E-MAGAZINE

Generative AI
Perspectives From
Databricks, Deloitte, & NVIDIA
Foreword
Generative AI will be transformative, with the potential for embedding human-like intelligence in countless
processes and decisions. At the same time, it will complement — not replace — existing analytics
and machine learning. It is an area of very active research and development, but has real, immediate
applications that enterprises should be working on already. You should thus start now, but expect
continuous evolution and innovation, loosely coupling your applications to underlying infrastructure and
providers.

With Dataiku — the platform for Everyday AI — teams can move beyond the lab and build real and safe
Generative AI applications at enterprise scale. Dataiku brings enterprise-grade development tools, pre-built
use cases, and AI-powered assistants to help everyone do more with Generative AI.

A pivotal piece of that puzzle is our ecosystem, from integrations with leading Generative AI providers to
partnerships with leading consulting firms. With our model- and provider-agnostic approach, teams will
always be able to leverage the most cutting-edge technologies, enabling maximum agility to meet their
business needs.

Just as Dataiku has solidified its position in the data ecosystem as the common backbone for Generative AI
applications (notably via the LLM Mesh), Databricks, Deloitte, and NVIDIA underscore the critical importance
of that ecosystem and are here as our trusted partners to share their various points of view and subject
matter expertise on Generative AI.

- Clément Stenac
Co-Founder and CTO, Dataiku

PREM PRAKASH BEENA AMMANATH SCOTT MCCLELLAN

How Generative AI Is Value With Generative AI Is Generative AI:

Transforming the Business a Matter of Trust Catching the Most
World A Helpful Framework for Promoting Transformative Wave of AI
Trust and Model Governance Innovation to Date

DATAIKU .3
How Generative
AI Is Transforming
the Business
World Prem Prakash
Bio
Prem Prakash is the Head of Product Marketing
for AI, Developer Experience and Learning at
Databricks. He previously led Product Marketing for
Azure AI services at Microsoft.

DATAIKU .5
 Through enhanced and
deeply engaging customer
experiences, automation
of routine tasks, and
A monumental shift is underway in today's out as a game changer, steering businesses toward
dynamic business environment, primarily driven uncharted territories and innovative possibilities.
by Generative AI. As we delve deeper into this Let's unpack the multifaceted impact of Generative

accelerated decision making,

digital transformation era, Generative AI stands AI on modern business.

Generative AI has firmly

The Rise of Generative AI in
Business Operations established its footprint
Gone are the days when AI was a buzzword that impact of Generative AI. It can reshape across sectors.
echoed in tech forums and conferences. Today, it conventional business models, from refining supply
is a cornerstone of innovation, transforming day- chain logistics to offering real-time analytics and
to-day business operations. Through enhanced creating unparalleled customer experiences.
and deeply engaging customer experiences,
automation of routine tasks, and accelerated Due to this unbounded potential, industry analysts
decision making, Generative AI has firmly estimate the benefits of Generative AI to be in the
established its footprint across sectors. trillions of dollars. McKinsey Global Institute
estimates that Generative AI will add between
Every sphere, whether it’s healthcare, finance, $2.6 and $4.4 trillion in annual value to the
retail, or real estate, is feeling the transformative global economy.

DATAIKU .6 7 DATAIKU
 The Investment
Generative AI’s Impact Across
Surge in
Every Organization
Generative AI
While several IT sectors grapple with budget
ELEVATING CUSTOMER
INTERACTIONS • Leveraging Customer Data: By imbibing
constraints, the appeal of AI investments is hard to
customer-specific data into models,
ignore. Despite financial hurdles, there’s a growing
The recent research spearheaded by MIT and organizations can proactively discern, foresee,
inclination toward Generative AI, underscoring its
Databricks sheds light on the future of Generative and cater to the unique requirements of
recognized enduring worth.
AI, emphasizing its pivotal role in customer individual clients, patients, or stakeholders.

