Generative AI & Large
Language Models
Revision Notes
Source: IBM Skills build —
Foundations in Generative AI, Introduction to LLMs, AI Ethics & Workplace
AI
https://cmis5.anudip.org/auth/ibm-student-nomination
1. Generative AI Fundamentals
●
Foundation Models — Large-scale models trained on massive
(terabyte-scale) unstructured data using self-supervised learning (predicting
patterns without human labels).
●
Large Language Models (LLMs) — A type of foundation model
specializing in generating and analyzing human language.
●
Core Mechanism — LLMs work by predicting the next word in a sequence,
generating responses one word at a time.
●
Tuning —
Adapting a foundation model to a specific task by introducing a small amount of
labeled data (e.g., classification).
●
Prompting — Applying the model to new tasks without additional training,
using instructions or questions (e.g., asking if a sentence is positive or
negative).
●
Advantages — High performance (vast training data) and productivity gains
(requires less labeled data).
●
Disadvantages — High compute costs to train/run; trustworthiness
issues (bias, toxic content, lack of transparency on training datasets).
●
Domains — Language, Vision (images), Code, Chemistry, and Climate.
●
Generative AI Tasks — Idea generation, creating content drafts at scale,
problem solving, summarizing/categorizing data, context understanding, and
prediction.
2. Prompt Engineering Techniques
●
Be Specific
○
Vague
prompts ("Summarize this") yield vague outputs. Add explicit details
about length, focus, and tone.
●
Include Examples (Few-shot Prompting)
○
Providing
examples of correct classifications (e.g., positive vs. negative reviews)
reduces misclassification.
●
Refinement Techniques
○
Add
the format the output should take (e.g., "list of bullets").
○
Add
the role the AI should take and the expertise to use.
○
Add
information describing exactly what you want the AI to do.
●
Prompt Components
○
Task
• Context • Role • Constraints • Output format.
3. AI in the Workplace
●
Automation
○
Handling
repetitive, predictable tasks (e.g., sorting files, generating routine
reports).
●
Augmentation
○
AI
supporting human thinking, decisions, and creativity (e.g., smart assistants
helping prioritize meetings).
●
Human Judgment is the Ultimate Authority
○
AI
provides speed; humans provide direction, context, and accountability. AI
output must always be reviewed and approved by a human — never treated as
final.
●
Case Study — IBM Bob (AI Developer Partner)
○
Bob
builds code foundations, suggests fixes, runs safety checks, documents code,
and refines old code.
○
The
developer always investigates, polishes, and makes the final decisions.
●
Other Roles
○
Project
Managers: AI organizes transcripts; managers prioritize and message
stakeholders.
○
Cybersecurity:
AI prioritizes alerts; analysts evaluate true risk.
○
Data
Analysts: AI translates natural language to queries; analysts validate
assumptions.
○
Educators:
AI generates content; educators shape meaning and build relationships.
4. Ethical Considerations for Generative AI
●
Three Pillars of AI Ethics
○
Transparency:
Users understand how AI works and can trust its outputs.
○
Accountability:
People and organizations take responsibility for AI outcomes.
○
Fairness:
Preventing biases from distorting results.
●
Common Ethical Risks
○
Plagiarism
— AI-generated content closely resembling an artist's/writer's work.
○
Bias
/ Lack of Fairness — training on historical patterns lacking diversity (e.g.,
favoring degrees or past-hire similarity).
○
Data
Privacy Violations — using real, anonymous customer data in training inputs.
○
Harmful
Misinformation — incomplete/outdated data used for decisions such as healthcare
diagnoses.
○
Lack
of Transparency — undisclosed algorithms filtering content, leading to
distrust.
●
Strategies
○
Train
on diverse demographic data, regularly test for equitable outcomes, provide
transparency, and manage bias.
5. IBM Granite Models & Data Analytics
●
Granite Code
○
Best
for generating efficient code, automating routine tasks, and explaining complex
code for onboarding.
●
Granite Instruct
○
Best
for analyzing customer reviews to identify trends and provide summarized
insights.
●
Granite Multilingual / Japanese
○
Specialized
for multi-language and Japanese text tasks (not code).
●
Data Analytics Transformation
○
Gen
AI shifts analytics from reporting on the past to collaborating in real time.
○
Users
query in natural language (e.g., "Which creators grew fastest?").
○
Generates
summaries, charts, code scripts, narratives, and dashboards — empowering
non-coders and data scientists alike.
●
Use Case — Personalization
○
E-commerce
companies use LLMs to recommend products based on individual customer
preferences and purchase history.
6. Quick-Reference Answer Key
|
Scenario
Cue |
Correct
Answer |
|
Personalization
based on purchase history |
Personalization
use case |
|
Generate
efficient code + explain code for onboarding |
Granite
Code |
|
Analyze
reviews for trends & insights |
Granite
Instruct |
|
Model
misclassifies without examples given |
Include
examples |
|
Vague
summary from vague prompt |
Be
specific |
|
Naming/generating
catchy ideas |
Idea
generation |
|
Producing
response word-by-word |
Prediction
step |
|
Determining
intent/tone of a request |
Contextual
understanding |
|
100
product descriptions quickly |
Creating
content drafts at scale |
|
Summarizing/categorizing
thousands of entries |
Problem
solving |
|
Wants
output as bullet list |
Add
output format |
|
Wants
more audience-relevant detail |
Add
role & expertise |
|
Response
too broad |
Add
exact task detail |
|
Diverse
training + bias testing |
Fairness |
|
Tool
favors certain universities despite balanced data |
Managing
bias |
|
Favors
candidates similar to past (non-diverse) hires |
Lack
of fairness |
|
Undisclosed
content-filtering algorithm |
Providing
transparency |
|
Chatbot
leaks another customer's data |
Data
privacy violation |
— End of Notes —
No comments:
Post a Comment