27.240 --> 00:00:33.280
We're incredibly excited to build systems
that can assist PwC’s

00:00:33.280 --> 00:00:39.120
domain experts
to tackle the most complex transactions.

00:00:39.600 --> 00:00:41.280
I think

00:00:41.280 --> 00:00:44.280
what is most exciting is solving

00:00:44.400 --> 00:00:49.640
real-world problems that are far beyond
the scope of any of the academic research,

00:00:49.640 --> 00:00:53.240
data sets or capability
of existing foundation models right now.

00:00:53.720 --> 00:00:57.240
So our vision for Harvey
is to bring AI to professional services.

00:00:57.480 --> 00:01:00.840
And what's really important about that is
we can't build it in a lab.

00:01:01.000 --> 00:01:01.320
Right.

00:01:01.320 --> 00:01:02.880
So in the same way that you can't build

00:01:02.880 --> 00:01:06.720
self-driving cars in a lab
because you run into all of these issues

00:01:06.720 --> 00:01:09.000
that you, you know,
you might not be able to predict,

00:01:10.520 --> 00:01:12.600
you kind of have to do the same thing
with professional services.

00:01:12.600 --> 00:01:15.960
Professional services
is incredibly complex, right?

00:01:15.960 --> 00:01:19.120
And a lot of the work that you're
doing is related to

00:01:19.320 --> 00:01:22.360
a particular client's needs
or a particular domain’s needs.

00:01:22.560 --> 00:01:26.160
And so you need the domain experts
to help you actually work

00:01:26.160 --> 00:01:27.360
through those problems.

00:01:27.360 --> 00:01:31.920
We are combining the best research talent
with the best domain experts.

00:01:32.120 --> 00:01:36.240
And that is exactly how you have to build
the best AI systems.

00:01:36.720 --> 00:01:40.720
Right now we work with a number
of the foundation model providers.

00:01:41.040 --> 00:01:44.760
We work super closely
with OpenAI and Microsoft.

00:01:44.760 --> 00:01:47.480
We're starting to expand
to work with other providers.

00:01:47.480 --> 00:01:50.080
We are also starting
to build our own in-house,

00:01:50.080 --> 00:01:53.040
and we are building everything
from foundation

00:01:53.040 --> 00:01:56.920
models to agents, RAG systems,
kind of anything under the sun

00:01:56.920 --> 00:02:00.400
that helps us help our clients
with their challenges.

00:02:00.880 --> 00:02:05.440
What we've found is the real way
to unlock the potential of AI

00:02:05.600 --> 00:02:08.520
is to actually train models and systems
to do these

00:02:08.520 --> 00:02:12.240
very specific workflows,
and then chain them together.

00:02:12.400 --> 00:02:16.280
Even the best foundation models right now
are nowhere close

00:02:16.280 --> 00:02:18.440
to where domain experts are.

00:02:18.440 --> 00:02:22.760
So figuring out which systems
we can use to build agents,

00:02:22.760 --> 00:02:26.280
and more complex systems that can aid
in these very complex domains,

00:02:26.600 --> 00:02:28.080
I think that's one.

00:02:28.080 --> 00:02:30.400
Another is security.

00:02:30.400 --> 00:02:33.280
In any of these domains,
you could have the best performing model,

00:02:33.280 --> 00:02:34.560
but if it's not secure,

00:02:36.000 --> 00:02:38.280
none of our clients
or their clients are going to use it.

00:02:38.280 --> 00:02:42.480
I think collaboration is the only way
that you can build these systems.

00:02:42.880 --> 00:02:45.600
We've hired incredible AI

00:02:45.600 --> 00:02:48.600
talent and incredible legal talent,

00:02:48.800 --> 00:02:51.800
but as we started expanding
into other domains,

00:02:51.800 --> 00:02:56.760
there is so much subtlety and nuance
and complexity to these problems

00:02:56.760 --> 00:03:01.160
that even just being able to tell
is this model output a good output

00:03:01.480 --> 00:03:06.000
is incredibly expensive,
the value per hour of

00:03:06.720 --> 00:03:09.400
PwC’s domain experts is so high.

00:03:09.400 --> 00:03:14.120
It was an obvious choice to choose
PwC to build this platform with.

00:03:14.120 --> 00:03:18.320
They’re the largest deal provider in the world.

00:03:18.920 --> 00:03:23.280
And the other thing that they have is
they have a very good way of kind

00:03:23.280 --> 00:03:27.480
of cross-sectionally looking at
problems, right, in various jurisdictions.

00:03:27.480 --> 00:03:30.840
So over time,
especially after the internet,

00:03:30.840 --> 00:03:33.000
deals are becoming more and more complex,

00:03:33.000 --> 00:03:36.000
and there's all of these different
jurisdictions that you have to work with.

00:03:36.120 --> 00:03:41.480
And PwC, more than anyone else
has this sense of collaboration across

00:03:41.480 --> 00:03:45.520
all of these jurisdictions and across
all of these areas of expertise.