Introduction

This is my personal report for the recently held Machine Learning for Good (ML4Good Bootcamp) Singapore, Sept 20-28, 2025.

ML4Good provides intensive in-person bootcamps to upskill people in AI safety. The bootcamps have been held in various parts of the world (mainly Europe and LatAm). ML4Good Singapore, to the best of my knowledge, was the first ML4Good bootcamp in Asia. You can see more information at their page.

There has been similar posts in the LW community as well (for example, see this and this) .

Curriculum and Schedules

The bootcamp covers a lot of topics related to AI safety, including

General ML prerequisites technical topics: optimizers and hyperparameters, transformers, agents
Topics that are relevant to AI safety, such as capabilities, forecasting, risks, and strategies.
AI safety technical topics, such as interpretability, RLHF + CoT. Most of these technical topics include hands-on workshops besides lecture.
Governance-related topics, e.g. reading latest AI bills

Our main book is the AI Safety Atlas written by CeSIA. The first three chapters were prerequisites for the bootcamp, and my impression is that the course was organized around those chapters.

A usual day started at 9 AM and ended formally at 7.30 PM. A single day usually consisted of mixes of lectures, hands-on technical sessions, and other workshops in the format of discussions. The mix was varied; for example, our first day was mostly filled with hands-on sessions, whereas some other days lectures and discussions are more common.

Besides the lecture-style sessions, we also had one-on-one sessions between participants and career planning. For the one-on-one session, each participant was assigned a partner and given some time to talk through each of their career plans and provide feedback. Career planning was done by the instructors to help the participants solidify their career plans and they provided feedback as well.

The last major component of the bootcamp was the final project. All participants were given roughly 2 days (10 hours) to do AI safety related topics of their interests. A large number of participants worked together to set up accountability systems for their current/future AI safety endeavors (e.g. fellowships, field building), and the rest did mixtures of governance and technical works on quite diverse topics, e.g. governance, eval a...

Posts

Wikitag Contributions

Comments