Module 3: Neural Nets and Large Language Models
Mandatory Watching - videos with quizzes - finish before 11 am, Tuesday, June 30
- 3.1 — Lecture: Neural Nets, 31 minutes (Uploaded March 31, 2026)
- 3.2 — Lecture: Large Language Models, AI Assistants, and Reasoning Models, 52 minutes (Uploaded March 31, 2026)
- 3.3 — Art of the Problem - How AI Learns, 15 minutes (Uploaded Nov. 14, 2019)
- 3.4 — Emergent Garden - Why Neural Networks Can Learn (Almost) Anything, 10 minutes (Uploaded March 12, 2022)
- 3.5 — 3blue1brown - Large Language Models Explained Briefly, 9 minutes (Uploaded Nov. 20, 2024)
- 3.6 — Sunny Israni - Claude Cowork vs OpenClaw (about AI agents), 18 minutes (Uploaded Feb. 19, 2026)
Written assignment - finish by end of Tuesday, June 30, give comments by meeting on Thursday, July 2
Access three LLM assistants and one pure LLM text completion engine. Get them to answer a question from the past assignment about your assigned fiction (see spreadsheet (link removed)). Copy and paste the conversations into files, say which you think is best, and submit here.
Further watching/reading
- How people are using different LLMs
- Interactive timeline of LLMs from my slides
- Market share of the different AI companies: by user chats (from FirstPageSage, updated regularly), by corporate spending (from Menlo Ventures, as of July, 2025)
- Anthropic just passed OpenAI in revenue, Jason Lemkin at SaaStr.ai (April 2026)
- The Anthropic Economic Index (March 24, 2026) - Anthropic’s data on how people are using Claude
- Marc Zao-Sanders, Top ten use cases for gen AI (2024, 2025)
- Andrej Karpathy, Deep Dive into LLMs like ChatGPT, 3 hours 30 minutes (February 2, 2025) - a really deep and detailed explanation that helped me learn a lot even after I had been teaching this class for several months! (I downloaded it to watch on an airplane.) It’s mostly accessible to a general audience.
- Welch Labs - ChatGPT is Made from 100 Million of These (the Perceptron), 24 minutes (Jan. 31, 2025) - an excellent introduction to how neural nets and perceptrons work and are trained, that uses a little bit of calculus
- Murray Shanahan, Talking about Large Language Models, (Feb. 12, 2024) - a really clear discussion of the ways in which the text produced by a LLM “looks like” human beliefs, but is philosophically quite different - though note that it also pre-dates modern “reasoning” models
- Understanding transformers
- 3blue1brown - Transformers (How LLMs Work) Explained Visually, 27 minutes (Uploaded Apr. 1, 2024)
- Timothy B Lee, A Jargon-Free Guide to How LLMs Work, in Wired, somewhat easier to follow
- Thomas Woodside and Helen Toner, several part series explaining how LLMs work
- Jay Alammar, The Illustrated GPT-2, very, very detailed explanation
- Steve Newman, An Intuitive Explanation of Large Language Models (April 25, 2023)
- Evaluations of early text-prediction LLMs:
- OpenAI’s article at the launch of GPT-2 (Feb. 14, 2019) - note that their responses were generated by the larger version rather than the small version we can still try, but even so, the results are much less coherent than anything generated by current models
- Back and forth between Scott Alexander and Gary Marcus
- Scott Alexander, Do Neural Nets Dream of Electric Hobbits? (Feb. 18, 2019)
- Gary Marcus, GPT-2 and the Nature of Intelligence, (Jan. 25, 2020)
- Scott Alexander, The Obligatory GPT-3 Post (June 10, 2020)
- Ernest Davis and Gary Marcus, Experiments in Commonsense Reasoning in GPT-3: Status Report from June 2022
- Scott Alexander, My Bet: AI Size Solves Flubs (June 6, 2022), a response to Davis and Marcus
- Vox, Computers Just Got a Lot Better at Writing, 7 minutes (March 4, 2020) - three years before ChatGPT, when GPT-3 was released, see how people were thinking about it
- Historic research papers:
- Vaswani, Shazer, et al, “Attention is All You Need” - the 2017 paper that introduced transformers
- Kojima et al, “Large Language Models are Zero-Shot Reasoners” (May 2022) - introduced reasoning models
- François Chollet, OpenAI o3 Breakthrough High Score on ARC-AGI-Pub (Dec. 20, 2024)
- Try the task yourself - see how much you “talk to yourself” while solving it (another site)
- “Spreadsheets are all you need” - an implementation of GPT-2 in a series of spreadsheets - it takes about 60 seconds to generate even one word of response, but you can see all the steps of the calculations it goes through for each word