>> Welcome to IBM THINK 2023! >> AI generated art,
AI generated songs. AI, what is that? It sure is a lot of fun. But when foundation models are
applied to big business, well, you need to think bigger. Because AI and business needs to
be held to a higher standard. Built to be trusted,
secured, and adaptable. This isn't simple
automation that is only trained to do one thing. This is AI that is
built and focused to work across your organization. This isn't committing
to a single system. This is hybrid ready AI that
can scale across your systems. This isn't wondering
where an answer came from. This is AI that
can show its work. When you build AI into the
core of your business, you can go so much further. This is more than AI. This is AI for business. Let's create. (MUSIC) >> Please welcome
Senior Vice President and Director of Research,
IBM, Dr. Dario Gil. (Applause) >> DARIO GIL: Hello. Welcome, welcome to the
last session of THINK. And I understand some
of you even had a drink. How special. So, I hope you've enjoyed
the last two days with us. And what an incredible
year it has been for AI. You can really feel the change
that is happening all around us. And there's just no denying that
the pace of this technology continues to be exhilarating
and that its implications are now so clear for all to
see around the globe. I am just fascinated by AI. And as a technologist, this
level of excitement really comes about only maybe once
or twice every decade. And I am just thrilled to see
all the possibilities that this technology is going to enable. Because it's really going
to impact every industry. From customer care to
transforming data centers and logistics to medicine,
to manufacturing, to energy, to the automotive
industry, to aerospace, communications, you name it. It's really going to impact
every one of our businesses and really touch every
aspect of our lives. So, it's really exciting. And while sometimes the pace
of this technology can feel, you know, daunting and scary,
the opportunities to harness foundation models and generative
AI with proper governance, the opportunities are immense. The emergence of foundation
models and generative AI is really a defining moment. And we need to recognize
its importance, we need to capture the moment. And my advice is, don't
be just an AI user. Be an AI value creator. Just think about it, as an
AI user, you are limited to just prompting
someone else's AI model. It's not your model, you
have no control over the model or the data. Just think carefully about
whether that's the world you want to live in. As an AI value creator, on
the other hand, you have multiple entry points. You can bring your own data
and AI models to Watsonx or choose from a library of
tools and technologies. You can train or influence
training, if you want. You can tune, you can have
transparency and control over the governing
data and AI models. You can prompt it, too. Instead of only one model, you
will have a family of models. And through this creative
process you can improve them and you can make them your
own, your own models. Foundation models that are
trained with your data will become your most valuable asset. And as a value creator, you
will own that and all the value that they will
create for your business. So, don't outsource that. You can simply control your
destiny with foundation models. So, let me show you how we
become, and allow you to become a value creator with Watsonx. Watsonx is our new integrated
data and AI platform. It consists of three primary
parts: First, Watsonx.data it is our massive curated data
repository that is ready to be tapped to train and fine-tune
models, with state-of-the-art data management system. Watsonx.ai, this is an
enterprise studio to train, validate, tune, and deploy
traditional machine learning and foundation models that provide
generative capabilities. And Watsonx.governance,
this is a powerful set of tools to ensure your AI
is executing responsibly. Watsonx.data, Watsonx.ai,
Watsonx.governance, they work together seamlessly
throughout the entire lifecycle of foundation models. And true to our commitment to
hybrid cloud architectures, Watsonx is built on top
of Red Hat OpenShift. Not only does it provide
seamless integration of Watsonx components, it allows
you to access and deploy your AI workloads in any IT
environment, no matter where they are located. WatsonX is the AI platform
for value creators. And look, I don't need to
tell you that deploying these technologies is not easy
at the enterprise level. But the platform changes that. So let's take a look now at how
an entire AI workflow end to end works in the platform. The lifecycle consists of
preparing our data, using it to train the model, validate the
model, tune it, and deploy in applications and solutions. So let's start
with data preparation. So say you're a data scientist
and want to access the data that is in a Public Cloud, some that
is on prem, some that may be in another external database, or in
a Public Cloud, a second one, or anywhere else outside your
hybrid cloud platform. So you access the platform
from your laptop and invoke Watsonx.data. It establishes the necessary
connections between the data sources so you can
access the data easily. We have been building our IBM
data pile combining raw data collected from public sources
with IBM propriety data. We are bringing data from
different domains, the internet, code, academic sources,
enterprise, and more. We have used Watsonx.data to
collect petabytes of data across dozens of domains to produce