An intriguing revelation from a joint study from experience and personalization for the upcoming

An intriguing MIT and Databricks shows that 88% of business years. Here’s how Generative AI stands to revamp
the customer engagement paradigm:
• Tailored Customer Engagements: Whether
it's policyholders nearing renewal deadlines,
leaders are tilting their investments toward
revelation from a Generative AI1. This isn’t a fleeting trend; it reflects
citizens settling taxes, or patients up for health
• check-ups, Generative AI offers automated
joint study from MIT AI’s vast transformative potential. Employee Productivity: Generative AI can
automate manual and repetitive tasks, be a solutions. These span from policy renewals and

and Databricks shows Furthermore, these investments aren’t just personal assistant (to organize calendar or
manage emails), and improve the pace of work
timely appointments to hassle-free product
returns.
whimsical gambles. Business executives
that 88% of business anticipate impressive dividends, spanning by suggesting code to write for developers or
• Automated Document Management: The
leaders are tilting their amplified operational efficiencies, substantial cost content to write for marketers.
technology can autonomously draft or refresh
curtailments, and groundbreaking innovations.
investments toward This confidence underscores the deep-seated • Enhanced Human Interactions: Generative AI
can craft text that mirrors human conversations,
critical documents, such as medical reports or
contracts, without human oversight.
faith in Generative AI’s capacity to spearhead
Generative AI. revolutionary changes across industries. adapting to various languages, conversational
nuances, and emotional undertones. Further, • Iterative Customer Engagement: The models
its integration into search engines ensures swift constantly assimilate feedback, ensuring
1 MIT and Databricks. (2023). (rep.). The Great Acceleration – Perspectives on and accurate answers, thus streamlining user that subsequent customer interactions are
Generative AI. Retrieved from https://www.databricks.com/resources/ebook/
mit-cio-generative-ai-report. queries. continually refined based on past experiences.

DATAIKU .8 DATAIKU .9
STRENGTHENING MODERNIZING SUPPLY
CYBERSECURITY CHAIN OPERATIONS

In today's interconnected digital realm, Generative AI's capabilities extend to optimizing

safeguarding sensitive organizational data is supply chain mechanisms. By assimilating and
critical. Generative AI emerges as a linchpin synchronizing data throughout the supply chain, it
in streamlining and bolstering cybersecurity bridges the gap between upstream anticipations
measures. and downstream deliveries.

• Automated Protection Mechanisms: By • Precision Forecasting: Generative AI mines

automating tasks like threat detection and real- historical data and market tendencies, aligning
time responses, Generative AI mitigates risks them with inventory stats, supplier delays, and
associated with human errors — a significant output to determine intricate patterns and
cause of data breaches. unveil opportunities often missed by traditional
methods.
• Empowering Cybersecurity Teams: A
consolidated AI platform democratizes the • Strategic Supplier Engagement: The
technology, which means even analysts with technology aids in selecting optimum
foundational skills can glean rapid insights, suppliers based on performance metrics,
optimizing costs and freeing seasoned experts quality benchmarks, and cost-effectiveness.
to innovate further. Furthermore, it equips enterprises with insights
to manage these supplier relations effectively.
• Proactive Code Security: Generative AI's
capabilities include scanning infrastructure • Predictive Maintenance Enhancement:
codes, pinpointing vulnerabilities, and Generative AI's feedback loop adjusts to
initiating automated remediations. This not equipment updates, environmental variations,
only maintains code quality but also preempts and usage patterns. This ensures timely
potential security threats. equipment maintenance and proactive
response mechanisms.
While enterprises often commence their Generative
AI voyage eyeing profit augmentation, ensuring
robust cybersecurity is paramount. Neglect here
could lead to spiraling costs tied to cyber incidents,
tarnished brand image, customer attrition, and
regulatory penalties.

DATAIKU .10 DATAIKU .11

 However, beneath this buoyant outlook lies a consensus: For Generative AI to deliver consistently at an
expansive scale and ensure a viable Total Cost of Ownership, certain foundational pillars are indispensable:

Generative AI’s
Potential
Generative AI, once an aspirational concept, has
now matured into a potent tool with tangible
#1
Comprehensive Data Access:
applications. Many open source and easily
customizable models are now within the reach The essence of Generative AI is its ability to tap into
of businesses, marking a pivotal moment in AI's diverse data – regardless of its nature, origin, or
evolutionary journey. Notably, functionalities real-time status.
that seemed distant a mere year ago are now

