Started on Wednesday, October 18, 2023, 5:48 AM

State Finish '

Completed on Wedn Back Next 5 AM

Time taken 27 mins 7 secs
Feedback Congratulations, you passed the quiz!

Question1

Correct

Points out of 1.00

Generative AI has been used by millions of users worldwide

for various tasks, from writing emails to marketing brochures,
to generating reports, to even passing university exams. A
client's Chief Information Officer (CIO) lamented that these
are interesting but hardly meaningful in a business context.
You assure this CIO that generative AI is not just useful for
grabbing interesting headlines but it is great for businesses as
well. What are some valid business use cases for generative
AI?

Perform strategic extrapolation on competing enterprises

based on their past behavior, acquisition history, and
gathering information about these companies from the
web.
Performing summarization of text, extracting of
insights, generating new information, classification for
sentiment or topics, and Question & Answering (Q&A).
Performing data sharing with different teams and
companies using the same encoding-decoding foundation
models. Users can pass in data and encode it with the
model. The recipient will decode and receive the data
when decoding it with the same model.
Automatically generate new company HR policies
whenever there is a question from an employee that has
no clear answer from existing policies.
Question 2

Correct

Points out of 1.00

Back Next

A client has been working with traditional AI and is dissatisfied

with the amount of time and resources that are needed to
train a model for each specific task using their proprietary
data. What is a valid reason why they should consider using
foundation models?

Foundation models refer to a huge collection of models

pre-built for all common verticals. Clients simply pick the
right foundation model to use.
Foundation models only need a small amount of
labeled data to fine-tune for a specific task.
Foundation models are much cheaper to run than
traditional AL Most foundation models can run on regular
CPU nodes.
Foundation models are customizable by all vendors.
Clients can supply the data and have the vendor build a
customized model from the ground up.

Question 3

Correct

Points out of 1.00

While generative AI has huge potential, there are also many

inhibitors. According to IBM Institute for Business Value
research: "Generative AI: The state of the market", what are
the 4 main concerns, inhibitors, and fears companies have on
adopting generative AI?

Explainability, Bias, Trust, and Ethics

Explainability, Transparency, Drift, and Ethics
Trust, Transparency, Tunability, and Ethics
Explainability, Ethics, Drift, and Tunability
Question 4

Incorrect

Points out of 1.00

Back Next

A client has experience working with traditional AI for many

years. They are now looking into generative AI and foundation
models. They know, despite all the press, that models won't
"just work" with their applications and workflows without
some modifications. They have done some research on
prompt tuning, prompt engineering, and fine-tuning. They
have come to you with questions about these processes.
Which of the following is a correct response to share with the
client?

Both prompt tuning and prompt engineering requires

inputting labeled data into the model for further training.
Fine-tuning involves adjusting a set of parameters such as
learning rate, number of epochs, etc.
Prompt engineering includes zero-shot learning. Prompt
tuning and fine-tuning both require user data. Fine-tuning
will change the underlying model weights.
In terms of complexity, fine-tuning is the most x
complex, followed by prompt engineering, and the
simplest to do is prompt tuning.
Prompt engineering will not change the weights of the
underlying model. Both prompt tuning and fine-tuning will
change the weights of the underlying model.
Question 5

Correct

Points out of 1.00

Back Next

The data engineers state that ChatGPT "can do all sorts of

amazing things" for their company. The Chief Information
Officer (CIO) is very excited about the opportunity and is
talking about incorporating ChatGPT to achieve new levels of
efficiency in business workflows. How would you caution the
CIO about selecting a generative AI platform and what are
some valid questions they should ask about the foundation
models they might consider using?

Foundation models require a massive amount of diverse

data to train that only the open-source community can
truly provide. Clients must look for platforms that support
open-source models for production. Proprietary models
are useful only when used with products from the same
vendor.
Ultimately, cost is the only thing that truly matters to an
enterprise. Everyone can claim to be able to do generative
AI, but this client should test to see what vendor model
can complete a task at the lowest cost.
Building foundation models require a huge amount of
resource and only hyperscalers have the true ability to
train the models. Clients must look for models that come
from IBM, Amazon Web Services (AWS), Microsoft Azure,
or Google only. Open-source models are only good for
testing new tasks.
Clients should understand that the quality of data used
to train a model is important. They also need to know
how the AI model detects and corrects for bias, how to
prevent hallucinations, and what data privacy and
regulatory compliance standards are in place.
Question 6

Correct

Points out of 1.00

Back Next

A Cloud Pak for Data client has been using services like
Watson Studio, Watson Assistant, and Watson Discovery and
they think that watsonx.ai is just a clever rebranding. You want
to clarify that watsonx.ai provides additional services on top of
what's available in Watson Studio today. What should you
mention?