trillions of tokens that we can use to train foundation models. And besides the raw data and
our proprietary data, we allow clients to bring their own data
to enrich and improve their purpose-built foundation models. It is all stored in .data. With granular metadata that
provides traceable governance for each file or document. So, now we took this and we move
to filter and process the data. First, we identify the
provenance and the idea of the data. Then, we need to categorize
it, we classify it, for example, in a pile for different
languages, let's say English, Spanish, German, and so on. A pile of code data that we
then separate by programming language, Java, Ansible,
COBOL, and so on. And any other
category that we have. Now we filter it, we
do analytics and get rid of duplicated data. Now identify hate, abuse,
and profanity in the data, and we remove it. We filter it for private
information, licensing constraints, and data quality. By annotating, we allow
data scientists to directly determine the right
thresholds for their filtering. Having done all of that,
the pile is now ready for the next step. We version and tag the data. Each dataset, after being
filtered and preprocessed, receives a data card. The data card has the name
and the version of the pile, specifies its content,
and filters that have been applied to it. And any other relevant content
to make it easy to manage and track different choices that of
the work and the right subsets of the data that we have used to
develop the foundation models. Now, we can have
multiple data piles. They coexist in .data and access
the different versions of data for different purpose
is managed seamlessly. So, we are now ready to
take the pile and start training our model. This is step 2 in
our AI workflow. So, we move from .data to .ai
and start by picking a model architecture from the five
families that IBM provides. These are bed rocks of
models, and they range from encoder only, encoder-
decoder, decoder only, and other novel architectures. Let's pick the encoder-decoder
sandstone to train the model and pick a target data pile version
from the piles that is .data. .ai allows training with
computing resources across the hybrid cloud. In this case it
runs on IBM Vela. Vela is the first of a kind
cloud native AI super computer that we built last year. It gives you bare metal
performance in the cloud with a virtualization overhead
that is less than 5%. And we are making it
available as a service. Watsonx.ai auto scales
the resources for the training being done. And the first thing that we
need to do is to tokenize the data according to the
requirements of the model. So, we first query the data
using the version ID for the pile we want to use. That materializes a
copy of the dataset on Vela for tokenization. What this means is that, for
example, we were building a large language model, the
sentences in the data are broken into tokens. And this process can
create trillions of them. And we use the tokens
to train the model. Now, training is a very complex
and time consuming task. It can require dozens,
hundreds, even thousands of GPUs and can take days,
weeks, and even months. Training in Watsonx.ai
takes advantage of the best open-source technology out there
to simplify the user experience. Built on code flare,
using PyTorch and Ray, it also integrates Hugging
Face to bring you a rich variety of open formats. Once training is done, the
model is ready for validation. So for each model we train,
we run an extensive set of benchmarks to evaluate
the model quality across a wide range of metrics. Once the model passes all
the thresholds across the benchmarks, it is packaged
and marked as ready for use. For each model, we create a
model card that lists all the details of the model. We will have many
different models, trained on different piles, with
different target goals. Next, we go to
Watsonx.governance to combine the data card that has the
detailed provenance information for the data pile that was used
for training, with the model card that has the detailed
information on how the model was trained and validated. Together, they
form a fact sheet. This fact sheet is cataloged
in .governance and all the other fact sheets for all the models
that we have available for use. Now let's go on to tune the
model that we just created, and what we mean by that is to adapt
it to new downstream tasks, which is the basis for the large
productivity gains that is afforded by foundation models. So, say, in this case, you are
a different person, and you are the application developer. So, you can access Watsonx.ai
and start by picking a model from the catalog to work with. We have a family of
IBM models specialized for different domains. But we also have a rich set
of open models, because we believe in the creativity
of the global AI community and in the diversity of
models it offers, and we want to bring that to you. In this case, we pick
sandstone.3b from the IBM language models, which is the
model that we just trained. We set up the options for
tuning, the tuning approach. We pick summarization
as an example, as the base model to use. Now, we can access and use
business proprietary data to tune the base model and for
the task that we choose, whether that business data is
located in anywhere in the hybrid cloud platform. So, now we send prompts and
tuning data, and that's used to tune the model in .ai. You get the outcome of
the prompt on the model. This process happens back
