ML4Good runs intensive in-person bootcamps intended to inspire and technically upskill AI safety hopefuls. The bootcamps are held globally, from the UK, France, and Germany, to as far as Brazil. There are aspirations to reach many more corners of the world, like India, the US, and the Philippines. They are free to attend with recent editions receiving funding from Open Philanthropy.
I attended their Germany 2024 camp. As someone who is new to AI safety—with a latent interest, but a need to expand my network, build technical skills and discover the landscape of opportunities—the experience was invaluable. I would highly recommend it to others in similar positions.
The course is intensive: there really is no weekend. Whilst the content is introductory, it is ambitious. The time-spent is split roughly fifty-fifty between technical and non-technical content.
The technical lectures cover transformers, reinforcement learning, RLHF, introductory mech interp, evals, and more. Coding exercises include the terrifying sounding “implement GPT-2 from scratch”. For those familiar, the technical content is based on a curated subset of ARENA.
The technical content was delivered well. I left the transformer talk sure that I could explain the concept to a maths undergraduate, and perhaps even a GCSE student. My recurring message in the endless feedback forms was that (1. There were too many feedback forms, but 2.) I wanted more of the technical content.
The non-technical content included timelines forecasting, scalable oversight, and governance methods. These were predominantly run as discussion groups.
The most unexpectedly valuable thing I got out of the camp was the opportunity to do focused self-reflection in a motivating atmosphere. A close second was the friendships made and diverse conversations had. Among the social highlights were plunging into an ice-cold lake and taking on a 2100-rated chess player in a two-against-one match. The final day was actually quite emotional.
The truth is you’re not going to become an expert in technical AI safety, or frankly even remotely competent, with just 10 days to learn. You might however ignite an internal spark that will lead you on that journey to expertise. You might leave with many ideas for where to go next, better knowledge of where your interests lie, and a strong accountability network made of friendly faces.
Reflections on ML4Good
ML4Good runs intensive in-person bootcamps intended to inspire and technically upskill AI safety hopefuls. The bootcamps are held globally, from the UK, France, and Germany, to as far as Brazil. There are aspirations to reach many more corners of the world, like India, the US, and the Philippines. They are free to attend with recent editions receiving funding from Open Philanthropy.
I attended their Germany 2024 camp. As someone who is new to AI safety—with a latent interest, but a need to expand my network, build technical skills and discover the landscape of opportunities—the experience was invaluable. I would highly recommend it to others in similar positions.
The course is intensive: there really is no weekend. Whilst the content is introductory, it is ambitious. The time-spent is split roughly fifty-fifty between technical and non-technical content.
The technical lectures cover transformers, reinforcement learning, RLHF, introductory mech interp, evals, and more. Coding exercises include the terrifying sounding “implement GPT-2 from scratch”. For those familiar, the technical content is based on a curated subset of ARENA.
The technical content was delivered well. I left the transformer talk sure that I could explain the concept to a maths undergraduate, and perhaps even a GCSE student. My recurring message in the endless feedback forms was that (1. There were too many feedback forms, but 2.) I wanted more of the technical content.
The non-technical content included timelines forecasting, scalable oversight, and governance methods. These were predominantly run as discussion groups.
The most unexpectedly valuable thing I got out of the camp was the opportunity to do focused self-reflection in a motivating atmosphere. A close second was the friendships made and diverse conversations had. Among the social highlights were plunging into an ice-cold lake and taking on a 2100-rated chess player in a two-against-one match. The final day was actually quite emotional.
The truth is you’re not going to become an expert in technical AI safety, or frankly even remotely competent, with just 10 days to learn. You might however ignite an internal spark that will lead you on that journey to expertise. You might leave with many ideas for where to go next, better knowledge of where your interests lie, and a strong accountability network made of friendly faces.