Description
This comprehensive, non-coding session provides an architectural walkthrough of Large Language Models (LLMs), bridging the gap between high-level AI concepts and practical model utilization. The session traces the historical evolution of Natural Language Processing (NLP), tracking how early statistical models (N-grams, Bag-of-Words) and sequential deep learning networks transitioned into modern self-attention paradigms. Participants examine the inner workings of the Transformer architecture, exploring tokenization, embeddings, multi-head self-attention, and the structural trade-offs between encoder-only, decoder-only, and encoder-decoder designs, while analyzing how context window limits dictate memory consumption and computational complexity.
Beyond core mechanics, the session evaluates the current LLM ecosystem through model benchmarks, comparing leading families (OpenAI, Anthropic, Google, Meta, and DeepSeek) across parameters like reasoning, cost efficiency, multimodal capabilities, and alignment. Finally, through guided visual demonstrations and concrete examples, participants discover how to access these models via proprietary APIs and open-source platforms (such as Hugging Face), identify key failure modes (hallucinations, bias, prompt injection), and learn some applied prompt engineering patterns to reliably direct model behavior.
Prerequisites
Familiarity with computers
Details
Learning Objectives
- Explain the core mechanics of Transformers, including text embeddings, multi-head self-attention mechanisms, and the structural differences between encoder-only, decoder-only, and encoder-decoder architectures.
- Analyze how raw text is broken into tokens and how context window constraints impact both computational complexity and model memory.
- Identify common model failure modes, including hallucinations, context loss, data privacy concerns, prompt injection vulnerabilities, and socio-cultural biases in training datasets.
- Trace the historical progression of Natural Language Processing from early statistical approaches (N-grams, Bag-of-Words) to sequential deep learning architectures and analyze how parallel processing and self-attention mechanisms in Transformers allowed modern LLMs to effectively replace these predecessors.
- Benchmark and contrast leading model families including OpenAI (GPT series), Anthropic (Claude series), Google (Gemini series), META (Llama) and DeepSeek across key selection criteria such as context window capacity, reasoning capabilities, multimodal inputs, cost efficiency, and alignment/safety strategies.
- Explore various ways to communicate with LLMs, including proprietary LLM APIs and leverage open-source tools such as Hugging Face transformers.
- Formulate effective prompt design patterns including zero-shot, few-shot, system persona constraints, and Chain-of-Thought (CoT) prompting to steer reasoning and enforce structured outputs.
Who Should Enrol
This workshop is for anyone interested in Generative AI (no coding required) looking to understand Transformer architectures, evaluate modern LLMs, and learn some practical prompt engineering techniques.