and forth, back and forth many times, and in the end,
you end up with a set of ideal prompts to use. The model is now specialized
and ready to deploy. This is the final step
in our AI workflow. The application where you want
to use the foundation model can live in the public cloud, it can
live on-prem or on the edge. And you can really
deploy and run foundation models efficiently
wherever you need them. And the deployed
model can be used in many different applications. So, for example, we've
embedded foundation models in Watson Assistant. For text generation in
Assistant, you describe the topic that you want the
assistant to handle, and it generates the corresponding
conversational flow. We have an inference stack
to scale the serving of the model in applications. It consists of state-of-
the-art technology that has been field tested for
scalable model serving. This is how Watsonx allows us to
go from data to a model that is trusted, governed, deployed
and ready to serve, and how we can scale that model
to different applications. Once models are deployed, we
continuously monitor them and update them in
both .data and in .ai. We call this constant process
our data and model factory. At Watsonx.governance monitors
the models, if there's any change that may impact how the
model can be used or performs, be driven because we
have new data that can be leveraged or there's a
change in some regulation or law or data licensing. Any change detected by the
.governance process guides and process the update to both
the data and the model. The idea of the model factory
is that is central to solid and proper governance of AI. Now, all of these updates
need to happen without disrupting the underlying
applications that are leveraging the foundation models. And this data and model factory
is in production today. We have already produced over 20
models across modalities like language, code, geospatial
and chemistry, and spanning different sizes of models
from hundreds of millions to billions of parameters. We have infused these
foundation models into IBM products, Red Hat products,
and our partners' products. At IBM, over 12 foundation
models are powering our IBM NLP library, which is used in
over 15 IBM products and is available to ISVs. Granite models train over
code are part of IBM Watson Code Assistant, which has
been applied in the Red Hat Ansible Automation Platform. And as you heard earlier in this
event, SAP has partnered with us and is infusing foundation
models into their solutions. So, Watsonx is really ready for
you to create value with AI. Now, to maximize what you can
do and the innovations at your disposal, we believe that
you should bet on community. Because, the truth is, one
model will not rule them all. And with the innovations and
models that it develops, the open community is super
charging the value that you will be able to create. To be true to our belief in the
diversity and the creativity of the open AI community, we are
proud to announce our new partnership with Hugging Face. So let's invite to the
stage cofounder and CEO of Hugging Face, Clem Delangue. (Applause) >> CLEM DELANGUE: Hey, Dario. >> DARIO GIL: Clem. >> CLEM DELANGUE:
Thanks for having me. >> DARIO GIL: First of
all, welcome to IBM THINK. We are just delighted
to have you here. So let's begin by, tell us
a little bit about yourself and how and when you got
interested in AI and how did Hugging Face get started. >> CLEM DELANGUE: Yeah,
thanks so much for having me. I, actually, started in
AI almost 15 years ago. I look at the room at the time
we couldn't have filled it. Maybe it would have been
one person, two persons in the room at most. As a matter of fact, we
weren't even calling it AI at the time, we were
calling it computer vision. I was working at French company
– I am French, as you can hear from my accents – and we
were doing computer vision on device, on mobile. The company went on to get
acquired by Google after. But I never lost my passion
and excitement for AI. So, seven years ago, with my
cofounders, Jillian Thomas, we gathered around this passion for
AI and started Hugging Face, right, what you see on
my T-shirt, basically. We started with something
completely different. We worked on conversational
AI for three years and as it sometimes happens for startups,
the underlying platform and technology ended up more
useful than the end product. When we started to release part
of it on GitHub, we started to see open-source contributors
joining us, we started to see scientist sharing models in
the platform, leading to what Hugging Face is today. >> DARIO GIL: So I mentioned the
power and the creativity of the open community creating in AI. Just share with us some
statistic, how big is it? How much energy is there in that
community and how much should we expect in the creativity
available to all of us? >> CLEM DELANGUE: Yeah, the
energy in open-source AI is insane these days. Just a few weeks ago I
was in San Francisco. I tweeted that I would be
around and that we could do some sort of a small get-together
for open-source AI people. We thought we would get maybe a
few dozen, few hundred people. And the more the days came, the
more people ended up joining. We had to change locations
three times to something at the end almost as big as
that, we had 5000 people. People started calling
it the Woodstock of AI, so that's just an example. We are competing with
that, the Woodstock of AI. Just proof of how vibrant the
open-source AI community is. We think the same thing
on Hugging Face, right? Since we started on the platform
