<aside>
👋 Welcome! Below are all the topics and readings we’ll cover during the BAIST AI Policy program this spring. For each session, we aim to do ~1 hour of reading and ~1 hour of discussion in-person (though you’re welcome to do the optional readings beforehand). No work will be required outside of our weekly 2-hour meeting.
This program will introduce you to the field of AI policy with a focus on safety and catastrophic risk prevention. The goal of the program is to equip you with the knowledge to begin a career in AI safety.
Apply here by Saturday, September 14
</aside>
<aside>
🚨 FOR PARTICIPANTS ONLY
Fill out this onboarding form (~5 mins). Most fields are optional, feel free to leave them blank. Make sure you select AI Safety Fundamental Governance for the Course track.
</aside>
Session 1: Introduction to Machine Learning
- But what is a neural network? (3Blue1Brown, 2017). Watch first 12.5 min, rest optional.
- The AI Triad and what it means for national security strategy (Buchanan, 2020) 21 min.
- Visualizing the deep learning revolution (Ngo, 2023). 15 min.
- 4 charts that show why AI progress is unlikely to slow down (Henshall, 2023). 8 min.
- **How ChatGPT Works, Technically** (ByteByteGo, 2023). 8 min.
Optional readings:
- **Gradient descent, how neural networks learn (3Blue1Brown, 2017). Video; 20 mins.**
- **A short introduction to machine learning (Ngo, 2021). 11 min.**
- **The bitter lesson** (Stutton, 2019)
Session 2: Overview of risks from advanced AI systems
- [Overview of how AI might exacerbate long-running catastrophic risks](https://aisafetyfundamentals.com/blog/overview-of-ai-risk-exacerbation?_gl=1*13q1d3h*_ga*MTkyNDUzNTAyNS4xNzA2MjIzMzI0*_ga_8W59C8ZY6T*MTcwNjIyMzMyMy4xLjEuMTcwNjIyNDYyOC4wLjAuMA..) (BlueDotImpact, 2023) 15 min. Alternate paper: An overview of catastrophic AI risks (Hendrycks et al., 2023).
- **2022 Expert Survey on Progress in AI (Stein-Perlman et al., 2022). Read “Summary of results.” 2 min.**
- Harms from increasingly agentic algorithmic systems (Chan et al., 2023). Read pgs. 2-6 (§ 1 & 2). 10 min.
- Why AI alignment could be hard with modern deep learning (Cotra, 2021). 20 min.