WEBVTT

00:00:00.450 --> 00:00:04.447
- Let's have faith in the pace
of progress of the technology,

00:00:04.530 --> 00:00:07.177
because if the models get good
enough to do it end to end,

00:00:07.260 --> 00:00:10.107
a year from now and only
then you start deploying it.

00:00:10.190 --> 00:00:12.607
There will be another two year delay
and that's, you know, that's two years

00:00:12.690 --> 00:00:14.257
during which all
the work that you're doing

00:00:14.340 --> 00:00:16.687
to benefit patients is not happening.

00:00:16.770 --> 00:00:21.217
Hello everyone. My name is Diogo Rau and
I'm Chief Information and Digital Officer

00:00:21.300 --> 00:00:24.427
of Eli Lilly and Company.
I'm joined here with Dario,

00:00:24.510 --> 00:00:26.767
who is the Founder and CEO of Anthropic.

00:00:26.850 --> 00:00:28.867
Dario, thanks for joining me today.

00:00:28.950 --> 00:00:31.327
- Thanks for having me, Diogo. - I know

00:00:31.410 --> 00:00:33.577
you're spending a lot
of time now thinking about

00:00:33.660 --> 00:00:34.837
how do you work better with enterprises?

00:00:34.920 --> 00:00:36.517
What's your enterprise strategy

00:00:36.600 --> 00:00:41.137
and how do you see Anthropic
different from other providers?

00:00:41.220 --> 00:00:42.997
- Yeah, I mean, you know,
I think we've made a number

00:00:43.080 --> 00:00:45.787
of choices that are different, right?

00:00:45.870 --> 00:00:49.777
So if I think about the incentives
given by consumer AI,

00:00:49.860 --> 00:00:53.767
their folks are in a competition
for engagement and growth, right?

00:00:53.850 --> 00:00:57.487
And so that drives a lot
of behaviors of the AI

00:00:57.570 --> 00:01:00.337
that I think are not ideal
from an enterprise perspective,

00:01:00.420 --> 00:01:02.617
For example, there's
this idea of model sycophancy

00:01:02.700 --> 00:01:05.337
where the model tells you whatever
you say is a good idea, - Right.

00:01:05.420 --> 00:01:08.947
right, and even on the consumer side
that can, you know, cause problems.

00:01:09.030 --> 00:01:10.597
We've seen stories of people
who are like, oh yeah,

00:01:10.680 --> 00:01:12.777
I've discovered a new
fundamental theory of physics.

00:01:12.860 --> 00:01:16.387
- And that's right. - Models, like, that's
great and maybe you don't want it to,

00:01:16.470 --> 00:01:18.817
maybe you don't want it to say that,

00:01:18.900 --> 00:01:22.327
but I think, you know,
of course on the enterprise side,

00:01:22.410 --> 00:01:27.637
you know, the problems are much greater
and clearer with that, where, you know,

00:01:27.720 --> 00:01:30.787
you really don't want the model to say,
oh yeah, this drug compound's great.

00:01:30.870 --> 00:01:32.767
Spend millions of dollars to, you know,

00:01:32.850 --> 00:01:34.897
I just think this is, you know,
I think your idea is great.

00:01:34.980 --> 00:01:37.447
I think it's really promising,
like, you want truth.

00:01:37.530 --> 00:01:38.363
- Yeah.

00:01:39.240 --> 00:01:41.737
- And so I think that incentive has led us

00:01:41.820 --> 00:01:43.447
to design our models
in a different way, right?

00:01:43.530 --> 00:01:47.707
I think it's more compatible
with making the model smarter,

00:01:47.790 --> 00:01:53.497
making them better at a wide variety
of economically valuable tasks.

00:01:53.580 --> 00:01:58.087
And it causes us to put a premium
on accuracy and reliability.

00:01:58.170 --> 00:02:01.147
One experiment, you know,
that I give to everyone,

00:02:01.230 --> 00:02:03.007
although it's particularly relevant

00:02:03.090 --> 00:02:05.077
because I'm talking to you,
is I say, you know,

00:02:05.160 --> 00:02:08.107
let's say I improve the model's
knowledge of biochemistry

00:02:08.190 --> 00:02:11.107
from undergraduate level knowledge
to graduate level knowledge.

00:02:11.190 --> 00:02:12.277
You know, if I go to consumers

00:02:12.360 --> 00:02:15.217
and say that, 99% of them
are going to say, you know,

00:02:15.300 --> 00:02:16.837
I didn't know what you
were talking about before.

00:02:16.920 --> 00:02:20.737
I don't know what you're talking about
now, but if I go to you, like you really,

00:02:20.820 --> 00:02:22.657
you care about that,
a lot. - Appreciate that.

00:02:22.740 --> 00:02:25.477
- Like, that's very important.

00:02:25.560 --> 00:02:27.697
- That's exactly right. Well,
actually, that gets into something else

00:02:27.780 --> 00:02:30.067
that you've launched as well,
which is Skills, right?

00:02:30.150 --> 00:02:32.047
There are a lot of skills
that you want in biology

00:02:32.130 --> 00:02:34.950
or even just skills,
like, as an enterprise,

00:02:35.810 --> 00:02:38.947
how you want to operate. Is that part
of the future for you as well?

