AI Safety Fellowship
Six weeks, online, in a small group. Each week takes one question about how advanced AI can go wrong and what people are doing about it, with short readings before a two-hour discussion. You don’t need to code, and you don’t need to have read anything yet.
One question a week.
How could this actually go wrong?
What alignment means, the ways advanced AI could cause serious harm, and how to describe a failure by what causes it rather than how it looks in a headline.
Why does a model tell you what you want to hear?
How models learn from human approval, and how that same training produces flattery, gamed objectives and behaviour that changes when a model thinks it is being watched.
How do you check work you can’t check?
Supervising a system that knows more than you do: debate, weak-to-strong generalisation, and reading a model’s reasoning to catch it cutting corners.
Can we see what a model is doing inside?
What interpretability research has actually shown so far, and where the claims run ahead of the evidence.
If you can’t trust it, can you still use it?
AI control, and the safety plans frontier labs have published: what they promise, and who checks that they keep to it.
Where is this going, and where do you fit?
Why people disagree about how fast this moves, and a plan for your next three months: what to apply for, what to build and who to write to.
How it works.
Sessions are two hours, once a week, on a video call at a time that suits the group. Before each one there is under an hour of reading, mostly short explainers written for newcomers, and a short piece of writing. Expect about four hours a week in all. There is no fee.
The group is about ten people, with the same facilitator for all six weeks. The last session is about what you do next, and anyone who wants to can take on a small project and present it at the end.
Who it’s for.
Anyone in India who wants to understand how advanced AI can go wrong and what people are doing about it. Students, engineers, researchers and people working in law, policy or economics all have something to bring.
You don’t need to be able to code or have a background in AI safety. What matters is curiosity, and being able to give it the time.
How we choose.
The application takes about ten minutes. Most of it is practical. The question that decides most of it asks you to pick an AI system you use and describe one way it could fail badly, and it rewards curiosity rather than what you already know.
Everyone who applies hears back either way. If we can’t take you this time, you’ll be the first to hear when the next cohort opens.