WEBVTT

00:00:00.120 --> 00:00:01.960
Matt, thank you so much for
joining us.

00:00:01.960 --> 00:00:05.160
So tell us a little bit about
the work that Clarifai is doing.

00:00:05.280 --> 00:00:06.360
Sure, thanks for having me.

00:00:06.360 --> 00:00:09.720
So Clarifai is really the best
way to think about Clarifai is

00:00:09.720 --> 00:00:11.200
as a full stack AI platform.

00:00:11.440 --> 00:00:14.920
So we have all the tools in one
place for you to take your idea

00:00:14.920 --> 00:00:16.880
and get it into production with
AI.

00:00:17.280 --> 00:00:20.400
And so that means tools for data
labeling, which is always the

00:00:20.400 --> 00:00:23.080
first step in trying to
customize the AI to recognize

00:00:23.080 --> 00:00:26.080
what you care about, managing
that data, training models to

00:00:26.080 --> 00:00:28.280
then recognize how you care
about your data.

00:00:28.640 --> 00:00:30.880
And then you really want to
evaluate them to see how

00:00:30.880 --> 00:00:31.640
accurate they are.

00:00:31.640 --> 00:00:34.960
And then all the production
scale inference you need to take

00:00:34.960 --> 00:00:37.960
that into production and get
feedback from your users,

00:00:37.960 --> 00:00:40.040
interacting with the AI in real
time.

00:00:40.400 --> 00:00:43.680
And just today actually, we
launched a new capability we're

00:00:43.680 --> 00:00:47.120
calling compute orchestration,
and it takes all those workloads

00:00:47.120 --> 00:00:50.680
and lets you run them anywhere,
on any cloud, your own cloud, or

00:00:50.680 --> 00:00:51.520
even on-premise.

00:00:52.360 --> 00:00:55.960
So what challenges do companies
face when taking AI from

00:00:55.960 --> 00:00:57.440
prototype to production?

00:00:58.120 --> 00:00:59.000
Yeah, there's quite a few.

00:00:59.560 --> 00:01:03.240
Some that come to mind are
people kind of overestimate how

00:01:03.240 --> 00:01:07.000
simple it is when they see on
social media, cool, you know,

00:01:07.000 --> 00:01:09.160
hack around, maybe OpenAI's
model.

00:01:09.320 --> 00:01:11.400
Ever since ChatGPT has been a
common trend.

00:01:11.640 --> 00:01:14.760
And you can prompt these models
to do interesting things really easily.

00:01:15.000 --> 00:01:17.080
But then you have to think
about, are you going to trust

00:01:17.080 --> 00:01:18.080
that in front of your users?

00:01:18.120 --> 00:01:20.960
Is it reliably giving you back
that same result that happened

00:01:20.960 --> 00:01:23.720
one time and you posted it on
social media and it's going to

00:01:23.720 --> 00:01:24.680
happen over and over?

00:01:24.680 --> 00:01:26.560
Or is it going to go off the
rails and tell your customers

00:01:26.560 --> 00:01:27.800
something you don't want them to
hear?

00:01:28.160 --> 00:01:30.360
Are you going to be able to run
that cost effectively?

00:01:30.720 --> 00:01:32.960
That's a huge concern with these
large models.

00:01:33.760 --> 00:01:36.160
And that's one of the reasons
why we have this compute

00:01:36.160 --> 00:01:38.440
orchestration, to let you run on
efficient compute.

00:01:39.280 --> 00:01:42.400
And then just thinking about how
you're going to go into day two

00:01:42.400 --> 00:01:44.880
operations. Getting to
production is one thing, but

00:01:44.880 --> 00:01:47.040
then how are you going to
maintain that with

00:01:47.040 --> 00:01:49.760
state-of-the-art, which is
evolving at such a rapid rate

00:01:49.760 --> 00:01:51.760
right now with all the
innovations in AI?

00:01:52.280 --> 00:01:54.800
So there's lots of challenges
going from that first idea all

00:01:54.800 --> 00:01:55.920
the way through production.

00:01:56.320 --> 00:01:58.040
And just to jump off that
really quick.

00:01:58.120 --> 00:02:00.720
So how do companies future proof
their projects?

00:02:01.160 --> 00:02:04.800
Yeah, we see often people are,
you know, building the tools to

00:02:04.800 --> 00:02:08.040
build AI rather than just
focusing on customizing the AI

00:02:08.040 --> 00:02:08.960
for their needs.

00:02:09.280 --> 00:02:11.560
And so they're building
infrastructure tools, they're

00:02:11.560 --> 00:02:13.560
building data-labeling tools,
training models.

00:02:13.560 --> 00:02:17.160
And it becomes like a whole team
just dedicated to trying to stay

00:02:17.160 --> 00:02:19.680
up to speed with
state-of-the-art rather than

00:02:19.680 --> 00:02:22.920
actually innovating with AI to
differentiate your business.

00:02:23.200 --> 00:02:26.040
So we urge people to think about
the long term.

00:02:26.080 --> 00:02:28.920
Do you want to be building AI
tools? Or do you want want to be

00:02:28.920 --> 00:02:30.680
differentiating from your
competitors?

00:02:31.040 --> 00:02:33.800
So that's one big aspect, just
deciding what you want to build

00:02:33.800 --> 00:02:34.600
and what you don't.

00:02:34.920 --> 00:02:38.120
The other is thinking about the
compute long term and your data.

00:02:38.360 --> 00:02:42.800
So compute is crucial and doing
it in a cost-effective way.

00:02:42.800 --> 00:02:46.280
We see this trend of people
purchasing more GPUs to bring

00:02:46.280 --> 00:02:50.040
that compute back on-premise so
that the cloud costs are not,

00:02:50.040 --> 00:02:51.280
you know, ballooning.

00:02:51.640 --> 00:02:55.800
And that's I think a good
strategy for the long term.

00:02:56.160 --> 00:02:59.160
And to be hardware agnostic, I
think, is really, really

00:02:59.160 --> 00:03:00.480
important for the future.

00:03:00.760 --> 00:03:03.480
There's great new types of
AI-specific processors that

00:03:03.480 --> 00:03:05.320
people should be paying
attention to.

00:03:05.800 --> 00:03:09.320
And then on the data front, data
is the lifeblood of AI.

00:03:09.560 --> 00:03:12.360
So the first step is always
getting your data organized and

00:03:12.360 --> 00:03:15.080
thinking about not just the
quantity of data you have, but

00:03:15.080 --> 00:03:15.920
the quality of it.

00:03:16.160 --> 00:03:18.760
Because the better the data, the
better the AI is going to

00:03:18.760 --> 00:03:19.360
ultimately be.

00:03:19.640 --> 00:03:21.080
So it's kind of those big
aspects.

00:03:21.080 --> 00:03:23.440
What do you want to build, where
you're going to run that stuff,

00:03:23.440 --> 00:03:24.680
and then the data that
powers it.

00:03:25.320 --> 00:03:26.240
Thank you so much.

00:03:26.640 --> 00:03:27.040
Thank you.