00:02:39.030 --> 00:02:40.597
- Yeah, I definitely think so.

00:02:40.680 --> 00:02:46.177
I mean, things ranging from skills to, you
know, we're in the process of launching

00:02:46.260 --> 00:02:49.447
various specialized Claudes
which are, you know, in some cases

00:02:49.530 --> 00:02:52.957
will be improvements to the model
itself, fine tunings of the model.

00:02:53.040 --> 00:02:54.997
But in some cases it'll be something

00:02:55.080 --> 00:02:58.657
that looks more like
wrapping the model with access

00:02:58.740 --> 00:03:00.097
to particular types of information.

00:03:00.180 --> 00:03:02.257
So when we did Claude
for Financial Services,

00:03:02.340 --> 00:03:07.807
you know, we connected to a lot
of the usual kind of indices and ratings.

00:03:07.890 --> 00:03:13.147
And so, you know, you'd be surprised
to connect Claude to those things

00:03:13.230 --> 00:03:16.687
and kind of use it in a way that's
aware of that knowledge is valuable.

00:03:16.770 --> 00:03:18.637
So I think, you know,
we're working on a Claude

00:03:18.720 --> 00:03:20.887
for Life Sciences
that will be some mixture

00:03:20.970 --> 00:03:25.867
of making the model inherently smarter and
wrapping it with various things, right?

00:03:25.950 --> 00:03:28.147
I don't know exactly what
the analogy will be here,

00:03:28.230 --> 00:03:30.967
but like, geez, there are
zillions of databases of,

00:03:31.050 --> 00:03:34.117
you know, proteins, compounds,
assays, like, you know,

00:03:34.200 --> 00:03:36.607
you probably want that
at the model's fingertips.

00:03:36.690 --> 00:03:39.697
- Well, any parting advice for those
of us that are working in this world

00:03:39.780 --> 00:03:41.917
of drug discovery and development?

00:03:42.000 --> 00:03:44.677
- You know, I would
say there's a temptation

00:03:44.760 --> 00:03:48.157
and it's, I think it's hard
to avoid starting this way

00:03:48.240 --> 00:03:51.097
of, you know, what are the small
things we can do with AI?

00:03:51.180 --> 00:03:53.707
Like, and in a way you just
kind of have to start there.

00:03:53.790 --> 00:03:56.557
I think one of my pieces
of advice is be very,

00:03:56.640 --> 00:03:59.917
very ambitious in terms
of where the models are going.

00:04:00.000 --> 00:04:04.177
I think you can get caught in a mode
where there's an existing process,

00:04:04.260 --> 00:04:09.877
it has 20 parts you want to swap
in AI, to part five and part 12.

00:04:09.960 --> 00:04:13.567
And, you know, that can actually be hard

00:04:13.650 --> 00:04:16.537
because part 12 has to,
you know, intersect with part 13

00:04:16.620 --> 00:04:19.657
and part 11, which are
not being done with AI.

00:04:19.740 --> 00:04:23.887
And you know, you look at it and you're
like, well, the AI models aren't

00:04:23.970 --> 00:04:28.027
where they could do, you know,
part zero to part 20, end to end,

00:04:28.110 --> 00:04:30.277
But in a year they might be.
- That's right. - And so,

00:04:30.360 --> 00:04:34.057
you should start thinking now,
don't get too seduced by, oh,

00:04:34.140 --> 00:04:36.727
we can make these little
hill climbing gains

00:04:36.810 --> 00:04:39.337
by, you know, doing
this part and that part,

00:04:39.420 --> 00:04:43.207
let's start preparing to do
the whole thing end to end.

00:04:43.290 --> 00:04:47.377
Let's have faith in the pace
of progress of the technology.

00:04:47.460 --> 00:04:50.077
Because if the models get good
enough to do it end to end,

00:04:50.160 --> 00:04:52.837
a year from now and only
then you start deploying it,

00:04:52.920 --> 00:04:55.207
there will be another two year,
- That's right. - There'll be

00:04:55.290 --> 00:04:56.947
another two year delay.
And that's, you know, that's two years

00:04:57.030 --> 00:04:58.657
during which all
the work that you're doing

00:04:58.740 --> 00:05:00.877
to benefit patients is not happening.

00:05:00.960 --> 00:05:05.527
Whereas if you go in parallel, if you
start preparing now for the large change

00:05:05.610 --> 00:05:09.607
as the models are getting better,
then you know, you may save years of time.

00:05:09.690 --> 00:05:12.157
- That's right. So don't
do two year long projects

00:05:12.240 --> 00:05:16.077
and expect that it's gonna be exactly the
same way in two years from now as this,

00:05:16.160 --> 00:05:19.627
- Yes, if you do two year long projects,
plan for where the AI is gonna, I mean,

00:05:19.710 --> 00:05:23.737
that sounds like an obvious thing
to say, but I think it actually takes

00:05:23.820 --> 00:05:25.177
a lot of courage and foresight to do that.

00:05:25.260 --> 00:05:27.907
- It does. For sure. Well, thanks
a lot for taking the time to chat today.

00:05:27.990 --> 00:05:32.500
Really appreciate it.
- Yeah, yeah. Thank you.