The world around us is changing fast with the integration of AI in our everyday life. Are you someone who is passionate about AI safety and/or has worried about how it is currently being built and deployed? We may be the last generation with a real say in how machine intelligence turns out, and this 6-week fellowship gives you the technical grounding to think seriously about it. No ML background required.
Timing: 5:30-7:30 (tentative) Monday
Location: UBC Campus (room details will be sent out)
Snacks will be provided!!! (this is your incentive guys)
Content
Week 1: Foundations
How modern AI actually works: neural networks, transformers, and how LLMs are trained. What "capabilities" means and how we measure them.
Week 2: Scaling and Forecasting
Scaling laws, benchmark prediction, and inference-time compute. Building your own model of how fast progress is likely to go and what could bottleneck it.
Week 3 : The Alignment Problem
Why capable systems don't inherit our goals by default. Specification gaming, mesa-optimization, and the gap between what we train for and what we get.
Week 4: Training Methods and Their Limits
RLHF, constitutional approaches, and where they break down. How does alignment faking work in frontier models?
Week 5: Evaluations and Catastrophic Risk
Dangerous capability evals, red-teaming, and ASL-style thresholds. Misuse, misalignment, and how complex systems fail.
Week 6: Current Affairs and Next Steps
Where the field stands right now: regulation, compute governance, lab commitments, and open research problems. How to get involved.
