1 Week 1: Setting the Stage – Intro to Machine Learning & AI (01/22 & 01/23) 1 Week 1: Setting the Stage – Intro to Machine Learning & AI (01/22 & 01/23)
1.1 Day 1 1.1 Day 1
1.1.1. “Visualizing the deep learning revolution” by Richard Ngo, 2023
This reading shows, by examples across a variety of application areas, the dramatic increase in capability of AI systems over the past few years. Extrapolate these improvements another few years, what will these systems be capable of then?
1.1.2. “Machine Learning: A Primer” by Lizzie Turner, 2018
This reading describes a pre-LLM view of machine learning and serves as foundation knowledge for ML that we'll build on over the semester. What problems or applications do you think the methods discussed in this reading are superior to transformer-based models (e.g., LLMs) for?
1.1.3. "What Is ChatGPT Doing … and Why Does It Work?—Stephen Wolfram Writings” by Stephen Wolfram, 2023
This reading is important to begin to understand how LLMs and ChatGPT work. The LLM underlying ChatGPT is simply generating text by repeatedly predicting the next token, given a history of tokens. What types of text do you think would be challenging to generate well in this way?
1.1.4. Perma | The Building Blocks of Interpretability
This reading is important to begin to understand how neural networks represent and combine concepts. What kind of interfaces into a neural network model would help you better understand how the model is operating? What kind of interfaces would help to increase your trust in how the model is operating?
1.2 Day 2 1.2 Day 2
1.2.1. A Path Toward Autonomous Machine Learning by Yann LeCun, 2022 (read pg 1-9)
This reading is a leading AI researcher's perspective on creating autonomous intelligent agents. To what degree do you think autonomous intelligence agents must go through a similar development trajectory as humans (e.g., Figure 1) to achieve the level of broad competency humans have? To what degree do you think having a biologically-inspired architecture (e.g., Figure 2) is important to developing autonomous intelligent agents?
1.2.2. “The Bitter Lesson” by Rich Sutton, 2019
This reading is a leading AI researcher's perspective on how the field of AI research oversteers toward "building in how we think we think we think" rather than relying on scaling computation. Can you think of counter examples where building in how we think that we think won't inhibit progress in the long run? What does the future of academic AI research look like if large, costly computational resources are required to make progress?
1.2.3. [Recommended] The Mythos of Model Interpretability by Zachary Lipton, 2017
This paper disambiguates various notions of interpretability. For LLMs, what definition of interpretability is most relevant? How does this vary by application area? For image generation models, what definition of interpretability is most relevant? How does this vary by application area?