#2
seamlessly executed by business users, even those
without deep technical know-how.
Industry leaders
unanimously concur In the aforementioned collaborative research
initiative with MIT, Databricks discerned a trend Robust Data Governance:
that Generative AI can of burgeoning optimism. A staggering 80% of
surveyed executives are confident that AI will Central to any successful AI blueprint is
yield dual benefits: amplify operational efficiency by a minimum governance. The strength and relevance of AI

substantial cost of 25% within the next two years. Moreover, a models hinge upon the integrity, accuracy, and
substantial one-third project this uptick to reach or timeliness of the data they draw from. Enterprises
reductions and novel exceed 50%. Such ambitious projections underline must maintain resilient data conduits and update
the enthusiasm surrounding Generative AI and and expand data sources as needed. The AI
revenue streams. its transformative potential across corporate framework.

#3
This is achievable landscapes.

by enhancing Industry leaders unanimously concur that

Generative AI can yield dual benefits: substantial
operational efficiency, cost reductions and novel revenue streams. This
Effective Data Sharing:
expediting insights is achievable by enhancing operational efficiency,
expediting insights derivation, broadening visibility

derivation, broadening scopes, and underpinning decisions with data- Realizing the transformative potential of Generative
driven methodologies. AI mandates capitalizing on the collective
visibility scopes, and strength of an organization's entire ecosystem. By

underpinning decisions Given the impressive ROI trajectories associated integrating essential third-party data channels, like
with Generative AI, it's hardly surprising that 88% the Databricks Marketplace, companies can enrich

with data-driven of executives are investing in it. Furthermore, over their models. This facilitates swifter, more insightful
a quarter of those executives are already in the decision-making, enabling businesses to transition
methodologies. implementation phase. from a passive stance to a proactive one.

DATAIKU .12 DATAIKU .13

 Databricks AI:
With Databricks, you also get LLMOps which is
management of all aspects of the lifecycle–from
data ingestion, model building, tuning, and
A Data-Centric productionization — from a single platform. A
notable feature is its innate governance and

Approach to monitoring capabilities, ensuring the creation

of models within a secure environment where

Generative AI data privacy, permissions, and compliance are

paramount.

For Generative AI to yield impactful outcomes, AI With Databricks MosaicML, businesses have the
models must grasp the specialized terminologies advantage of constructing expansive AI models on
and distinct data sources characterizing each their proprietary data within their private domains,
industry. This underscores the necessity for ensuring swiftness, security, and cost efficiency.
industries — and even individual entities within MosaicML is the most cost-effective way to fine-
them — to approach Generative AI in a manner tune or pretrain models, up to 7x lower costs than
that's uniquely tailored, considering how their pre-training other Generative AI models.
specific data sets can influence the creation of
Generative AI tools.

What stands out is the need for businesses to utilize Unlocking

Success
their distinctive data to carve out Generative AI
strategies that set them apart in the marketplace.

Enter Databricks AI. With its specialized emphasis

Generative AI is not merely a technological trend;
on data, Databricks AI equips users with the tools
it's a revolution. It’s redefining the contours of
to craft and launch AI models swiftly, efficiently,
industries, making operations more innovative,
and with a robust governance framework. This
efficient, and customer centric. As we look ahead,
allows businesses to use their proprietary data as a
businesses that understand, adopt, and innovate
starting point, swiftly constructing and introducing
with generative AI will not just thrive but lead.
generative AI tools tailored to their sector and
niche.
The future is not just automated; it's generative,
promising, and tailored to unique needs, ensuring
Be it the aspiration to create Q&A chatbots
businesses remain relevant, competitive, and
leveraging open source LLMs combined with
successful.
Retrieval Augmented Generation (RAG) or the
ambition to devise bespoke models by fine-tuning
an existing pre-trained LLM with proprietary data,
Databricks AI is equipped for it all.

DATAIKU .14 DATAIKU .15

Value With
Generative AI Is a
Matter of Trust
A Helpful Framework for Promoting
Trust and Model Governance
Beena Ammanath
Bio
Beena leads Trustworthy AI & Technology Trust
Ethics at Deloitte. She is the author of “Trustworthy
AI,” a book that can help businesses navigate
trust and ethics in AI. Beena has extensive global
experience in AI and digital transformation,
spanning across e-commerce, finance, marketing,
telecom, retail, software products, services, and
industrial domains with companies across a variety
of industries.

