
This presentation was created using AI as assistant.
All content is carefully checked and arranged by hand
- Brainstorming: OpenAI ChatGPT4
- Research: OpenAI ChatGPT4 + Bing, Bing, perplexty.ai
-
Generate Images: DallE3 in OpenAI ChatGPT4 and Bing/create
do you recognize all generated images?
AI Disclaimer
What is AI?


lost Group of wanderers
no sight, no directons, no map
Target: home alley


LLMs
Computer uses
natural language


Computer solves complex tasks
What is AI?



Behavior is trained on data
What is AI?




Search for the smallest Error
What is AI?





Neuronal Net
What is AI?




Neuronal Net
complicated net of simple computation
What is AI?





AI seeks configuration
with smallest error
What is AI?






find valley in
multidimensional space
What is AI?
What is AI?


lost Group of wanderers
no sight, no directons, no map
Target: home alley
Generative Pre-trained Transformer
Generates new text
by transforming input text
exactly as trained previously
What are LLMs?
What are LLMs?

Artikel: medium.com

Complete next word in sentence

What are LLMs?



Completes sentence word by word
Artikel: medium.com


What are LLMs?


Transforms Request into Answer
Artikel: medium.com


What are LLMs?


Transforms Request into Answer
Artikel: medium.com


What are LLMs?
Mechanical Parrot
Does not repeat
but give expected answer




What are LLMs?
LLMs trained on
- high quality science content


What are LLMs?
LLMs trained on
- high quality science content
- Twitter of Trump



Range of Quality
perfect "Bullshitter"
Wants to sound smart
doesn't care for thruth
What are LLMs?


Why is GPT so good?

Architecture
- Vast Training Data
570 GB Text ~ 800.000 Books
- Transformer Architecture
handles data sequences
Attention: learns relation of words
Why is GPT4 better than GPT3?

Architecture
- 10x More Training Data
- 10x Bigger Architecture
- >10x Bigger Context Window
- Better Mixture of Experts (MoE)
Mixture of Experts
Mixture of Experts

GPT:
a single Assistant?

Mixture of Experts

GPT:
a Team of Assistants


Factcheck
Ethics
Code of Conduct
Mixture of Experts

GPT:
a Team of Assistants


Factcheck
Ethics
Code of Conduct
MoE improved constanly
Evolution of LLMs

Evolution of LLMs

LLM Moore's Law: Every year model size increases by 10x
Cost of LLMs
- 8 GWh (700 years US household)
- 3 Month
- 300 t CO2 (1 Mio km Car)
Cost of LLMs
Sam Altman: 40x Cost reduction for ChatGPT in 1 Year
(Unconfuse Me with Bill Gates)
- 8 GWh (700 years US household)
- 3 Month
- 300 t CO2 (1 Mio km Car)
- 0.01 kWh (15xGoogle, 100ml Tea)
- few seconds
- 0.4 g CO2 (10 Breath)
= 4 t CO2 / 10 Mio requests (daily)
My use of ChatGPT
GPT Power User since April '23
- Brainstorm Ideas
- prepare talks & lectures
- ehance study material
- coding
- search for reference, incl academic
- research in science
- produce videos
- ...

My use of ChatGPT
GPT is the Assistant I've been waiting for
GPT Power User since April '23
- Brainstorm Ideas
- prepare talks & lectures
- ehance study material
- coding
- search for reference, incl academic
- research in science
- produce videos
- ...

GPT is the Assistant I've been waiting for


Artikel: DevOps.com

AI enhances
- productivity
- satisfaction
- efficiency
simplify and improve our work lives
GPT is the Assistant I've been waiting for


Artikel: DevOps.com
AI enhances
- productivity
- satisfaction
- efficiency
- not success
- but speed

Use of LLMs






costumer service
Summarizing
write code
Assisting



Use of LLMs






costumer service
Summarizing
write code
Assisting





















- Jack Ma, Satya Nadella, Elon Musk, Andrew Ng, Ginni Rometty, Marc Benioff, Kai-Fu Lee, ...


- Jack Ma, Satya Nadella, Elon Musk, Andrew Ng, Ginni Rometty, Marc Benioff, Kai-Fu Lee, ...



- Jack Ma, Satya Nadella, Elon Musk, Andrew Ng, Ginni Rometty, Marc Benioff, Kai-Fu Lee, ...




I made this with DallE-3

LLMs - 24/7 software engineer


LLMs - 24/7 software engineer

LLMs - 24/7 software engineer




LLMs - 24/7 software engineer





LLMs - 24/7 software engineer






LLMs - 24/7 software engineer


LLMs - 24/7 software engineer


LLMs - 24/7 software engineer




LLMs - 24/7 software engineer

Churned code
3.3 %
7.1 %


LLMs - 24/7 software engineer




LLMs - 24/7 software engineer





LLMs - 24/7 software engineer


Future vision:
- Programming language designed by AI
- co-work pseudo code with LLM
- Required skills:
abstraction, problem solving,
logics, adaptability,
system & algorithm design

LLMs - 24/7 software engineer


Future vision:
- Programming language designed by AI
- co-work pseudo code with LLM
- Required skills:
abstraction, problem solving,
logics, adaptability,
system & algorithm design

AI has no aim or purpose

How to Code with AI





How to Code with AI



GPT3.5 Output:
- 59.7% plagiarized
- 45.7% copied
- 46.5% paraphrased
- 27.4% Minor Changes
Basically: No GPT Output is original



GPT3.5 Output:
- 59.7% plagiarized
- 45.7% copied
- 46.5% paraphrased
- 27.4% Minor Changes
Basically: No GPT Output is original



LLMs can do serious comedy!?!
Internet dies of AI

Internet dies of AI



AI dies of bad (Internet) Data




AI dies of bad (Internet) Data






AI dies of bad (Internet) Data




!!! Explicitly label AI generated text !!!




























Improvement of AI
Improvement of AI
GenAI brings Images to live
GenAI brings Images to live

Improvement of AI


Improvement of AI




Future of Mankind
Future of Mankind


Future of Mankind



Future of Mankind



Future of Mankind


Future of Mankind


Achievements of AI


Achievements of AI



Achievements of AI




Achievements of AI





Achievements of AI




Achievements of AI





Achievements of AI






Achievements of AI







Achievements of AI
Pivotal Role of LLMs?
Achievements of AI
Pivotal Role of LLMs?


Achievements of AI


Pivotal Role of LLMs?

Achievements of AI


Pivotal Role of LLMs?


Future of AI
Future of AI
- talk & reason with machines
- explanation for failure
- clarify questions
- explain new task
- personalize

Future of Work
- Creativity
- Innovation
- Leadership & Mentorship
- Collective Intelligence
- Connecting Dots

Responsible Use
- collaborate with AI
- Iterate
- cite when & for what you use AI
- label AI generated content
- don't overuse AI
Conclusion
- AI - revolution in the making
- LLMs - Universal Interface man - machine
- huge increase in productivity
- already transforming entire industries
- humans lead the way of AI
- we are just getting started...

Selbstlernkurs
Didaktik mit KI

Grundwissen zu KI
Offene Diskussion
KI in der Lehre
Jeden 1. Montag im Monat
12:00 - 13:00
Mathematikzentrum
LLMs - Revolution or Hype?
By Stefan Hackstein
LLMs - Revolution or Hype?
- 1,675