Generative AI On Amazon Web Services Ebook
Generative AI On Amazon Web Services Ebook
potential of
generative AI with
Amazon Web
Services
Introducing generative AI 3
Why Fractal? 31
Resources 32
Introducing
generative AI
Introducing generative AI
4
History of generative AI
AI Artificial Intelligence
Artificial Intelligence
The field of computer science that seeks to
1956 create intelligent machines that can replicate
or exceed human intelligence
Machine Learning
Machine Learning
Subset of AI that enables machines to learn
1997 from existing data and improve upon that
data to make decisions or predictions
Deep Learning
Deep Learning
A machine learning technique in which layers
2017 of neural networks are used to process data
and make decisions
Generative AI
Generative AI
Create new written, visual, and auditory
2021
content given prompts or existing data
5
Key benefits of generative AI
6
Key benefits of generative AI (Continued)
7
Generative AI applications
Generative AI has many
applications across different
domains, such as marketing, data Content marketing: Create personalized and engaging content for your
analytics, software engineering, website and customers, such as web copies, blog posts, customer
sales, and cross-functions. stories, and social media posts. Generative AI can also generate catchy
slogans, headlines, and product descriptions.
Through its ability to create new
Data analytics: Provide insights and predictions based on business data
content from various input types that can help companies make informed decisions, identify patterns and
such as text, images, voice, and trends, and optimize business operations. Generative AI can also help
computer code, generative AI has generate data visualizations and reports for your data analysis.
applications across most business
functions and processes. Software engineering: Simplify and automate various software
development tasks such as generating code from natural language
descriptions or pseudocode, completing code snippets, fixing bugs,
suggesting improvements, and explaining code logic.
8
The different types of generative AI models
1 Language models
Large Language Models (LLMs) process and understand natural language in text (or through speech recognition, voice)
format. These models are trained on massive amounts of text. The most popular type of generative AI models
currently, LLMs can not only be used for creative, academic, and business writing, but also for translation, grammatical
correction or analysis, computer code writing, and more. The most well-known LLM is ChatGPT from OpenAI.
3 Voice models
These models can generate natural sounding speech or music based on input such as a text, a melody, or a voice.
Voice models can be used for composing, songwriting, dubbing, speech recognition, sound editing, and more. For
example, Jukebox is a voice model that can generate music, including rudimentary singing, as raw audio in a variety of
genres and artist styles. Jukebox can also generate lyrics conditioned on a genre, an artist, or a user prompt.
9
Key Large
Language Model
(LLM) generative
AI providers
Key LLM platforms and models providers: summary table
• Anthropic APIs
Anthropic • In-house model: Claude 2, Claude Instant • Bedrock APIs
• Claude chat webapp
Meta • In-house model with Llama 2 (various sizes) • Stand-alone open-source models
11 Note: This table represents an August 2023 technology snapshot of a rapidly evolving ecosystem
Avoiding the halo effect
In December 2022, ChatGPT shook the world. Initially, it was mostly the
technology world that understood its potential impact, especially for
developers. Then, in early 2023, ChatGPT spread quickly to education and to
many knowledge worker roles.
It also showed its limitations when a lawyer went to court with a ChatGPT-
generated argument that was full of so-called “hallucinations” (i.e., made-up
content not rooted in reality).
Soon after ChatGPT came out, the race for better models picked up with LLMs
from Anthropic (Claude), Meta (Llama), Google (PaLM model through the Bard
service), and OpenAI’s own GPT-4.
The cost associated with training and operating those larger and larger
models, however, is quickly becoming prohibitive.
So much so that even OpenAI’s CEO, Sam Altman, was quoted in April 2023
saying that “I think we’re at the end of the era where it’s going to be these giant
models, and we’ll make them better in other ways.”
12
Avoiding the halo effect (Continued)
As the battle was being fought over which one of those giant models was
better, Facebook open-sourced its Llama model.
This changed the dynamic again, and the focus switched from creating ever-
larger generic models to creating smaller or customized ones.
For instance, even if text translation generated from LLMs was marginally
better than from “traditional” AI machine translation models, customers
quickly realized that the cost differential didn’t justify switching to LLMs for
their large-scale machine translation processes.
13
Avoiding the halo effect (Continued)
As always, for any new ground-breaking technology, it’s easy to fall into the
“hammer in search of a nail” trap. Generative AI is and will remain
transformational across most business processes for many years to come, no
doubt about it.
LLMs’ capabilities will continue to improve in the coming years. However, once
the initial momentum of rapid successive quality improvements recedes (as it
always does with technology) and that progress becomes more gradual, there
will be two likely outcomes:
• One (or a few) players will dominate the field with a quality level way
above all the rest of the LLM pack.
