Topic 7D - Large Language Models

Learning Objectives

By the end of this topic, you should be able to:

• Explain what a large language model is: a neural network trained on massive text corpora using next-word prediction.
• Describe the three phases of LLM development: pretraining on unlabeled text, supervised fine-tuning on human-written examples, and reinforcement learning from human feedback (RLHF).
• Explain what hallucination is and why it is a systematic property of how LLMs work rather than a correctable bug.
• Identify at least two other systematic limitations of LLMs and explain why each arises from how LLMs work.
• Given a proposed use of an LLM in an educational context, evaluate it by identifying relevant capabilities and limitations.

Learning Activities

To help you meet the learning objectives, we have prepared three readings. Please complete them in order.

Readings

• Reading 1 - What an LLM Actually Is — transformers, attention, scale, and what the pretraining process actually produces inside the model
• Reading 2 - How LLMs Are Made Useful — the gap between a pretrained model and a helpful assistant, instruction fine-tuning, RLHF, and why the same base model can become very different products
• Reading 3 - What LLMs Can and Cannot Do — genuine capabilities, systematic failure modes, what this means for teachers, and the questions worth asking about any LLM deployment in education

These readings intentionally build on each other, so please complete them in order.

Checking for Understanding

Review the Learning Objectives at the top of this page. The questions below will help you check your understanding before the Week 7 small group discussion and in preparation for the AI Competency Demo.

What an LLM Is

1. Explain in your own words what a transformer's attention mechanism does. Why does the ability to weigh the relevance of distant words matter for understanding language?
2. An LLM is trained by predicting the next word across hundreds of billions of words of text. In what sense does this task require the model to learn something deeper than statistical word sequences? What does successfully predicting the next word require the model to "know"?
3. The textbook discusses ELIZA, which produced understanding-seeming outputs by pattern-matching sentences and reflecting them back as questions. In what ways is a modern LLM like ChatGPT similar to ELIZA? In what ways is it fundamentally different? Does the difference matter for how we should use it?

How LLMs Are Made Useful

4. Explain the difference between a pretrained LLM and a fine-tuned one. What would interacting with a raw pretrained model look like? Why is it not useful as a product?
5. RLHF uses human raters to score model outputs, and those scores serve as the reward signal for further training. What are the risks of this approach? What happens if the raters have systematic biases or if the model learns to optimize for high ratings rather than genuinely good answers?

What LLMs Can and Cannot Do

6. A student submits an essay written with ChatGPT's help. The essay contains a confident, plausible-sounding citation to a research paper that does not exist. Explain why this happened. Is it a bug that will eventually be fixed, or a systematic property of how LLMs work?
7. A teacher wants to use an LLM to provide personalized feedback on student writing. Identify two genuine capabilities of LLMs that make this potentially useful, and two systematic limitations that make it potentially problematic. On balance, what questions would you want answered before adopting this approach?
8. A student says: "ChatGPT knows everything — it's smarter than any human." Based on what you learned in this topic, identify three specific things that are wrong or misleading about this statement, and explain each one.

It is completely fine to revisit the readings as you work through these questions.

Extend Your Learning