Watsonx.ai's added value is in providing a unified

interface to all of IBM's Watson AI tooling such as Watson
Studio, Watson Assistant and Watson Discovery. Clients
can access them all in one single place.
Watsonx.ai comes with a library of pre-trained
foundation models. It provides a Prompt Lab where
clients can try out different models and performance
prompt with prompt engineering.
Watsonx.ai incorporates Guardium and QRadar, and this
provides additional security features that are not available
in Watson Studio.
Watsonx.ai is a platform that allows users to take models
from SageMaker, Azure Machine Learning, and Google
Vertex AI and run it natively. It allows clients to deploy
one single AI platform and run all of their existing models.
Question 7

Correct

Points out of 1.00

Back Next

A client is training a group of data scientists on foundation

models and generative AL They understand it is important to
work with prompts and gain some experience with a model's
output. This client is interested in the IBM watsonx.ai Prompt
Lab. What are some of the parameters they can try to modify
in the Prompt Lab?

Choice of foundation models, percent of base data to be

included, number of maximum tokens, maximum time for
creating response to prompt.
Types of decoding, repetition penalty, stop sequence,
number of minimum output tokens.
Types of decoding, timeon value for cluster graphical
processing unit (GPU) usage, stop sequence, minimum
time in 10 seconds increment that the model must use
before providing results.
Repetition penalty, percent of memory usage lower and
upper limit, number of prompts allowed.

Question 8

Correct

Points out of 1.00

Generative AI is a consideration for many clients' executives

when they are considering an AI adoption and deployment
strategy. Executives should consider an AI vendor with a
sound strategy that includes data used for model training,
governance, regulation, and compliance. What are the
components of IBM watsonx that seek to address these end-
to-end strategic requirements?

watsonx.ai, watsonx.query, and watsonx.governance

watsonx.studio, watsonx.governance and watsonx.data
watsonx.ai, watsonx.data, and watsonx.governance
watsonx.foundation, watsonx.generative, and watsonx.ai
Question 9

Correct

Points out of 1.00

Back Next

Clients are increasingly concerned with being locked into a

single cloud platform since many enterprise businesses have
data in multiple clouds as well as on-premises. Having an AI
platform that is anchored on a particular cloud is not
desirable. How does IBM watsonx.ai (and in general watsonx
and IBM Cloud Paks) differentiate itself in this regard?

Watsonx.ai (as well as watsonx and IBM Cloud Paks)

are built on Red Hat OpenShift, it can be deployed on¬
premises or anywhere in the cloud.
Watsonx.ai (as well as watsonx and IBM Cloud Paks) are
available on VMware images. These can be deployed
wherever the clients have VMware set up.
Watsonx.ai (as well as watsonx and IBM Cloud Paks)
installable images are created for Linux (Ubuntu, RedHat,
and Debian) as well as Windows 11. They can be deployed
anywhere with these virtual OS images.
Watsonx.ai (as well as watsonx and IBM Cloud Paks) are
written in C and can be easily re-compiled to be deployed
on other cloud platforms
Question 10

Incorrect

Points out of 1.00

Back Next

Your client has been building their own AI models for a long
time and has successfully applied these models to solve their
business requirements. They have heard that foundation
models are trained with "massive amounts of data" and feel
they are the same as the models they have been building.
What would be a correct statement to explain the difference
between traditional models and foundation models?

The client is correct that both require training with

massive amounts of data. Traditional AI models are
language-specific, but foundation models are multi¬
lingual.
Traditional AI models are trained with massive amounts of
unlabeled data, whereas foundation models are trained
with a massive amount of labeled data.
The client has incorrect information. Both traditional AI
models and foundation models are trained with
*
massive amounts of unlabeled data. Traditional AI
models are task-specific, but foundation models are
not.
Traditional AI models are trained with massive amounts of
labeled data, whereas foundation models are trained with
a massive amount of unlabeled data using self-supervised
learning.
Question 11

Correct

Points out of 1.00

Back Next

Amazon Web Services (AWS), Microsoft Azure, and Google

Cloud Platform (GCP) have all entered the foundation model
race with their own platforms. What is true about these
offerings?

Azure OpenAI is not multi-cloud-enabled, but it can be

made available on-premises via Azure Stack.
Google Generative AI Studio stores everything on Google
Big Query and is available on-premises or in other clouds
via Google BigQuery Omni.
AWS Bedrock, Azure OpenAI Service, and Google
Generative AI studio are only available on their
respective cloud and nowhere else.
AWS Bedrock is a fully integrated suite with governance
already built in via AWS Glue.
Question 12

Correct

Points out of 1.00

Back Next

Trust in foundation models begins with trust in the data. If a

model is trained with "bad" data, then it will provide poor
results. One of the strengths of watsonx.ai is in the quality of
the data used to train IBM's proprietary foundation models.
What makes IBM's data superior to other data that is used to
train foundation models offered in the market today?

