0% found this document useful (0 votes)
9 views28 pages

Transitioning to Creative AI Technologies

Generative AI represents a shift from analytic to creative capabilities, enabling the generation of diverse content types such as text, images, and audio. Key advancements include the transformer architecture, which enhances data processing and scalability, and the use of reinforcement learning for output alignment. The future of AI includes autonomous agents capable of performing tasks independently, highlighting the ongoing evolution in AI technology.

Uploaded by

ffreakscience
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
0% found this document useful (0 votes)
9 views28 pages

Transitioning to Creative AI Technologies

Generative AI represents a shift from analytic to creative capabilities, enabling the generation of diverse content types such as text, images, and audio. Key advancements include the transformer architecture, which enhances data processing and scalability, and the use of reinforcement learning for output alignment. The future of AI includes autonomous agents capable of performing tasks independently, highlighting the ongoing evolution in AI technology.

Uploaded by

ffreakscience
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

What is Generative AI?

➜ Shift in Paradigm: Moving from Analytic AI (predicting


labels) to Creative AI (generating content).

➜ Content Types: Capable of producing high-fidelity text,


high-resolution images, code, and audio.

➜ Zero-Shot Learning: Ability to perform tasks without


specific prior training examples or fine-tuning.
Training at Scale

Pre-training: Reading the open internet to learn


grammar, facts, and underlying world logic.

SFT Fine-Tuning): Human experts teaching the


model how to follow specific instructions and
formats.
RLHF Reinforcement Learning from Human
Feedback to align outputs with helpfulness and
safety values.
The Transformer Revolution (2017)

"Attention is All You Need": The landmark Google


paper that introduced a new way to process data.

Parallelization: Processes entire sentences at once


rather than word-by-word RNN style).

Scalability: The breakthrough that allowed models to


grow from millions to trillions of parameters.
The "Attention" Mechanism

Contextual Spotlight: AI assigns "weights" to words in a


sequence to determine their mutual relevance.

Disambiguation: Understanding that "Bank" in "River Bank" is


different from "Bank Loan" based on surrounding context.

Global Context: Relates words at the start of a 500-page book to


words at the very end seamlessly.
Tokens & Embeddings

Numerical Language: Words are chopped into "tokens"


(sub-word units) and converted into numbers.

Vector Space: Concepts are mapped into a high-dimensional


mathematical space based on their semantic meaning.

Semantic Proximity: Concepts like "King" and "Queen" are


mathematically closer than "King" and "Apple".
Context Windows

Short-Term Memory
The total volume of information the AI can "hold" and reference

2.0M
within a single conversation cycle.

Range: 4,000 GPT3 legacy) to 2,000,000 Gemini 1.5 Pro).

Impact: Determines if the AI remembers a paragraph or a


1,000-page encyclopedia.

MAX TOKEN LIMIT


ChatGPT (OpenAI) Evolution

2020 2022 2023 2024+


GPT3 Launch ChatGPT 3.5 GPT4 Power GPT4o Multimodal

Strategy: Rapid iteration and public deployment for real-world stress testing.

GPT Defined: "Generative Pre-trained Transformer".


Multimodal Reasoning

Vision Voice Reasoning


Interpreting complex charts, Real-time conversations with Unified model that sees, hears, and
handwriting, and live feeds. emotional inflection. speaks simultaneously.
The "o1" Thinking Models

Chain of Thought
AI that "pauses" to think through complex logic before
answering.

Reasoning Power: Exceptional performance in PhD-level


science and competitive math.

Self-Correction: Ability to find and fix its own errors during the
thought process.
The ChatGPT Ecosystem

Custom GPTs DALL-E Data Analysis


Specialized versions for niche tasks. Integrated high-quality image Spreadsheet processing and
generation. visualization.
Claude AI
The Ethical Intellectual by Anthropic
The Context King

High Retrieval: Renowned for the "needle in a haystack"


test—finding a single fact within 200,000 tokens.

Synthesis: Perfect for legal contracts or 1,000-page


research papers.

Consistency: Doesn't "forget" instructions given 50 pages


ago in the same chat session.
Coding & Writing Nuance

Developer Choice: Often rated as the most reliable for complex


software engineering logic.

Natural Tone: Writing that sounds less "robotic" and more like a
sophisticated human colleague.

Artifacts Feature: Dedicated side-window for viewing and


editing code/docs in real-time.
Anthropic's Mission
Public Benefit Corporation: Focused on Safety and Interpretability.
Built-in safeguards to avoid harmful stereotypes, ensuring outputs are Helpful, Honest, and
Harmless.
Google Gemini: Natively Multimodal

Built Different: Trained on all media types (text, code, image,


video) simultaneously from Day 1.

Video Understanding: Can analyze an entire 1-hour video to


summarize key timestamps or find specific objects.

Heritage: Powered by Google DeepMind—the team behind


AlphaGo and AlphaFold.
The 2,000,000 Token Window

2M
TOKEN CAPACITY

Infinite Memory: Analyze 30,000 lines of code or entire libraries in one prompt.
Google Ecosystem Integration

Workspace Live Tools Android


Directly reads Gmail, Google Drive, Integrated with Maps, Flights, and The successor to Google Assistant
and Docs. YouTube. on billions of devices.
Infrastructure Power

AlphaGo Heritage: Using "Tree-Search" algorithms to solve


logic and scientific puzzles.

Global Scale: Leveraging Googleʼs TPU Tensor Processing


Unit) chips for massive training.

Discovery: Technology being used to map proteins and


predict global weather patterns.
The Head-to-Head Comparison

Model Best For... Key Feature Safety Focus

ChatGPT Creativity & Logic Voice & Ecosystem High RLHF

Claude Coding & Writing 200k Context Window Constitutional AI

Gemini Video & Search 2M Context Window Ecosystem Integrated


Challenges & Ethics

Hallucinations IP & Bias Sustainability


Why AI sometimes lies with total Debate over training data and fair Massive energy and water needs
confidence. representation. for data centers.
The Future: AI Agents

From Talking to Doing: AI that doesn't just suggest a flight, but


books it for you.

Autonomous Workflows: Systems that can run marketing


campaigns or code apps independently.

Personalized Models: AI that learns your specific preferences


and habits over time.
Questions?
"The best way to predict the future is to build it."

Thank you for your attention!


Image Sources

[Link]
[Link]
Source: [Link]

[Link]

Source: [Link]

[Link]

Source: [Link]

[Link]

Source: [Link]

[Link]

Source: [Link]

[Link]

Source: [Link]
Image Sources

[Link]

Source: [Link]

[Link]
[Link]
Source: [Link]

[Link]

Source: [Link]

[Link]

Source: [Link]

[Link]

Source: [Link]

[Link]

Source: [Link]
Image Sources

[Link]

Source: [Link]

[Link]

Source: [Link]

[Link]
f6
Source: [Link]

[Link]
wZw
Source: [Link]

[Link]

Source: [Link]

[Link]

Source: [Link]
Image Sources

[Link]

Source: [Link]

[Link]

Source: [Link]

You might also like