Artificial Intelligence Lecture

Introduction

• AI has been around since post-WWII with the advent of digital computers.
• Progress was slow until the 21st century, significantly improving around 2005.
• AI is broad but Machine Learning (ML) notably became practical and useful post-2005, especially around 2012.

Machine Learning (ML)

• Misleading name: doesn't involve a computer self-learning like a human.
• Focus of the lecture: Understanding ML's workings and applications.

Alan Turing and Facial Recognition

• Example: AI used in facial recognition to identify faces such as Alan Turing's.
• Supervised Learning
  • Requires training data: Input-output pairs (pictures labeled with names).
  • AI uses pictures of Alan Turing labeled as 'Alan Turing' to learn.
• Importance of Training Data
  • Social media labelings help train ML algorithms.

Classification Tasks

• Classification task: identifying and labeling data, such as recognizing faces or understanding text.
• Applications in medical imaging (tumors), autonomous driving (Tesla), etc.
• Difference from Generative AI: Classification vs. Generating new outputs.

Neural Networks

• Concept: Artificial neurons mimicking biological neurons in the brain.
  • Human brain: ~86 billion neurons, each connected to up to 8,000 others.
  • Each neuron detects simple patterns and sends signals based on that detection.
• Implementation in AI
  • Based on ideas from the 1940s; practical realization in the 21st century.
  • Requires: Big Data, Deep Learning methods, and significant computer power.
• Training Neural Networks
  • Adjust network to produce desired outputs through extensive training data.
• Takeoff in Technology
  • Advances around 2005; supercharged around 2012 with better computing power and Big Data.

GPT and Large Language Models (LLMs)

• Transformer Architecture
  • Introduced in the paper "Attention is All You Need"; key innovation: attention mechanism.
• GPT-3
  • Released June 2020 by OpenAI; ~175 billion parameters, trained on 500 billion words.
  • First impactful LLM; stark improvement over predecessors.
  • Applications: Prompt completion tasks (autocorrect, summaries, etc.).
• Training Data
  • Includes downloaded text from entire web, scrapes PDF documents, etc.
• Emergent Capabilities
  • Showed capabilities not explicitly trained for, surprising AI researchers.
• Limitations
  • Needs extensive computing power, not feasible for universities.

Issues with AI

• Misinformation
  • LLMs often generate plausible but incorrect information.
• Bias and Toxicity
  • Training data from places like Reddit introduces biases and toxic content.
  • Companies employ guardrails but these are often not deep fixes (e.g., gaffa tape).
• Intellectual Property
  • Absorbs copyrighted material, raising legal issues.
  • Example: Recognizing text from books.
• GDPR Compliance
  • Neural networks can't exclude specific individuals' data as they're not structured that way.
• Outside Training Data
  • Neural networks struggle with scenarios not covered in training data (e.g., Tesla misidentifying stop signs).

General Artificial Intelligence (AGI)

• Types of General Intelligence
  • Type 1: Machine as capable as a human in all facets, including physical tasks.
  • Type 2: Cognitive tasks only, no physical manipulation.
  • Type 3: Any language-based task.
  • Type 4: Augmented LLMs calling specialized subroutines.
• State of the Art
  • NLP is advanced, but other human intelligence aspects (e.g., manual dexterity) are not.

Machine Consciousness

• Debates
  • Google's Lambda case by Blake Le Moine claimed AI sentience, debated widely.
  • Chat-GPT and Co. mimic conversations but don't possess self-awareness.
• Consciousness Understanding
  • Hard problem: Electrochemical brain processes Vs. subjective experiences.
  • AI does not have subjective experiences or personal perspectives.

Conclusion

• Current AI advancements mark a new era in AI research and applications.
• Understanding and limitations critical for future developments.

Final Note: It's essential to use current AI technologies responsibly and understand their functional boundaries.