Beena is also the Founder of the nonprofit, Humans

for AI, an organization dedicated to increasing
diversity in AI. Beena also serves on the Board
of AnitaB.org and the Advisory Board at Cal Poly
College of Engineering. Prior to her joining Deloitte,
she was a Board Member and Advisor to several
technology startups. Beena thrives on envisioning
and architecting how data, artificial intelligence,
and technology in general, can make our world a
better, easier place to live for all humans.

DATAIKU .17
 As the marketplace of
The world has entered Generative AI solutions
a new era with AI. matures, models that can
create text, images, code,
Beginning in earnest in 2022, the release of several Of course, new technologies often conjure
consumer-facing, headline-grabbing Generative new questions around risk and trust. Even as
AI applications captured public attention and organizations are contemplating how to leverage
stimulated interest among organizations and
across industries. Innovation has come fast,
and today, novel Generative AI applications are
Generative AI for efficiency gains, tailored offerings,
product improvements, new experiences, and
business transformation, attention should also be
video, and much more will
emerging in areas such as hiring and workforce
management, medical diagnosis and treatment
given to the governance and risk management that
makes these use cases trustworthy and valuable. serve as an enabler of work
and creativity.
recommendations, supply chain efficiency and
constraint management in manufacturing, and To pursue confidence and trust in Generative AI,
internal controls for financial reporting. there are a variety of considerations and strategic
actions that can help govern the model lifecycle
The potential with Generative AI is grand in and balance commercialization, regulation, ethics,
part because of how this type of AI functions. co-creation, risk management, and ultimately,
Whereas deep learning, for example, is valuable trust. The most impactful approaches are found
for discovering patterns and making predictions, in the familiar areas of people, processes, and
Generative AI is characterized by its capacity to technologies.
create data that appears as coherent as its training
set. As the marketplace of Generative AI solutions
matures, models that can create text, images, code,
video, and much more will serve as an enabler of
work and creativity.

DATAIKU .18 19 DATAIKU

Preparing the
Workforce for the
Generative AI Era
As well as AI literacy, stakeholders may also
need new skills specific to their practice
area:
The technical function of Generative AI models may
only be deeply understood by the organization’s
• Compliance and regulatory leads need
data and IT professionals, but governing AI is
familiarity with the technology so as to
a whole-of-enterprise activity. The onus for
understand the potential impact from
identifying and managing risk is shared throughout
AI laws and regulations emerging in
the organization, including with implementation
regions where the business operates.
leads, development leads, data governance
managers, peer reviewers, security experts, and line
• Marketing and communications
of business users.
professionals need an understanding
of how to leverage Generative AI
What is more, in a time when workforces are capabilities given questions around
evolving, stakeholders can also include vendors, intellectual property and copyrights.
partners, and contingent workers. For these
professionals to take part in Generative AI • Human relations leaders need to
governance, they require a clear understanding of see how Generative AI applications
their roles and responsibilities in AI development, can improve workforce programs,
deployment, and outcomes. employee support, and third-party
engagement.
Stakeholders also require AI literacy as it relates to
Generative AI applications. They need a working • Business domain experts need to
understanding of the AI model and its use cases, collaborate with computer and
its drawbacks and limitations, as well as the risks data scientists to translate business
it may create. To promote Generative AI literacy, objectives into value-driving AI use
business leaders can leverage existing workforce cases.
training and knowledge transfer sessions between
technical and non-technical stakeholders. • The C-suite and the organization’s
board need to understand the vision
The organization may also offer or promote online for how Generative AI enhances the
education, industry conferences, or seminars, all business and the risks it may create
to nurture a culture of continuous learning. This along the way.
is particularly important with Generative AI, a
technology that is changing and improving at a
rapid pace. Taking part in risk management and
governance requires stakeholders to keep up with
developments in Generative AI.

