← 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 3 • Discussion

💬 Week 3 — Discussion Questions

To think about before class

These are example discussion points for you to think about before class. You are not expected to engage with all of them — pick the ones that speak most directly to your own research, and bring two or three rough answers to the in-class session. The full description of how to use these pages, including what the question tags mean, is on the Week 1 Discussion page.

Sub-lessons

What Does AI Actually Consume

  1. Calibrate Pick a recent media claim about AI energy use (per-query cost, comparison to a household, an analogy to flying). Place it on the spectrum from “defensibly bounded” to “hand-waving” using the lesson's framing. What single number would change your placement?
  2. Apply Estimate the energy and water cost of one week of your own current AI-assisted research workflow, using the lesson's per-interaction numbers. Where are you most uncertain, and which uncertainty would actually change your behaviour if resolved?
  3. Critical The lesson emphasises a “transparency problem”: vendors don't publish the numbers you would need. Is the appropriate response (a) push harder for disclosure, (b) build your own measurement, or (c) accept the opacity and use proxy metrics? Defend the trade-off.
  4. Connect Week 2 traced what happens during training: the staggering compute cost of pre-training, the optimisation pipeline, the data scale. This sub-lesson traces what happens during inference. Looking at both halves together, which is the more honest place to focus environmental concerns — the one-time training of frontier models, or the rolling per-query cost spread across billions of users?

Critical Minerals and AI

  1. Calibrate The lesson maps the supply chain from mine to data centre. Pick one frequently-cited material (e.g. cobalt, lithium, rare earths) and trace what specifically about its supply chain is most fragile or most ethically charged. Where do the assumptions about “clean tech” AI break down?
  2. Apply If you were writing a one-paragraph “materials and methods” statement that acknowledged the supply-chain footprint of the AI tools used in your research, what would it actually say? What would it have to leave out for honesty's sake?
  3. Critical The geopolitical-risk framing for critical minerals can be deployed (a) to argue for ethical sourcing or (b) to argue for resource nationalism. The same supply-chain facts support both. How do you tell which framing a given commentator is making?
  4. Connect Week 2 covered what is needed to build and train modern AI models — staggering compute, careful data curation, expensive optimisation. This sub-lesson covers what is needed to manufacture the underlying hardware. Putting the two stacks side by side, where do the bottlenecks each sub-lesson identifies actually interact, and where do they appear to be independent?

Infrastructure Scale and the Rebound Problem

  1. Calibrate The rebound problem is that efficiency gains keep being absorbed by greater consumption. Pick one specific recent efficiency announcement (better chips, smaller models, more efficient training) and read it through the rebound lens. Is the announcement net-positive, neutral, or net-negative for total consumption?
  2. Apply For a researcher trying to minimise the environmental footprint of their AI use, which lever genuinely helps and which is a feel-good distraction? List two of each and defend the categorisation.
  3. Critical The lesson presents embodied carbon (cost of building the hardware) as comparable in significance to operational carbon. If you doubled the lifetime of the hardware, how much of the picture changes? Where are the brittlest assumptions?
  4. Connect Week 2 introduced scaling laws — the empirical claim that more compute keeps producing better models. This sub-lesson introduces the rebound problem — the empirical observation that gains in efficiency keep being absorbed by greater consumption. Read scaling laws and the rebound problem against each other: do they describe the same dynamic from different angles, or are they actually in tension?

Sustainable AI: What Can Be Done

  1. Calibrate The lesson lists practical choices that “make a difference.” Pick one practice from your own work that you genuinely think reduces your AI footprint, and one that you suspect is performative. Honestly defend the second one as performative.
  2. Apply Pick one of the AI-for-environmental-solutions cases discussed and ask: would you trust it in your field? If the answer is yes, what makes it credible; if no, what would change your answer?
  3. Critical The policy-landscape section covers various instruments (disclosure rules, carbon taxes, etc.). Which of these would most directly affect how you do research, and which would barely register? Why?
  4. Connect Week 2's discussion of RLHF and alignment frames training as a place where labs make value-laden trade-offs invisible to users. This sub-lesson asks the same question about how labs publish (or don't publish) sustainability data. Are these two instances of the same opacity problem, or genuinely distinct cases that need different remedies?