← Back to Contents
Note: The first drafts of these discussion questions were generated using Claude (Anthropic's AI assistant) and then reviewed and edited for the in-class use of this course.
Week 1 • Discussion

💬 Week 1 — Discussion Questions

To think about before class — and the canonical guide to how these pages work

How these pages work (read this once)

Each week of this course ends with a Discussion Questions page like this one. They are designed for you to think about before the in-class session. They are not exam questions; most do not have a single right answer.

Each sub-lesson has four example questions attached, tagged so you know what kind of thinking each one is asking for: Calibrate applies the lesson's framework to a specific case; Apply connects the lesson to your own research practice; Critical asks where the lesson's argument might be wrong or incomplete; Connect links the sub-lesson to earlier or later material in the course.

These are example discussion points, not a checklist. You are not expected to engage with all of them. Pick the ones that speak most directly to your own research interests, and bring two or three rough answers — single paragraphs, not essays. You are welcome (and encouraged) to push back on the questions themselves if you think a framing is off; saying clearly why is itself a useful answer.

This page also stands as the canonical reference: the other Discussion pages across the course assume you have read this one once at the start, and they link back to it.

Sub-lessons

A Lightning Tour of AI

  1. Calibrate Watch the seminar and pick one claim or framing that has aged well between when it was recorded (in 2024) and now (May 2026), and one that hasn't. What changed in the gap, and what does that say about how to read AI talks recorded even a year or two ago?
  2. Apply The seminar threads together AI history and mathematical foundations. As a non-CS researcher, which thread — the historical narrative or the mathematical underpinning — is currently doing more work in your understanding of contemporary AI? What would change if you doubled down on the other one?
  3. Critical AI-history presentations often present a clean linear progression (Turing → expert systems → neural nets → LLMs). What is being smoothed over by that linearity, and what do the skipped or de-emphasised threads (symbolic AI, statistical methods, hardware) add when you put them back in?
  4. Connect Before this course started, what was your own working answer to “what is AI?” Hold it next to the seminar's framing. Where did the seminar reinforce what you already thought, and where did it overturn something you were taking for granted?

An Introduction to Transformers (3Blue1Brown video series)

  1. Calibrate After working through the videos, write your own one-sentence answer to “what does attention actually do?” Test that answer against the abstract of a recent transformer-based paper in your field. Where does the sentence hold, and where does it stretch?
  2. Apply The transformer architecture sits underneath nearly every LLM you will use this course. Which two or three properties of the architecture (context window, attention pattern, layer depth, tokenisation, etc.) most directly affect how you should read or trust LLM outputs as a researcher?
  3. Critical 3Blue1Brown's visual style makes hard concepts feel intuitively clear. Where might that visual intuitiveness be misleading — i.e. which mathematical move is dramatised in the animations in a way that doesn't quite match what transformers actually compute?
  4. Connect The history-of-AI lesson elsewhere in this week traces the path from biological neurons to modern neural networks. Where does the attention mechanism shown in these videos sit in that long arc — is it an incremental refinement of what came before, or a genuine break with it?

But What Is a Neural Network? (3Blue1Brown)

  1. Calibrate The video frames a neural network as a system that “learns to recognise patterns.” What is the strongest, broadest extension of that framing that is still accurate — and at what point does the framing start to do work it shouldn't?
  2. Apply Identify a concrete use-case in your own field where a small, well-understood neural network would be more appropriate than a frontier LLM. Be specific about the input, the output, and why the smaller model is the right call.
  3. Critical The video's pedagogical move is to build understanding “from the ground up” using visual analogy rather than formal mathematics. What does this approach gain, and what does it lose? What is the smallest amount of formal mathematics you would still want a researcher in your field to learn alongside it?
  4. Connect The History of AI lesson elsewhere in this week traces the long path that led to modern neural networks. Where in that history does the conceptual jump from “a neural network does pattern recognition” to “a network does something we'd be tempted to call understanding” actually start, and which intervening idea do you most want to understand better?