DATAIKU .21
Assessing New Risks to
Inform New Processes
Generative AI will likely be an integral part of the future of work, with The first step is to understand the risk domains and where gaps exist by A trustworthy use case is assessed for its capacity to be fair and impartial,
widespread adoption across industries and lines of business. This will conducting a risk assessment and analysis. No two AI applications are robust and reliable, transparent and explainable, safe and secure,
amplify human-machine collaboration in the Age of With™, as Generative precisely the same, fueled as they are by proprietary enterprise data. In accountable and responsible, and respectful of privacy. Identifying which
AI enables tools that supplement and enhance the human workforce. This conducting risk assessments, a helpful tool is Deloitte’s Trustworthy AI™ of these risk domains is relevant to a given use case serves as the basis for
will lead to changes in how work is accomplished, which has implications framework, which helps the enterprise identify elements of trust that may designing risk mitigation controls, responses, and other AI Governance
for business processes and workflows that support Generative AI risk apply to a Generative AI deployment. processes. It also informs key governance questions, including:
management and governance.

DATA GOVERNANCE AND PROCESSING: How does the

organization obtain, store, transfer, and curate the data used
for Generative AI? Is third-party data free of bias, and how is
that determined? Has personally identifiable information been
appropriately obscured so that it does not “leak” through the
deployment?

SECURITY: How does the organization secure data used by AI

models, as well as the models themselves? What kinds of security
vulnerabilities exist, and what are the tactics a bad actor may
employ to compromise the model (e.g., prompt-spoofing)? At
which points in the Generative AI lifecycle is safety and security
addressed, and how is it assessed and maintained over time?

ORGANIZATIONAL GOVERNANCE: How does the organization

define testing objectives? How does it monitor and evaluate
results from controls and processes? How are stakeholders
empowered to provide input, feedback, and concerns
throughout the AI lifecycle? Are there additional roles that could
be created to support stakeholder participation and input, such
as an AI ethics board or even a Chief AI Ethics Officer?

MODEL DEVELOPMENT AND EVALUATION: How does the

organization validate AI model performance? Who is involved
in that endeavor, and what tools and technology platforms can
help?

DATAIKU .22 DATAIKU .23

 Looking Ahead to
the Future With
Leveraging
Generative AI
Technology for These approaches for orienting people, processes,

Model Governance
and technologies can help organizations
operationalize trustworthy Generative AI.

Today’s Generative AI models often come with a However, in an increasingly crowded marketplace
disclaimer that the outputs may be inaccurate. of solutions, organizations are challenged to not
The model itself would be forgiven as Generative only implement governance for these powerful
AI creates data that looks like real data. It has no new tools but also keep up with advances and even
capacity to understand whether the generated data anticipate capabilities and use cases just over the
is valid. That task falls on the human stakeholders. horizon. A professional services collaborator with
deep experience in trustworthy AI development
The issue of reliability points to a larger challenge and application can offer critical support as the
in AI Governance, which is how to understand, enterprise establishes Generative AI Governance.
evaluate, and measure the performance of a tool
whose technical function may not always be Deloitte helps clients take a three-pronged

By deeply understandable. The “black box” problem with approach to Generative AI Governance:
other types of AI is perhaps even more perplexing

understanding how with Generative AI models. Large language models, • Evaluating roles and responsibilities,
for example, can have billions of parameters. The supporting change management, and
a model performs, scale defies human comprehension. conducting workforce training.
Sources:
•
an organization can Offering services around AI strategy and
To help ensure accountability, organizations governance, risk management, and AI

enhance other aspects may leverage technology solutions that help assessments. • Responsible Use of Generative AI
evaluate and track model performance. This allows • Using tools to monitor AI function and • Trustworthy AI Governance in Practice
of Generative AI the enterprise to evaluate whether an AI tool performance, such as Dataiku’s platform for
• Services: Trustworthy AI™
Governance, including performs as intended and aligns with the relevant assessing, managing, and documenting each
• Report: Trustworthy AI in Practice
dimensions of trust. step of model development and operation.

necessary workforce • Article: Proactive Risk Management in

Generative AI
Importantly, the product of technology-enabled Across the spectrum of professional services —
upskilling, stakeholder assessments is not just improvements to the from strategy consulting to advisory support • Article: The Implications of Generative AI for

engagement, model itself. By deeply understanding how a model to audit and assurance — Deloitte can offer the Businesses
performs, an organization can enhance other approach and offerings that help you take the right
• Document: Generative AI Is All the Rage
and process aspects of Generative AI Governance, including steps, make sound decisions, and implement the
• Article: Benefits and Limitations of
necessary workforce upskilling, stakeholder processes and programs that help you maximize
transformation. engagement, and process transformation. value and trustworthiness with Generative AI. Generative AI

DATAIKU .24 DATAIKU .25

Generative AI:
Catching the Most
Transformative
Wave of AI Scott McClellan
Bio
Innovation to Date
Scott McClellan is a senior director of product
management at NVIDIA, focused on data
science and other AI workflows. Before joining
NVIDIA, Scott was the chief technology officer
of PRGX Inc. He has been chief technologist and
led engineering and product development at
companies including RedHat and HP, where he
guided strategies across HPC, cloud, big data and
AI solutions. Scott holds a Bachelor of Science from
the University of Iowa.

