Generative AI refers to a class of artificial intelligence systems capable of creating novel content, such as text, images, audio, video, and code, based on patterns learned from training data. Unlike discriminative AI models that classify or predict based on input, generative models produce new data instances.
Building upon Machine Learning, GenAI is one facet of Artificial Intelligence that allows generating content. This includes Large Language Models, Generated Art, and GenAI Voice Cloning.
Key Technologies
- Large Language Models (LLMs): Models like GPT, PaLM, and Llama trained on vast text datasets to generate human-like text, translate languages, write different kinds of creative content, and answer questions.
- Diffusion Models: Used primarily for image generation (e.g., Stable Diffusion, DALL-E 2/3, Midjourney), these models learn to reverse a process of adding noise to images to generate new ones from text prompts or other inputs.
- Generative Adversarial Networks (GANs): Consist of a generator and a discriminator network competing against each other to produce realistic synthetic data. Widely used for image generation before diffusion models became prominent.
- Transformers: The underlying architecture for many state-of-the-art LLMs and some vision models.
Applications
- Content Creation: Generating articles, marketing copy, scripts, music, artwork, and virtual environments.
- AI Co-pilots: Assisting humans in tasks like coding, writing, and design.
- Synthetic Data Generation: Creating artificial data for training other AI models, especially when real data is scarce or private.
- AI Companionship: Powering chatbots and virtual assistants capable of engaging conversations.
- Drug Discovery & Material Science: Generating novel molecular structures or material designs.
Notable Examples
- Generated Art: AI-generated visual content using tools like Midjourney, DALL-E, and Stable Diffusion
- GenAI Voice Cloning: Replicating human voices for various applications
- AI-generated porn: Controversial use case raising ethical and consent concerns
- Large Language Models: Text generation systems like ChatGPT, Claude, and others
Challenges & Considerations
- AI Ethics: Concerns about misinformation, deepfakes, bias in generated content, copyright infringement, and job displacement.
- AI Content Labeling: The need to identify AI-generated content to maintain transparency and trust.
- Content Provenance: Tracking the origin and modifications of generated media using standards like C2PA Content Credentials.
- AI Safety: Ensuring models do not generate harmful, biased, or unsafe content.
- Computational Cost: Training large generative models requires significant computing resources.
Controversies and Real-World Issues
Voice Cloning and Copyright
- Stephen Fry Case: Actor Stephen Fry’s voice was stolen from Harry Potter audiobooks and replicated by AI, demonstrating the risks of voice cloning without consent. Reddit discussion
AI in Film Production
- Lisa Kudrow on Tom Hanks’ “Here”: Actress criticized the film as “an endorsement for AI”, raising concerns about AI replacing human actors
Commercial Exploitation
- TeePublic T-Shirt Copying: Generative AI systems have been found searching for popular shirts and trying to copy them without full context, demonstrating automated copyright infringement. Examples: Penguin Baseball shirts page 1, page 2
Articles and Resources
Technical Developments
- Amazing Updates to Midjourney AI
- Training ChatGPT with Custom Libraries Using Extensions
- The Dual LLM Pattern for Building AI Assistants That Can Resist Prompt Injection
Tools and Frameworks
- OpenAI SHAP-E: Generate 3D Objects Conditioned on Text or Images
- Artists for AI Prompts: Resources for prompt engineering with artistic styles
- Retrieval Augmented Generation (RAG): Reducing Hallucinations in GenAI Applications
Connections
- Relies heavily on Machine Learning and deep learning techniques
- Builds upon Artificial Intelligence foundations
- Raises significant AI Ethics questions
- Necessitates AI Content Labeling and Content Provenance solutions like C2PA Content Credentials
- Drives development in AI Co-pilots and AI Companionship
- Faces AI Regulation Challenges globally, such as the EU AI Act
- Requires AI Transparency Requirements for responsible deployment
- Includes Large Language Models
- Enables Generated Art
- Powers GenAI Voice Cloning
- Related to AI-generated porn
- Utilizes Artists for AI Prompts
- Connects to AI Safety concerns