Module 4: Classifiers and Bias
Mandatory Watching - videos with quizzes - finish before 11 am, Thursday, July 2
- 4.1 — Lecture: Expert Systems, 30 minutes (April 6, 2026)
- 4.2 — Lecture: Neural Classifiers, 37 minutes (April 6, 2026)
- 4.3 — Computerphile - Tricking AI Image Classification, 12 minutes (July 27, 2022)
- 4.4 — Wall Street Journal - How TikTok’s Algorithm Figures You Out, 13 minutes (July 21, 2021)
- 4.5 — Crash Course AI: Algorithmic Bias, 11 minutes (Dec. 13, 2019)
- 4.6 — Kate Cassidy: How AI Is Manipulating Your Mind With Flattery, 18 minutes (Aug. 11, 2025)
Written assignment - finish by end of Thursday, July 2, give feedback by meeting on Tuesday, July 7
Use Claude to “vibecode” a working computer game, inspired by your assigned fiction. Note that a free account on Claude only lets you do so much every 5 hours, so you should probably start this one a bit early. Let me know if you need an extension.
Further watching/reading
- History of the 1980s boom
- Interactive timeline of movies about AI, documenting the early ’80s boom
- 1985 Congressional Quarterly report on artificial intelligence
- More on machine learning classification
- Welch Labs - The Moment We Stopped Understanding AI [AlexNet], 17 minutes (July 1, 2024)
- Hilary Mason Explains Machine Learning at 5 Different Levels, 26 minutes (Aug. 18, 2021)
- Crash Course AI: How YouTube Knows what You Should Watch, 11 minutes (Nov. 22, 2019)
- Timothy B Lee, How Computers Got Shockingly Good at Recognizing Images, in Wired (Dec. 18, 2018)
- Welch Labs, The F=ma of Artificial Intelligence (Backpropagation), 30 minutes (June 10, 2025)
- Emergent Garden, Gradient Descent vs Evolution, 24 minutes (March 1, 2025)
- Sycophancy and psychosis
- TypeBulb - “You’re Absolutely Right!” - a test that measures how much different AI models agree with people even when they shouldn’t. Scroll down and click on individual lines to see what good and bad responses to given prompts look like.
- Dr. John Kruse - AI Induced Psychosis, 25 minutes (Sept. 14, 2025)
- Anthropic, Towards Understanding Sycophancy in Large Language Models (Oct. 23, 2023) - a research paper demonstrating how the training process makes models treat the customer as always right
- Ethan Mollick, Personality and Persuasion (April 30, 2025) - a discussion of a moment when ChatGPT accidentally became a complete sycophant, always telling people that they are great at everything
- Adele Lopez, The Rise of Parasitic AI (Sept. 10, 2025) - a diagnosis of the kinds of AI interactions that led to the most psychosis (particularly with ChatGPT 4o, from April to August of 2025)
- More on the failures of machine learning
- These stickers make computer vision systems hallucinate, article at The Verge (Jan. 3, 2018), also Magic AI: These are the optical illusions that trick, fool, and flummox computers (Apr. 12, 2017)
- Janelle Shane - The Danger of AI is Weirder than you Think, 10 minutes (Nov. 13, 2019)
- Adversarial attacks on deep learning models, more details on how adversarial images can be generated
- Vox, Are We Automating Racism?, 23 minutes (March 23, 2021)
- josh:), I Trained an AI to Make Me Laugh - It Worked Too Well, 8 minutes (June 10, 2025) - shows how training works, but also how it can get stuck in suboptimal strategies
- Golden Gate Claude
- Researchers at Anthropic published a paper in May 2024 about how they interpreted some of the neurons in Claude
- They also released a version of Claude where the “Golden Gate Bridge” neuron was turned up, and it kept bringing up the bridge no matter what you asked it about. See this Reddit thread of examples.
- You can explore features they identified - the neuron for Amelia Earhart was somehow not far from the ones for Christopher Columbus, the Hindenburg, parachutes, and the Roswell incident