• All the main players will converge within a small distance of each-others
quality-wise.
Whatever the end game might be, it’s too early to bet your company’s
generative AI multi-year strategy on one specific LLM.
So, what is the most logical strategy to move forward today given those
uncertainties?
14
Future-proof your Generative AI strategy
Knowing which LLM will win, if any, is not a bet you want to take. So, what’s the best approach to strategically position your
company to take this generative AI tectonic shift seriously while still not getting bogged down by (ever-changing) LLM-level
details? To future-proof your generative AI strategy, it’s crucial to avoid locking yourself in a specific LLM, or to assume a single
approach could be used for all your business needs.
First, you need to clearly define the most appropriate LLM per business process. Should you go with the one that offers the
largest context window to be able to consume long text, or should it be the one that has the best mathematical or data
analytics capabilities? Or maybe it’s the model’s coding, debugging, and software language translation that’s the most
important?
Also, you need to ensure that the model can be customized with your data to better align with your company’s unique
situation. Depending on the business process that you will redesign and augment with generative AI, the best LLM to select
can change dramatically. You need to keep the flexibility of which LLM to use now and in the future.
Regardless of the LLM selected, you also want to ensure you can swap it out seamlessly if better options become available.
Therefore, your generative AI solution must be architected in a way that insulates the application from the LLM, like you would
ensure that any business application would not be locked to a specific database.
Your overall generative AI platform, such as Amazon Web Services’ Bedrock platform, must allow for selecting the right model
(LLM, visual, code, voice, etc.) for the right use case. It should enable not only customization capabilities but, at least as
importantly, guarantee your company’s data security and privacy.
15
Building a multi-
year generative
AI strategy
Remember Amara's Law We tend to overestimate the effect
of a technology in the short run
and underestimate the effect in
the long run.
– Roy Amara
Impact of a new
technology
Fractal has identified four key stages - Crawl, Walk, Run, and Fly - that organizations should consider following for
their generative AI journey. Each stage requires investments in technical and human capabilities to maximize the
potential of generative AI and ensure internal adoption while guaranteeing enterprise data privacy and information
security.
Crawl (Test)
In the “crawl” stage, organizations experiment with easy-to-deploy use cases and then quickly measure the
ROI of those use cases. They also identify potential at-scale use cases that show promise for further
exploration and development.
Walk (Deploy)
In the “walk” stage, organizations deploy a few (one to three usually) use cases at scale. They again measure
their ROI and use the results to expand generative AI to more business processes. This unlocks further
potential for model customization, optimization, and scaling up.
Run (Scale)
In the “run” stage, organizations analyze most or all their existing processes to build a multi-year AI
transformation roadmap that comprehensively integrates generative AI across the enterprise.
19
End-to-end generative AI
Walk journey: “Walk” stage
Custom models developments: Deep AWS data and ML/AI
stack (including Bedrock) knowledge In the “walk” stage, organizations progress further by
moving from experimental projects to the practical at-scale
Cloud and data engineering deployment of generative AI-powered solutions.
20
End-to-end generative AI
Run journey: “Run” stage
Center of Excellence
During the “run” stage of their generative AI journey,
organizations focus on enterprise-wide scale-up (size) and
Business cases prioritization scale-out (number of projects).
Model productization and integration: To support this growth, DevOps and MLOps practices are
Software engineering, application integration implemented to ensure generative AI model integration,
monitoring, and management of AI models in their overall IT
Enterprise-wide scale up (size) and scale out (number of
and Data Science infrastructure.
projects): DevOps, MLOps
Model productization and integration become key priorities,
Custom models developments: Deep AWS data and ML/AI requiring strong software engineering capabilities and the
stack (including Bedrock) knowledge utilization of platforms such as Power Platform to integrate
those models across existing enterprise business
Cloud and data engineering processes.
Long term industry experience & solutions With the support of third-party partners, such as Fractal,
organizations can build a Center of Excellence (CoE) that can
help centralize expertise, increase knowledge sharing, and
enforce governance.
21
End-to-end generative AI
Fly
journey: “Fly” stage
AI AI-native design
In the “fly” stage, organizations make a shift towards
becoming an "AI native" organization. In this stage,
Center of Excellence organizations can design new processes or redesign existing
ones using AI as a core building block.
Business cases prioritization
Instead of adding AI to existing processes, they leverage AI
from the ground up to maximize results.
