#01 Orient
Why this, why now
The case for taking advanced AI seriously: capabilities, timelines, and what it would mean for things to go well or badly.
SAISI-F26-FEL · Fall 2026
The introfellowship
An eight-week reading group on how powerful AI systems fail, who is working to prevent it, and what a Swarthmore student can actually do about it.
CatalogSAISI-F26-FEL
TermFall 2026
Length8 weeks
WhenThursdays · 6:30–8:30 PM
WhereRoom TBA, Science Center
First sessionOct 1, 2026
Cohort12–16 fellows
CommitmentTwo hours a week. All reading happens in the room.
PrerequisitesNone. No CS, no math, no prior reading.
FoodDinner provided
DeadlineSep 25, 2026
Syllabus8 weeks
Reading list · Discussed in the room
Week by week
#02 Open
How the machines are built
Pretraining, fine-tuning, RLHF, and scaling laws, explained without code. Enough to read the rest of the syllabus.
#03 Break
How they fail
Specification gaming, reward hacking, goal misgeneralization. Real incidents, real transcripts.
#04 Doubt
Deception and control
Situational awareness, sandbagging, and whether we can keep a system that is smarter than its overseers in check.
#05 Look
Inside the model
Interpretability: what circuits, features, and probes have found so far, and what they have not.
#06 Test
Evaluations
Dangerous-capability evals, red teaming, and the organizations whose job is to say no before a launch.
#07 Govern
Labs, states, and rules
Compute governance, safety frameworks, export controls, and who actually decides.
#08 Act
What you do next
Research programs, policy fellowships, career paths, and a short project you present to the cohort.
Questions
- Do I need a CS background?
- No. The fellowship is written for anyone at Swarthmore who is curious and willing to argue. Philosophy, econ, bio, and CS majors all get something different out of it.
- How much work is it?
- Two hours a week in the room. There is no homework. Readings are short and we do them together.
- Is it a class?
- No credit, no grades. A small cohort, one facilitator, dinner.
- Who pays for this?
- SAISI is supported by Coefficient Giving through Project Kairos's Pathfinder fellowship, the same program behind groups at Harvard, MIT, Yale, and Berkeley.
- Can Haverford or Bryn Mawr students apply?
- Yes. We are the first Tri-Co group and want the cohort to reflect that.