Scribd
Upload
16 views

1 - Intro To FMs

The document provides an introduction to foundation models and generative AI. It discusses what foundation models are, how they work, their capabilities and limitations, and the lifecycle of developing and improving foundation models including techniques like pretraining, fine-tuning, and using human feedback.

Uploaded by

ALEX ROMBUCI
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
16 views

1 - Intro To FMs

The document provides an introduction to foundation models and generative AI. It discusses what foundation models are, how they work, their capabilities and limitations, and the lifecycle of developing and improving foundation models including techniques like pretraining, fine-tuning, and using human feedback.

Uploaded by

ALEX ROMBUCI
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 14

Introduction to

Foundation Models

Generative AI Foundations on AWS


Emily Webber, Principal ML Specialist SA at AWS
Lesson 1 – Level 300
Topics in our journey together
1. Primer on
1 4. Pretraining
foundation models

7. Distributed
hosting
2. How to pick a
2 4 7
foundation model

6 6. Reinforcement
5 learning with
3. Prompt
3 human feedback
engineering & fine-
tuning

5. Prepare data and


train at scale
Today’s activities
• What foundation models are
• Where they come from
• How they impact Generative AI
• Foundation model lifecycle
• Hands-on walk through: Foundation
models on Amazon SageMaker
Let’s say I asked you to learn everything on the internet.
How would you do it?

Structure Storage Time

5.74 B pages x 52 seconds


= ~83,000 hours A foundation model can
do this in a few months.
=> ~41,000 human years
You can do a lot with foundation models!

Text Q&A Text Text Paraphrase Search


generation summarization extraction rephrase

Code Image Image Audio Video


generation generation classification generation generation
Many ML tasks can now be re-cast as generative,
and most will benefit from foundation models.
Text: I am not into this house; it’s way too Text: I am not into this house; it’s way too expensive and
expensive and too far from the train line! too far from the train line!

Classify this sentence into positive or negative sentiment:

Sentiment:
Model negative Agent: Negative sentiment
Model

Traditional classification Using generation to classify text


There are many ways to customize a foundation model

Pretraining

Fine-tuning *
Accuracy

Retrieval
Prompt augmented
engineering generation

Complexity and cost


The best generative models are built on human feedback

1 2 3

Pick any Send in your


generative model prompts Send the responses
to your users

6 5 4

Use the reward model Update your


Train a
to improve your original training datasets
reward model
generative model
Foundation model spotlight: Stable Diffusion
Prompt: landscape of the beautiful city of paris rebuilt near
the pacific ocean in sunny california, amazing weather, sandy
beach, palm trees, splendid haussmann architecture, digital
painting, highly detailed, intricate, without duplication,
concept art, matte painting, trending on artstation

Negative prompts: trees, green

Stable Diffusion-generated image


Timeline of language foundation models
GPT = Generative Pretrained Transformers

PaLM
BLOOM
GPT-1 OPT
BERT GPT-3 ChatGPT

Transformer GPT-2 InstructGPT LlaMa ++


Megatron-LM Cohere GPT - 4
AI21 Anthropic
GPT - J Titan
Stable Chat

2017 2018 2019 2020 2021 2022 2023


Foundation model spotlight: AI21 Jurassic-2 Jumbo Instruct
Foundation model lifecycle

1 2 3

Pick a base Use prompt


foundation model engineering Evaluate performance
with your users

6 5 4

Update your original Fine-tune your model Store human responses


foundation model and preferences
https://bit.ly/sm-nb-1

Hands-on demo
Thank you!
Type: Corrections, feedback, or other questions?
Contact us at https://support.awsamazon.com/#/contacts/aws-academy.
All trademarks are the property of their owners.

You might also like