Course (Part 314)

• Describe what ChatGPT is and how it uses the transformer architecture (GPT‑4).
• Explain core NLP concepts relevant to conversational AI.
• Identify practical use cases: customer support, content creation, education, healthcare, and more.
• Demonstrate basic prompt engineering techniques to shape model output.
• Perform a simple integration using the OpenAI API and configure a basic chatbot flow.
• Recognize ethical concerns and implement basic safeguards (privacy, bias mitigation, human oversight).

ChatGPT is a pre-trained large language model (LLM) that predicts and generates text using patterns learned from extensive training data. It relies on the transformer architecture to model attention across tokens, enabling context-aware responses. Core NLP elements include tokenization, embeddings, sequence modeling, context windows, and fine-tuning. Prompt engineering shapes the model’s behavior by providing context, instructions, and examples. Integrations typically use an API to send messages and receive completions; developers control parameters like temperature and max tokens to balance creativity and determinism. While powerful, the model can hallucinate or reflect biases from its training data, so verification and human-in-the-loop designs are essential for sensitive applications.

English

ChatGPT is an advanced conversational AI built on GPT‑4 that generates coherent, context-aware text. It supports applications from chatbots to content generation, and integrates via API with developer controls for behavior and output.

বাংলা (Bangla)

ChatGPT হল GPT‑4 ভিত্তিক উন্নত কথোপকথন-ভিত্তিক কৃত্রিম বুদ্ধিমত্তা যা প্রাসঙ্গিক ও স্বচ্ছন্দ টেক্সট তৈরি করে। এটি চ্যাটবট থেকে কনটেন্ট নির্মাণ পর্যন্ত বিভিন্ন অ্যাপ্লিকেশন সমর্থন করে এবং API-এর মাধ্যমে ইন্টিগ্রেট করে ডেভেলপাররা আউটপুট ও আচরণ নিয়ন্ত্রণ করতে পারে।

• Always verify model outputs for factual accuracy before using them in decisions or publishing.
• Design systems with human oversight for high-risk or sensitive domains (medical, legal, safety-critical).
• Protect user privacy: minimize data retention, anonymize inputs, and follow applicable regulations.
• Tune parameters (temperature, max_tokens) to match use-case needs: lower temperature = more deterministic.
• Prompt clarity and context improve result quality—include role, format, and examples when needed.

1. ChatGPT always returns 100% factual answers. — False
2. Prompts with clear instructions produce more reliable outputs. — True
3. API integration is required to run ChatGPT on a website. — True (or an embedded product using API)
4. Lower temperature typically increases randomness in responses. — False (lower temperature reduces randomness)

1. GPT stands for Generative Pre-trained Transformer.
2. “Prompt engineering” means crafting inputs to guide the model’s output.
3. Use temperature to control creativity and max_tokens to limit length.
4. Human-in-the-loop is important for mitigating _____ (answer: errors/biases).

1. Which architecture underlies ChatGPT?
   • A. RNN
   • B. Transformer ✓
   • C. SVM
   • D. Decision Tree
2. To make replies more deterministic, set temperature to:
   • A. 1.0
   • B. 0.8
   • C. 0.2 ✓
   • D. 2.0
3. Which is a recommended ethical practice?
   • A. Share all user data publicly
   • B. No human oversight ever
   • C. Inform users they interact with AI ✓
   • D. Ignore data privacy

Describe a safe deployment pipeline for a customer support chatbot: collect minimal user data, route sensitive requests to human agents, log anonymized interactions for monitoring, apply validation checks on model responses (facts, links), implement rate limits, and provide a clear escalation path and user disclosure that they are interacting with AI.

Prompt challenge: Write a 60‑word product description for an eco-friendly water bottle in a playful tone and include a one-line CTA. Example prompt to the model: “You are a witty brand copywriter. Produce a 60-word playful product description for an eco-friendly stainless steel water bottle and end with a single-line call-to-action.”

1. Hands-on: Create a simple FAQ chatbot using the OpenAI API and host it on a static site.
2. Prompt lab: Iterate three versions of a prompt for the same task and compare outputs.
3. Ethics discussion: Review a case study where an AI produced biased output and propose safeguards.

• Transformer — neural architecture using attention mechanisms
• Token — a unit of text (word or subword) used by models
• Prompt engineering — designing inputs to guide outputs
• Fine-tuning — training a pre-trained model on task-specific data
• API key — credential used to access a remote service

1. Register for an OpenAI account and read the API quickstart guide.
2. Draft three prompts (instructional, contextual, comparative) for a chatbot you’d like to build.
3. Prepare a one-page ethics checklist to use when deploying conversational AI.

Build a mock “store assistant” conversation flow: design five user intents (product inquiry, order status, returns, shipping, feedback), craft example prompts and expected responses, and simulate a 6‑turn dialog for each intent. Submit transcripts and a short reflection on where human escalation is required.