IBM encrypts all data before putting it in watsonx.data.

Only this data is used to train models. The integration with
watsonx.ai allows the data to stay encrypted for training -
hence there is no chance of corrupted data for model
training.
IBM only uses in-house data. Because the data is used by
IBM, it is already vetted and curated, and it can be trusted
since IBM has one of highest quality standards in the
industry.
IBM only uses data from open-source repositories
(GitHub, Hugging Face, OpenAI). These elevate IBM
models to the same standard as OpenAI's GPT 4.0 model
which is much better tuned than ChatGPT.
IBM performs data deduplication (exact and fuzzy
dedup) as well as extensive data annotation (filtering
out hate, abuse, profanity), checking data quality,
detecting copyright and license material.
Question 13

Correct

Points out of 1.00

Back Next

A client is eager to test watsonx.ai and has signed up to try it.

They have asked you to provide some information about how
to use the prompt lab. Which of the following is a correct
statement about the watsonx.ai Prompt Lab?

The Prompt Lab is where clients perform prompt

engineering and prompt tuning. The Tuning Studio is only
for fine-tuning.
On the watsonx.ai user interface, clients will go to the
Prompt Lab to run prompt tuning and use the Tuning
Studio for prompt engineering.
The Prompt Lab is where clients can perform prompt
engineering: working with zero-shot, one-shot, and
few-shot prompting.
The Prompt Lab is where users can provide their labeled
data to fine-tune the model.
Question 14

Incorrect

Points out of 1.00

Back Next

Since ChatGPT was released by OpenAI in November 2022

many other vendors have entered the generative AI space. A
client is feeling pressure to implement AI into their business
workflows or fall behind the competition. They want to know
why they should consider IBM watsonx.ai over other
generative AI vendors. Which of the following are valid
watsonx.ai differentiators?

IBM watsonx.ai supports both traditional and x

generative AI. IBM offers open-source models from
Hugging Face, proprietary models from IBM and
vendors such as AI21, Anthropic, and Stability AI.
IBM watsonx.ai is part of IBM Cloud Pak for Security
which gives clients access to all the security features
available from IBM Cloud Pak for Security, including
Guardium and QRadar.
IBM watsonx.ai is built on open technologies, designed to
be trusted, targeted toward business use cases, and
empowers value creators.
IBM has a partnership with Hugging Face and all of the
models available from Hugging Face are available and
integrated with Watson Studio
Question 15

Correct

Points out of 1.00

Back Next

A client has been working with Watson Studio and Watson

Machine Learning for years and have developed skills;
however, they are skeptical about generative AI - seeing it as
more of a fad than having the ability to deliver true business
value. They ask you to show them the difference between
traditional AI and generative AI. Which of the following is true?

Traditional AI models do not require supervised learning

and are adaptable to different tasks. Generative AI
foundation models require supervised learning.
Both traditional AI models and generative AI foundation
models are adaptable and transfer well to support
different tasks. Generative AI models are just less
expensive and less time-consuming to train.
Traditional AI models are general, once trained with
labeled data they apply to many different tasks.
Generative AI foundation models cannot be trained for
specific use cases.
Traditional AI and the models built with this method
are typically task-specific. Generative AI foundation
models are pre-trained with unlabeled data and can be
adapted to multiple tasks.
Question 16

Correct

Points out of 1.00

Back Next

Your client listened to an IBM webinar about generative AI and

they are confused about the trade-off and differences
between prompt engineering and prompt tuning. What are
some important points you should highlight about prompt
engineering and prompt tuning?

With prompt engineering, a client does not need to

provide labeled data, and less preparation and training
time is required. In prompt tuning, a client will pass in
a set of labeled data to tune the model.
As generative AI is a new field, different people tend to
use different terms. But in truth, these two concepts are
the same.
These are two steps to the tuning process. Clients begin
by providing a small set of data for prompt engineering.
Once it is trained, prompt tuning involves changing
parameters to get the best price-performance by finding
the optimal combinations of model size and other
parameters.
Both prompt engineering and prompt tuning require
modifying the model with labeled data. However, in
prompt tuning, clients need to supply a much larger set of
labeled data (in the order of lOx to lOOx more) but can
get a much better-trained model in the end.
Question 17

Correct

Points out of 1.00

Back Next

A client approaches you and mentions that they know

watsonx.ai is providing different open-source models. They
want to try out these open models to perform a variety of
tasks such as summarizing, generating, and extracting. Which
open models should you suggest the client try out to best suit
these types of tasks?