four years ago, we grew to now having over 15,000 companies
using the platform including very large companies like
Google, like Meta, like Bloomberg, all the way down
to smaller companies like Grammarly, for example. And collectively they have
shared over 250,000 open models on the platform,
50,000 datasets, and over 100,000 open demos. Just last week 4000
new models have been shared on the platform. So, that shows you kind of like
the magnitude and energy in open-source AI community. >> DARIO GIL: Just think about
that, 4000 models in one week. So, one of the myth busting
things that we were chatting about is that the element
of one model will not rule them all, right? There's going to be a huge
amount of innovation that is happening from so many sources. So, perhaps, you could share
with us, what are some examples of innovation that you see? We have seen scale. But what are some examples
that really caught your eye or you think were
particularly powerful? >> CLEM DELANGUE: Yeah, I mean,
it's interesting because since the release of ChatGPT,
right, and some people have said, okay, ChatGPT is a
model to rule them all. 100,000 new models have been
added on Hugging Face, right? And, obviously, companies,
they don't train models just to train models, right? They would prefer not to
do it because it costs money to train models. But the truth is, if you look at
how AI is built, when you can build smaller, more specialized,
customized models for your use cases, they end up being
cheaper, they end up being more efficient, and they end up being
better for your use case, right? Just the same way every single
technology company learned how to write code, right, and to
have a different code base than their competitors or than
companies in other fields. We are seeing the same
thing for AI, right? Every single company needs to
train their own models, optimize their own models, learn how
to run these models at scale. Every single company needs to
build their own ChatGPT because if they don't, they won't be
able to differentiate, they won't be able to create the
unique technology value that they have been building for
their customers, and they lose control, right, if they
start outsourcing it. That's what we are seeing
on Hugging Face and in the ecosystem as a whole. >> DARIO GIL: It's back to the
philosophy of don't be a prompt tuner user, right, be a value
creator with all of this. So let's talk about our
partnership for a minute. Why are you excited about
bringing the power of all of this community into Watsonx,
in the context now of an enterprise, you know, need and
meeting the needs of our clients that are here listening? >> CLEM DELANGUE: Yeah,
obviously, Hugging Face and IBM share a lot of the same DNA,
right, around open-source, open platform, kind of,
like, providing extensible tools for companies. For me, one of the most iconic
collaboration partnership of the last decade is IBM plus Red Hat
and hopefully we are just at the beginning of it, but with
this collaboration, we can do the same thing for AI. I think with this integration
between Watsonx and Hugging Face, you kind of like get the
best of both worlds in the sense that you get the cutting
edge and the community and the numbers of models,
datasets, apps of the Hugging Face ecosystem, and
you get the security and supports of IBM, right? For example, you mentioned,
we mentioned all the models. The IBM consultants can help you
to pick the right models for you at the time that is going to
make sense for your company. So, you really get, kind of,
like, the perfect mix to get to what we were saying, meaning
every one of you being able to build your own internal ChatGPT. >> DARIO GIL: So, tell
us, this is just great. I am just delighted about
those opportunities. So tell us a little bit about
what's next for Hugging Face when you look over the
next year or so, what excites you the most? >> CLEM DELANGUE: Many, many
exciting things for us. We have seen a lot of adoption,
a lot of companies using us for text, for ODO, for image. And now we are starting to see
that expand to other domains, for example, we are seeing
a lot of video right now. We are seeing a lot of
recommender systems; we are seeing a lot of time series. We are starting a
lot of bioG chemistry. We are excited about it, we
think ultimately AI is the new default to build all features,
all workflows, all products. It's kind of like new
default to build all tech. So we are excited for this
expansion to other domains. Also, we are seeing a lot of
work around chaining different models and, in fact, at Hugging
Face we released today a transformer agents which is a
way to chain different models to build more complex systems
that are achieving kind of like better capabilities. These are some of the
things that we are the most excited about. >> DARIO GIL: So a lot
there so thank you, Clem. Thank you so much. Congratulations. >> CLEM DELANGUE:
Thank you so much. (Applause) >> DARIO GIL: Thank you. So, while you saw how the
platform works to enable the foundation model creation
workflow end to end. And we talked about data,
we talked about model architectures, the computing
infrastructure, the models themselves, the importance
of the open community. So, now let me show you
how to use and how you would experience Watsonx. And we are going to go inside
the studio, inside Watsonx.ai. And from the landing page you
can choose two prompt models to fine-tune models or deploy and
manage your deployed models. So here's an example of how