Model productization and integration:
Software engineering, application integration This approach parallels the shift seen during the early days
Enterprise-wide scale up (size) and scale out (number of of digital transformation, where companies added digital
projects): DevOps, MLOps elements to existing processes instead of designing them as
digitally native ones from the ground up.
Custom models developments: Deep AWS data and ML/AI
stack (including Bedrock) knowledge Similarly, AI-native companies build processes by
incorporating AI from the ground up.
Cloud and data engineering
The fly stage represents a transformative phase where
generative AI becomes a driving force, enabling
SenseForth Fractal GPT customization and optimization organizations to unlock new business opportunities and
create unique customer experiences.
22
Barriers and risks
associated with
generative AI
Barriers to generative AI Adoption
Data quality & availability Lack of business alignment Regulatory & ethical challenges
Limited access to reliable and relevant Organizations use generative AI for Concerns around privacy, security,
internal and external data makes it multiple purposes, like innovating a bias, and ethical implications create
difficult to effectively use generative product/service or using it to improve regulatory barriers and raise
AI algorithms for accurate results and existing processes. questions about the responsible use
insights. However, to harness its potential, of generative AI.
organizations must have a clear Also, a lack of awareness leads
vision, strategy, and roadmap that to misconceptions and fears about AI
align with their goals, needs, and taking over human jobs rather than
values. augmenting their ability to improve
their productivity and efficacy.
24
Barriers to generative AI Adoption (Continued)
Culture & change management Technical complexity & expertise Lack of infrastructure & resources
Resistance to change, organizational The complexity of implementing and Generative AI models require
culture, and lack of awareness about managing generative AI systems can substantial storage capacity, the right
the benefits of generative AI can pose challenges in effectively data platform, network bandwidth,
impede adoption and hinder the leveraging the technology if the and a skilled workforce to run
necessary cultural shift for its effective organization lacks the necessary new efficiently and reliably.
use. and specialized expertise.
25
Risks and their mitigation strategies
Hallucination or Customization
Cost overrun
errors, bias needs
26
Risks and their mitigation strategies (Continued)
Generative AI models may Generative AI models can Generative AI carries the risk of
require access to sensitive or generate content that may being detected by AI systems or
personal data, which can be infringe on the intellectual experts, leading to the potential
compromised or misused by property or creator rights. identification of fake or
third parties. harmful content.
Defining ownership and usage
Implementing strong data rights through legal agreements, To mitigate this, organizations
protection measures, complying employing authentication must create systems and
with privacy regulations, and mechanisms, monitoring implement processes to ensure
considering alternative systems, and using digital rights generative AI tools are creating
approaches like federated management (DRM) techniques ethically appropriate content
learning or differential privacy can help protect both IP and including deceptive or illegal
can help mitigate those privacy creator rights. content. The must also comply
risks. with applicable legal
and regulatory frameworks
governing the use of generative
AI.
27
Getting started with generative AI
Fractal offers different ways to help you quickly get
started on your generative AI journey.
28
Start creating with FractalGPT
FractalGPT
29
Generative AI in 30 days
The generative AI in 30 days offer is designed to help organizations analyze generative AI use cases and rapidly prepare for
the deployment of generative AIs. The offer achieves this by leveraging predefined resource deployment procedures and
templates from AWS and Fractal.
Offer outcomes
Phase 1 Phase 2 Phase 3
Landing zone Generative AI proof • Identify generative AI use
Design workshop deployment of concept cases in discovery workshop
• Identify generative AI • Establish AWS landing • Transfer knowledge on • Deployed POC of customer-
scenarios for proof of zone and generative AI data platform, generative defined generative AI use
concept capabilities AI capabilities & best case via customer’s AWS
practices subscription
• Gain an understanding of • Working session to
the business use case customize landing zone • Develop approach to tailor • Documentation and
• Review the current state of • Population of landing LLM models knowledge transfer
data and advise on data zone, showcasing • Provide customer with a • Experienced AI experts to
preparation methodology implementation best go-forward plan aligned guide your team along the
• Create initial architecture practices with resourcing and scenario refinement &
• Use migration accelerators architectural needs for the feature definition path
next phase
• Establish initial • Proposal for implementation
architecture to get data of production solution
flowing
31
Why Fractal?
As a Data and Analytics AWS Partner with decades of experience developing and
deploying AI solutions at scale, Fractal can help you analyze, design, and deploy your
generative AI-based solution quickly and effectively.
32
Resources
• AWS Bedrock
• Anthropic Claude
• Azure Open AI
• Google Vertex AI
33
fractal.ai
One World Trade Center Suite 76J, New York, NY 10007 | +1 (646) 547 1600