Skip to main content

AI in a nutshell

From AI models, over basic concepts, to AI engineering.

Summary​

AI has undergone a transformative shift since 2020, primarily driven by the scaling of models. This growth has led to increased capabilities, enabling numerous new applications and significantly boosting productivity and economic value. However, training these massive models requires vast amounts of data, computing resources, and specialized expertise—resources typically available only to large organizations. Consequently, the model-as-a-service paradigm has emerged, democratizing access to AI by allowing individuals and smaller teams to build powerful applications using pre-trained foundation models provided as a service.

Types of AI Models​

  • Language Models (LMs): Models that predict the likelihood of a word or token appearing in a given context. They form the foundation of many AI applications, such as text generation, summarization, and translation.
  • Large Language Models (LLMs): Extremely large-scale language models trained on vast datasets with billions of parameters, enabling sophisticated tasks such as content generation, coding, and open-ended reasoning.
  • Embedding Models: AI models that transform text, images, or other data into numerical vector representations, enabling tasks like semantic search, recommendation systems, and retrieval-augmented generation (e.g., CLIP for text-image embeddings, word2vec, sentence transformers).
  • Generative AI Models: Models designed to generate new content, such as text, images, music, or code, based on learned patterns.
  • Foundation Models: Large, general-purpose AI models capable of handling multiple data modalities (text, images, audio, video) and supporting various downstream tasks with minimal fine-tuning. Examples include GPT-4, Gemini, and multimodal models like CLIP and DALL·E.

More specialized models are:

  • Text Generation (e.g., GPT-4, Claude)
  • Image Generation (e.g., DALL·E, Midjourney, Stable Diffusion)
  • Music & Audio Generation (e.g., Jukebox, MusicLM)
  • Conversational AI Models: Specialized LLMs optimized for interactive, dialogue-based applications, such as virtual assistants, chatbots, and customer support agents (e.g., ChatGPT, Google Bard, Meta’s Llama).
  • Transformer-Based Models: AI models leveraging transformer architectures for superior efficiency and parallel processing. The transformer structure powers most modern AI models, including LLMs, multimodal models, and generative AI (e.g., BERT, GPT, Vision Transformers [ViTs]).

Basic AI concepts​

  • Tokens: Fundamental units (words or parts of words) used by language models to process text efficiently.
  • Vocabulary: The complete set of tokens recognized by a model.
  • Natural Language Processing (NLP): AI field enabling computers to understand, generate, and respond to human language. Applications include translation, summarization, sentiment analysis, classification, and conversational agents.
  • Multi-Modality: refers to AI systems that can process, integrate, and understand multiple types of data (e.g., text, images, audio, video) simultaneously. These models enable cross-modal interactions, allowing for tasks such as image captioning, video analysis, speech-to-text conversion, and multimodal AI assistants (e.g., GPT-4V, Gemini, CLIP).
  • Masked Language Models (e.g., BERT): Predict missing words using context from both directions. Suitable for context-rich tasks like sentiment analysis or code debugging.
  • Autoregressive Language Models (e.g., GPT): Predict sequential tokens based solely on prior text. Ideal for generating coherent, extended text and conversation.
  • Self-supervision: A method where models learn from unlabeled data by predicting parts of the input itself, overcoming limitations of manual labeling. This technique allows language models to scale by leveraging abundant unlabeled text.
  • Machine Learning (ML): The broader AI field where algorithms learn from data to make predictions or decisions without being explicitly programmed.
  • Deep Learning: A subset of ML using neural networks with multiple layers (e.g., transformers in LLMs) to recognize patterns and generate complex outputs.
  • Generative AI: AI that creates new content (text, images, music) rather than just analyzing or classifying existing data (e.g., DALL·E, ChatGPT).

AI Engineering vs. ML Engineering​

  • ML Engineering involves building and deploying machine learning models focused on specific tasks, usually with labeled data.
  • AI Engineering is the emerging practice of leveraging powerful, general-purpose foundation models. Key techniques include:
    • Fine-tuning: The process of adapting a pre-trained model to a specific task by further training it on a smaller, domain-specific dataset. (further adapting models to specific tasks or contexts)
    • Retrieval-Augmented Generation (RAG): A technique that enhances AI-generated responses by integrating real-time external knowledge from databases or documents. (integrating external data sources)
    • Prompt Engineering: The practice of designing effective prompts to guide AI models toward generating desired responses. Techniques include structured prompts, few-shot examples, and iterative refinement for optimal results. (crafting effective instructions for models)

Growth and Impact of AI Engineering​

AI engineering has rapidly become one of the fastest-growing tech disciplines. The reduced costs and improved ease of development have encouraged widespread adoption across industries, enabling innovative applications in content generation, marketing, coding assistance, and beyond. As AI continues evolving, these approaches will only become more integral to technological advancement.