This client should try the mpt-instruct2 model.

The client should try the gpt-neox model.
This client should try either the flan-ul2 or the mtO-xxl
model.
The client can use any model of the open-source model
from watsonx.ai - all work equally well.

Question 18

Correct

Points out of 1.00

A client has been using Watson Assistant and is interested in

taking the next step with generative AI. They have tried out
some open-source models and are not convinced that prompt
engineering is sufficient because the models would not get
enough information or context for downstream tasks. You
introduced the client to prompt tuning so they can input their
data to help tune the model. What format, initially, is required
to pass data into the watsonx.ai Tuning Studio?

The data can be in PARQUET, AVRO, or CSV format.

The data can be in JSONL or JSON format.
The data can be in PARQUET or JSON format.
The data must be in an ICEBERG table cataloged in
watsonx.data.
Question 19

Correct

Points out of 1.00

Back Next

ChatGPT has put generative AI into mainstream awareness

very quickly, growing to over 100 million users within just a
couple of months. What kind of foundation model is ChatGPT
built on?

Natural Language Utilization Model

General Purpose Trained Model
OpenAI WIKINET Model
Large Language Model

Question 20

Correct

Points out of 1.00

A client has run some queries using ChatGPT and other open-
source foundation models from Hugging Face, but they are
unsure about what a decoder-only model is. What is true
about a decoder-only foundation model?

Decoder-only models are used with chat platforms to

generate new responses, answers, or text based on
simple input from users.
Decoder-only models are translation models. Input
languages to decoder models are encrypted and
tokenized, then decoded to provide a translated output.
Decoder-only models are widely used for non-generative
tasks like classifying customer feedback and extracting
information from long documents.
Decoder-only models are used to deconstruct coding
examples. Users pass in a code snippet and the model will
describe what the code is doing.
Question 21

Correct

Points out of 1.00

Back Next

A client is interested in leveraging the power of generative AI.

However, they have had no prior experience and are a bit
unsure how they can learn to do prompt engineering. You
assure them and state that IBM provides a lot of examples of
prompt engineering for various use cases in the Prompt Lab.
What are some of these examples?

Examples include: image generation, Question &

answering (Q&A), and classification.
Examples include: audio generation, video creation, and
generating email.
Examples include: summarization, classification,
generation, and extraction.
Examples include: generating a graph from data points,
classifying sentiments, and explaining code.

Question 22

Correct

Points out of 1.00

Watsonx.ai gives clients access to popular open-source

foundation models (FMs), as well as a set of IBM FMs built on
highly curated data. Which of the following is true about the
IBM proprietary FM offerings?

Instead of offering decoder models and encoder models,

IBM's FMs are all encoder-decoder. This reduces the
complexity for clients as there is no need to separate the
two.
IBM offers decoder-only models, this includes the Slate,
Granite, and Sandstone IBM FMs.
IBM Granite is a decoder-only model and Sandstone is
an encoder-decoder model.
IBM offers encoder-only and encoder-decoder models.
IBM will not provide a decoder-only model but will
provide access to industry-leading models like GPT.
Question 23

Correct

Points out of 1.00

Back Next

A client uses traditional AI extensively but is unfamiliar with

generative AI. You suggest they can become familiar with how
foundation models can be trained with simple prompt
engineering. Which statement is true about prompt
engineering?

Prompt engineering requires updating and changing the

underlying model to fit specific tasks.
Prompt engineering is the process of providing a labeled
data set (minimum of 100 million, maximum depending
on the vendor) to train the model.
There are three types of prompt engineering: zero-shot,
one-shot, and infinite-shot.
Prompt engineering is a new discipline for finding the '
optimal prompt to use with a foundation model for the
best performance.

Question 24

Correct

Points out of 1.00

Generative AI, large language models, and foundation models

have become popular topics to discuss with clients. Your client
has been looking at different vendors including IBM watsonx,
Amazon Bedrock, and Anthropic. The client is very impressed
with the Anthropic 3Hs. What do the 3Hs stand for?

Hugging Face-based, Helpful, Honest

Helpful, Harmless, Honest
Hugging Face-based, Harmless, Honest
Homogeneous, Harmless, Honest
Question 25

Correct

Points out of 1.00

Back Next

A client is interested in using generative AI in their workflow.

They are aware that there are many open-source models
available in the market and they want to know what models
are available. Which watsonx offering provides a library of both
open-source and IBM proprietary foundation models?

Watsonx.ai includes a library of foundation models.

Watsonx.query provides access to remote models held by
Hugging Face.
Watsonx.governance catalogs and provides a library of
foundation models.
Watsonx.library is the service that provides a library of
foundation models.