DATAIKU .27
 The amount of data required
to train a foundation model
varies but is typically large,
often involving millions or
Today, nearly every executive or decision maker
responsible for IT or application technology
FOUNDATION MODELS even billions of examples, to
achieve high performance
strategy has seen an avalanche of excitement about A Generative AI system “learns” underlying
Generative AI or Large Language Models (LLMs). characteristics from training data and will use that
“knowledge” to generate a model based on those
Generative AI and LLMs represent two of the latest
waves of innovation in AI research. This feature
gives an overview of both technologies, highlighting
characteristics — this model is called a “foundation
model.” The amount of data required to train a and generalization across
diverse tasks.
foundation model varies but is typically large, often
some key considerations for safely surfing the involving millions or even billions of examples,
Generative AI/LLM technology wave. to achieve high performance and generalization
across diverse tasks.
Generative AI
Foundation models have broad, general
Generative AI refers to a class of AI capable of capabilities and can be customized to perform
generating net new content based on a variety specialized tasks. The potential need for
of inputs. The ability to create an artifact that customization and the methods for customizing
does not exist in the training data explicitly is the models are discussed in more detail later in this
fundamental distinction between Generative AI and feature.
more traditional AI.

DATAIKU .28 29 DATAIKU DATAIKU .29

 Large Language Models
LLMs are a form of Generative AI capable of
generating coherent and relevant text based on the
input they receive. These models can be used for
various natural language generation tasks, such as
text summarization, language translation, and even
creative writing.

In the case of an LLM, the foundation model is

trained on massive amounts of diverse text data.
The use of self-supervised training has drastically
expanded the scale and scope of data LLMs are
trained on, effectively allowing foundation models
Figure 1: Generative AI
to “learn” from the vast and diverse information
available on the internet. Even with the most
advanced techniques for parallelizing the training,
and taking advantage of GPUs to greatly accelerate
TRANSFORMER ARCHITECTURES MULTIPLE MODALITIES the training process, training a foundation model Even with the most
advanced techniques
on “internet scale” data takes a long time (days to
Inputs and outputs to these models can include weeks) and costs millions of dollars.
Foundation models are based on “transformer
architectures,” a type of neural network designed text, videos, sounds, images, animation, 3D models, for parallelizing the
or other types of data. Using an LLM in an application involves taking a
to take an input sequence and use it to generate an
output sequence one element at a time.
sequence of data called a prompt, feeding that training, and taking
advantage of GPUs
Generative AI can be multi-modal, meaning a given prompt to the LLM (in a process called inference),
AI pipeline can be designed to understand input and getting back a generated sequence of text
For example, a transformer could be used to
translate a sentence in English into a sentence data in multiple or mixed formats and can generate which is the model’s response to that prompt. to greatly accelerate
output in multiple or mixed formats.
in French. The process of translation could take
the sequence of words in English as input, and
The LLM generates the response by predicting a the training process,
training a foundation
new sequence based on the prompt it was given
the transformer would “predict” a corresponding
as input. The response is new content generated
sequence of words in French.
by the LLM. This process differs from search, where model on 'internet
a search engine finds existing text that matches
The key innovation in a transformer architecture
is the self-attention mechanism, which allows the
instead of generating new text as a response. scale' data takes a long
model to weigh the importance of different parts of
Through this process, LLMs are able to capture the time (days to weeks)
the input data when making predictions, improving
the quality/fidelity of predictions at a given
underlying patterns and semantic relationships, and costs millions of
gain a contextual understanding of language, and
computational budget.
use that to generate very human-like responses. dollars.
DATAIKU .30 DATAIKU .31
 TECHNICAL SKILL, DOMAIN
EXPERTISE

Similarly, users could choose a topic that required

technical or domain expertise — picking a question
they were confident ChatGPT could not answer.
The ChatGPT Effect Below is an example where ChatGPT is asked to
write code in the programming language Rust.
The watershed moment for many major technology
trends can be traced to the availability of a “killer
application” which showcases the power and
value of that technology. OpenAIs’s ChatGPT
application made it possible for anyone to ask
ChatGPT their favorite “stumper” question.

