2025-11-06 | RRG
Goal: Recursion Reading Group ML Session 2
Summary: Design/teach curriculum for motivating multiply add (dot product) intuition
Work sessions
| In | Out |
|---|---|
| 11:30 | 12:30 |
| 20:15 | 21:30 |
AI Safety
- Kicked off the meeting going over the abstract of Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
Neural Networks from Scratch
- During the second meeting (See first meeting: 2025-10-30) of Neural Networks from Scratch (a new chapter of Recursion Reading Group (RRG) discussions), I met up with my co-lead to motivate ML from first principles before diving into the fancy math and libraries
- Specifically, we derived the intuition for multiplication as a similarity metric and addition as a way to aggregate similarity over pattern matching of an \(x\) and \(w\) vector.
- Since the cohort consists of Computational Linguistics, CS, and Cognitive Science majors, we used C++ to program a dot product calculator (on paper, note that USC's intro CS classes are in C++) and mapped the Dot Product to the first layer of a Neural Network
- Hopefully, by making this event unplugged and based on intuitions first, it provides the cohort mates a chance to talk and converse about AI Safety in addition to making sure no one is left behind