Artificial Intelligence Notes
What is AI?
- Rule-based expert systems: Systems that use defined rules to make decisions.
- Machine-learning-based systems: Systems that learn from data to improve over time.
- Neuro-symbolic methods: Combines expert knowledge and flexibility of machine learning.
- Currently, AI is often associated predominantly with Machine Learning.
Brief History of AI
- Alan Turing: Pioneer in AI thought, proposed the Turing Test.
- Dartmouth Workshop: The formative conference where AI was first coined.
- Perceptron: Early neural network model developed in the 1950s.
- Expert Systems: Software that mimics human decision-making capabilities.
- Simple statistical ML: Early machine learning approaches that used basic statistics.
- Deep Learning: Advanced machine learning utilizing neural networks to process data at high levels of abstraction.
Modern AI vs. Traditional AI
- Expert System Features:
- Defined inputs (X) and functions (f).
- Operation similar to a vending machine.
- Modern ML/AI Features:
- Utilizes vast data sets on inputs and outputs (X and Y).
- Transfer of knowledge to a model that learns the function (f) rather than having it explicitly defined.
- Functions become a complex, often opaque process (black box).
Features of Modern AI
- Deep Learning Models: Primary driving force behind modern AI.
- Functions as a large, complex black box.
- Versatility of Use: Applications span numerous fields, from healthcare to finance.
- Accuracy: Generally achieves high prediction accuracy but often lacks interpretability.
- Example Issues: Classifying images as dogs or cookies can be challenging due to the black-box nature of deep learning.
Machine Learning Applications
Supervised Learning:
- Input Examples:
- Picture (Are there human faces? 0 or 1)
- Loan application (Will they repay the loan? 0 or 1)
- Ad with user data (Will the user click? 0 or 1)
- Output Examples:
- Speech recognition (transcription)
- Language translation (English to French)
- Preventive maintenance (predictive analysis for machinery)
- Self-driving cars (navigation and obstacle avoidance).
Generative AI
- Definitions: Models that can generate content (text, images, video, or sound) based on input prompts.
- Examples:
- Text: ChatGPT (OpenAI), Gemini (Google), LLaMA (Meta)
- Images: Dall-E (OpenAI), Mid Journey, Stable Diffusion
- Videos: Sora (OpenAI), Runway AI
- Voices: ElevenLabs.
Training LLMs (Large Language Models)
- Training Process: Models are trained on large text datasets (books, articles, websites).
- Tokens: Text is broken down into numeric representations (example: “Apple” may be represented as 10001).
- Inference: Models predict the next token in a sequence based on learned patterns, leading to more accurate and context-aware text generation.
Limitations of Tokenization
- LLMs function based on tokens rather than individual characters.
- Some tokens, especially those from programming, can lead to inaccuracies if the model hasn't been specifically trained on them.
Prompt Engineering
- Purpose: The art of crafting prompts to elicit improved responses from LLMs.
- Techniques:
- Chain of Thoughts (CoT): Directing step-by-step thought processes within prompts.
- Few-shot Learning: Providing examples in prompts for context and guidance.
Image Generation Models
- Proprietary Models: Dall-E 3, Mid Journey, which generally perform well out of the box.
- Open Source Models: Stable Diffusion, which can be customized and adapted for particular uses.
- Example: Combining Stable Diffusion with control networks for generating specific images.
Video Generation
- Challenges: Video generation is computationally intense with issues of consistency and control over output.
- Example Tool: OpenAI's Sora generates high-fidelity videos from text prompts.
Voice Generation/Cloning
- Resources: Websites like ElevenLabs allow text input for voice synthesis, including cloning features.
Collaboration with AI
How LLMs May Augment Human Capabilities
- Modes of Collaboration:
- Ghostwriting: AI generates full text for review or publication.
- Soundboard: AI critiques, aids brainstorming, or enhances language style.
- Experiment Insights: Using LLMs as ghostwriters may diminish creative potential due to anchoring effects; however, using them as soundboards increases performance.
Risks and Pitfalls of AI
- Data Reliance: Exclusively relying on data can lead to significant blind spots and biases.
- Hallucinations: LLMs can generate plausible but false content due to lack of understanding or meta-knowledge.
Responsible AI Use
- Important to evaluate tasks suitable for AI and the collaboration modalities between users and AI systems.
- Understanding and mitigating the societal implications and biases associated with AI, particularly in decision-making processes, is critical.
Bias in Image Generation Models
- Bias arises from training data leading to misrepresentation and inaccuracies.
- Strategies for addressing bias include data-debiasing methods or using explicit bias mitigation prompts during generation.
Conclusion: Societal Implications of AI
- Generative AI's societal impacts include ethical considerations, misinformation, financial loss, and the dynamics of human-like interactions with AI, necessitating vigilant and responsible use.