HUMAN-LIKE QUALITIES

Skeptical users would not expect a computer

program to demonstrate human qualities such as
reasoning, intuition, empathy, humor, creativity,
etc. For example, asking ChatGPT to write a poem
or a short story.
Figure 3: ChatGPT code generation example

Note how ChatGPT was able to take a very simple

prompt (which included a typo — “poen” not The ChatGPT output above was truncated, but
“poem”) and was able to generate a poem about the prompt generates working code, as well as a
Maui that stayed within the requested word count. description of how to install and use the code.

This example illustrates a simple form of prompt

engineering, using a detailed prompt, delivered to
ChatGPT to generate functionally complete source
code. The prompt is not perfect but ChatGPT
clearly understood what the user wanted. The
user could also ask ChatGPT to modify the code,
or ask for an explanation of the code, showcasing
how an application built on LLMs can be very
Figure 2: Example poem generated by ChatGPT conversational and natural to use.

DATAIKU .32 DATAIKU .33

 Use Cases Adoption Considerations CUSTOMIZATION

When using any form of Generative AI for a given Here we will characterize use of Generative AI
The consensus is that Generative AI has the
enterprise use case, it will likely be necessary to do by putting it into one of three buckets based on
potential to completely transform many
some degree of customization. How quickly and degree of customization.
enterprises, creating competitive differentiation,
improving existing business, and in some cases easily you can take advantage of LLMs will depend
creating whole new business opportunities. on how complicated your use case is and the
degree to which you need to customize.
Below are a few examples of enterprise use cases
for AI by industry.

FINANCE HEALTHCARE

Fraud Detection Molecule Simulation

Personalized Banking Drug Discovery
Investment Insight Clinical Trial Data Analysis

MEDIA &
ENTERTAINMENT
MANUFACTURING

Character Development Factory Simulation

Video Editing & Image Creation Product Design
Style Augmentation Predictive Maintenance

RETAIL MANUFACTURING Less Customization Moderate Customization Extensive Customization

How quickly and For some use cases, the requirements can Prompt-based learning and fine-tuning Finally, some use cases will require
Personalized Shopping AI Virtual Assistants
easily you can take
be met using a pre-trained foundation introduce a comparatively small number extensive fine-tuning or a new custom
model, without customization of the of parameters or layers to an existing foundation model. The infrastructure
Automated Catalog Descriptions Network Performance Tuning model itself. Prompt engineering can be foundation model. This requires an costs will be significantly higher and will

advantage of LLM
used to optimize the prompts sent by additional training step with additional require much more data. Creation of a
Automatic Price Optimization Remote Support Capabilities
applications to the LLM (during inference) specific data, but allows the model to new foundation model is much more
to get optimal results. The application cover new use cases and provide higher complex to train and deploy and requires

will depend on how FEDERAL ENERGY

can use a hosted model accessed via an
API. This minimizes the infrastructure
accuracy than prompt engineering
alone. Moderate customization will
more organizational maturity and
experience with AI in production systems.
requirements and allows the application have additional compute and data
complicated your use to use a cloud-based service, paying for
infrastructure on a “pay-as-you-grow”
curation requirements, and thus higher
infrastructure costs.
Document Summarization Knowledge Base Q&A
case is and the degree
model.
Audit Compliance Predictive Maintenance

to which you need to AI Virtual Assistants Customer Service

customize. Figure 4: Generative AI Use Cases by Industry

DATAIKU .34 DATAIKU .35

 Document Curation

As outlined in the above, there are use cases where

you will need to augment an existing pre-trained
model, or potentially, create your own foundation
model.

Injecting additional data into the workflow,

introduces several topics related to curating that
data.