Current Generative AI Landscape

  1. Calibrate The page argues you should “track by model families, modalities, context limits, and agent/tool reliability — not by a single leaderboard.” Pick one recent leaderboard claim you have seen and rewrite it through this multi-dimensional lens. Does the claim survive?
  2. Apply Across the four major categories (LLMs, image, video, code/agentic), where will you invest most of your remaining attention budget over the next twelve months for your own research? What evidence would change that allocation?
  3. Critical The page says “agent reliability beats raw scores.” Construct the strongest counter-argument: when does raw model capability genuinely matter more than reliability of the surrounding agent system?
  4. Connect This snapshot is dated May 2026 and is explicitly described as ageing. What habit should you build, starting in Week 1, so that your reading of the AI landscape stays useful even as the specific page goes out of date?

Foundations of Generative AI

  1. Calibrate The lesson traces an arc from Turing through to modern generative AI. Pick the single biggest discontinuity in that arc — a place where progress wasn't merely incremental — and defend the choice against the obvious alternatives.
  2. Apply Assume you are reading AI-related papers in your own field critically for the next year. What is the minimum vocabulary — five to seven terms — you actually need from this lesson? Justify each choice in one line.
  3. Critical The lesson assumes no technical background. Where might that pedagogical choice hide something that actually matters when you go from reading AI papers to using AI in your own research?
  4. Connect Before this course, where did your sense of what AI “is” come from — popular press, product marketing, an undergraduate course, a colleague? Which sources were closest to the framing in this lesson, and which were furthest off?

Hands-On Exploration

  1. Calibrate In Activity 1 you will run the same prompt across three different tools. Where the responses diverge most, what do those divergences tell you about each tool's training data, fine-tuning, or guard-rails? Be specific.
  2. Apply Do the Research Relevance Mapping (Activity 3) for your own dissertation or current project. Which two parts of your workflow are the strongest candidates for AI assistance, and which two should you keep firmly under human-only control? What makes the difference?
  3. Critical Hands-on exploration is structurally biased toward what tools do well — you stop when you get a satisfying output. How would you design a complementary exercise that systematically surfaces what these tools do badly in your specific field?
  4. Connect The Timeline Exercise (Activity 2) asks you to chart AI's historical evolution. How does the historical view affect how you read present-day claims about AI capability — does the long view make you more skeptical, more receptive, or both?

History of AI: From Neurons to Neural Networks

  1. Calibrate The lesson covers two AI winters. What evidence would convince you that the current period is heading into a third, and what evidence would convince you it isn't? Try to make both lists about equally credible.
  2. Apply Pick one historical figure or moment from the lesson whose framing of AI you find most relevant to your current research thinking. Explain the connection in concrete terms.
  3. Critical The narrative is structured around “neurons → networks → modern AI.” Which other historical strands (symbolic AI, statistical methods, hardware progress, cognitive science, formal logic) are de-emphasised, and what does putting them back in change about the present picture?
  4. Connect The lesson covers two AI winters between the 1970s and 2006. Compare the way present-day AI is described in this week's other readings (the Lightning Tour seminar, the Current Generative AI Landscape page) with how earlier moments in the history were described at the time. Where are the rhetorical patterns the same?

Understanding How Generative AI Works

  1. Calibrate The lesson draws a sharp distinction between discriminative AI (“which category?”) and generative AI (“create new content”). In your field, where does that distinction get blurry in practice — e.g. where a generative model is actually being used as a sophisticated classifier?
  2. Apply The lesson explains key architectures (transformer, diffusion, GAN, VAE, etc.) in plain language. Which one is most directly relevant to what you would want to do with AI in your own research over the next year? What would you need to learn about it next?
  3. Critical The “plain language” approach drops most of the formal mathematics. Where might a researcher reading only this lesson form a mental model of these systems that misleads them when they actually try to use the tools?
  4. Connect Compare the conceptual vocabulary in this lesson with the way the History of AI lesson elsewhere in this week talks about the same systems. Where do the two framings use different words for what is essentially the same idea, and where do they actually mean different things?