you can use the prompt lab to do a summarization task. You give the model the text
as a prompt, and the model summarizes it for you. In the case of a customer care
interaction, it gives you the customer problem and the
resolution according to the transcript of the interaction. In the tuning studio, as we
saw before, you can set the parameters for the type of
tuning that you want to do and the base model and
you can add your data. The studio gives you detailed
stats of the tuning process and allows you to deploy the tune
model in your application. It's that simple. We took the complexity of
the process away so you only need to worry about creating
value for your business. And here are some of our
current AI value creators. SAP will use IBM Watson
capabilities to power its digital assistant in
the recipe solutions. You have been hearing about Red
Hat, how it's embedding IBM Watson Code Assistant into the
Ansible Automation Platform, BBVA is bringing their
enterprise data to use with their own foundation model
for natural language. Moderna is applying
IBM's foundation models to help predict
potential MRNA medicines. NASA is using our language
models together with US spatial models we have created
together to improve our scientific understanding
and response to earth and climate related issues. And WiX is using foundation
models to gain novel insights for customer care as they meet
the needs of their customers. So, what I encourage you is to
join them and embrace the age of value creation with AI. A year ago, I stood on a stage
just like this, closing THINK. And I shared with all of the
attendees that what was next in AI was foundation models. And maybe at the time it seemed
a little bit abstract and, you know sort of, like, this
intellectual disposition about where things were going. But, boy, what a
year it has been. And it has been a big
year for AI at IBM. So, as we close our event this
year, let me remind of you of all of the things we have
created and announced. We have announced Watsonx, a
comprehensive platform that allows you to create and
governor AI in real time so that you can move with urgency
and capture this moment. We announced a set, a family of
foundation models, including IBM models, open community
models, and how you can even create your own models. We announced our data model
factory, using petabytes of data across multiple domains to
create trillions of tokens to create our family of foundation
models and show how the factory continuously updates them when
conditions change and brings a regular cadence of models to
ensure proper governance. We told you about products
where we have infused our foundation models over 15 of
them, including digital labor, Red Hat products like Ansible
Automation Platform, our partner products like ACP solutions. We announced important
collaborations to advance AI and bring it to the enterprise,
Hugging Face and long-standing collaborations and initiatives
like PyTorch and Ray. We showed you some of the
organizations that have become AI value creators with us. We are bringing IBM Vela or
cloud native AI super computer to train foundation models
with bare metal performance while giving us the
flexibility of the cloud. And we announced that
we are making it available as a service. Last year we launched
the Telum NC16. It's an engineering marvel
and IBM's first processor to have on chip accelerator
for AI inferencing. It can process 300
billion inference per day with one millisecond latencies. This means now you can infuse
AI into every transaction in Z16 for applications
like fraud detection and others in real time. Using the same core architecture
as Telum, we built the IBM Research AIU, which is optimized
to give superior performance for foundation models and enable
with Rat Hat software stack. And at IBM Research we
are incubating powerful AIU systems designed and
optimized for enterprise, AI inference, and tuning. So a truly fantastic year and
this is just the start of all the amazing things that we are
building and developing for you and that we will be sharing
with you in the coming years. So, today more than ever before,
it's important to have a business strategy in AI. And in closing, as you think
about how to harness foundation models for your business, let me
offer you some tips to consider. First, act with urgency. This is a transformative
moment in technology, be bold and capture the moment. Second, be a value creator,
build foundation models on your data and under your control. They will become your
most valuable asset. Don't outsource that and
don't reduce your AI strategy to an API call. Third, bet on community. Bet on the energy and
the ingenuity of the open AI community. One model, I guarantee you,
will not rule them all. Run everywhere efficiently,
optimize for performance, latency, and cost by building
with open hybrid technologies. And finally, be responsible. I can't stress this enough. Everything I have mentioned
is useless unless you build responsibly, transparently,
and put governance into the heart of your AI lifecycle. Continuously governor the data
you use and the AI you deploy. And co-create with trusted
partners, trust is your ultimate license to operate. If you map your AI business
strategy against these recommendations, you will be in
a prime position to do amazing things with foundation
models and generative AI. We have built Watsonx so
that you can do just that. And I hope you join us, because
we cannot wait to get started on this journey with you. Thank you. (Applause)