Augmenting with additional

DATA VARIETY:
SECURITY
CHALLENGES:
data adds complexity, but
also creates the opportunity
Data from multiple
Ensuring valuable
sources, in a
proprietary data is
multitude of formats.
not exposed.
AUGMENTING WITH ADDITIONAL
DATA for an enterprise to use
There will be cases where you will want to use
additional data to get the desired results. DATA
their own data and domain
knowledge as an asset — to
DATA PRIVACY: GOVERNANCE:
The simple intuition here is that an LLM can’t Ethical and The additional data
generate quality output for domains or topics not regulatory issues may have different
covered in the data used to train the foundation
model. This is particularly true for domains where
new data brings new context on a regular basis.
related to private or
confidential data.
rules that govern
how it can and provide differentiated value
in their application.
cannot be used.
Augmenting with additional data adds complexity,
but also creates the opportunity for an enterprise
to use their own data and domain knowledge as
an asset — to provide differentiated value in their SCALE
DATA VERSIONING
application. CHALLENGES:
AND EXPERIMENT
Associated with MANAGEMENT:
For example: If you were building an application manipulating and
Necessary to support
to provide advice on building maintenance, and storing data for fine-
the iterative model
you had thousands of technical documents, tuning or building
development/
maintenance logs, etc., you could fine-tune a pre- your own foundation
deployment process.
trained model with that proprietary data, adding to model.
the knowledge the foundation model already has.

DATAIKU .36 37 DATAIKU DATAIKU .37

 NVIDIA AI
Solutions
NVIDIA has a complete catalog of software for

Summary
accelerating all aspects of AI — everything from
data processing to traditional machine learning
and deep learning training and inference —
and a multitude of frameworks to support
different general and industry specific use cases. Generative AI represents the latest and most
Collectively, this is referred to as NVIDIA AI promising breakthrough in AI. In a very real
software. sense, Generative AI is the crowning achievement
of decades of research and development in
NVIDIA NeMo Framework is an end-to-end, cloud- deep learning, building on many of preceding
native framework to build, customize, and deploy breakthroughs in AI.
Generative AI models anywhere. It includes training
and inferencing frameworks, guardrailing toolkits, LLMs are a specific type of Generative AI, capable
data curation tools, and pre-trained models, of generating human quality responses to textual
offering enterprises an easy, cost-effective, and fast prompts — fundamentally changing the way
way to adopt Generative AI. humans and computers interact.

NVIDIA NeMo Framework is included with NVIDIA Given the vast potential, all enterprises need to be
AI Enterprise, the enterprise-grade software that exploring how to take advantage of Generative AI
powers the NVIDIA AI platform and accelerates and LLMs.
enterprises to the leading-edge AI with security,
stability, manageability, and support. NVIDIA AI Adoption and deployment of these technologies
Enterprise offers extensive frameworks, pretrained requires careful consideration, notably with
models, and development tools to streamline respect to model customization, and operational
development and deployment of production challenges like data curation.
AI, enabling organizations to harness the game-
changing insights of AI instead of maintaining a Solutions like NVIDIA’s NeMo Framework will
high-performance software platform. help enterprises with these key challenges, and
NVIDIA AI Enterprise will provide enterprises with
Click here to get the latest information from NVIDIA the stability, security, and support they need to use
on Generative AI. these AI in critical production applications.

DATAIKU .38 DATAIKU .39

 Everyday AI,
Extraordinary People
Dataiku is the platform for Everyday AI, enabling data experts and domain experts to work together to
build data into their daily operations, from advanced analytics to Generative AI. Together, they design,
develop and deploy new AI capabilities, at all scales and in all industries.

Elastic Architecture
Built for the Cloud

Machine Learning Visualization Data Preparation

Name Sex Age

Natural lang. Gender Integer

Braund, Mr. Owen Harris male 22

Moran, Mr. James male 38

Remove rows containing Mr. female

ques Heath female
Keep only rows containing Mr.
y male
Split column on Mr.
Mr. Robert male
Replac
ry e Mr. by ...
D Kingcome)

Remove rows equal to Moran, Mr. James

Keep only rows equal to Moran, Mr. James

Clear cells equal to Moran, Mr. James

Filter on Moran, Mr. James

Filter on Mr.

Toggle row highlight

Show complete value

Governance
DataOps & MLOps Analytic Apps Generative AI

©2024 dataiku | dataiku.com

DATAIKU .40
©2023 DATAIKU | DATAIKU.COM