Dear reader, welcome to the AI Safety Cape Town newsletter!
I’m Leo Hyams, co-founder of AI Safety Cape Town (AISCT). I’ve been building AISCT over the past year alongside my co-founder Benjamin Sturgeon, and we are really excited about it. In this post, I aim to communicate what our organisation is about, what we have been up to, and where we intend to go.
Why are we doing this?
We believe that ensuring that powerful AI systems are beneficial to humanity, and to life in general, is the most important problem of our time. When you stop and think about it, this moment in human history is truly bizarre. We are at the dawn of a new era; we are potentially creating a new species that may surpass us. We have whistled past the Turing test and a dozen other grand challenges in AI (and some in science more generally), without as much as a pause to consider what we are doing. Not only are the philosophical implications of AI profound, but we also find ourselves in a precarious situation. Companies like OpenAI, Meta, and Google Deepmind are racing ahead with frontier AI development while a considerable portion of AI researchers consider powerful AI systems to be potentially catastrophic for humanity. We think that bringing people together to consider the implications of this monumental shift for our species is a spectacular use of our time. Moreso, we think that South Africa has especially awesome and underrecognized talent that could contribute to AI Safety. Our mission is to help such people (perhaps you reading this!) orientate towards an impactful career in this field.
We now have a fast-growing community of around 100 members, a small research lab, and some awesome international partnerships, such as the European Network for AI Safety and Apart Labs. However, I think that we have only explored a fraction of the potential network in Cape Town, let alone South Africa.
What have we done in the past?
We recently had significant involvement in the Deep Learning IndabaX 2024 held at the University of Witwatersrand, where Ben, Myself, and Ahmed (a prominent member of our community) ran an introductory workshop on Technical AI Safety. The workshop had great engagement, and if you were there we thank you very much for coming!
I also facilitated a panel discussion titled “AGI Horizons: Utopia, Dystopia, or Something In-Between”, which featured Ben, Lydia De Lange (Executive Director of the Deep Learning IndabaX South Africa), Professor Benjamin Rosman (Head of RAIL Labs and a top AI professor in Africa), and Joshua Maumela (Senior Machine Learning Engineer at Vodacom). The discussion, as suggested by the title, was wide-ranging and generally served as a litmus test on leading professionals' attitude and thoughts about the future of the field.
Earlier in the year, we ran an AI Safety Fundamentals Fellowship in collaboration with AI Safety Sweden. We had over 200 applicants from African countries, accepted the top 100, and had 75 people complete the program. The program included two tracks, one AI Governance in Africa, and one on Technical Alignment. I thoroughly enjoyed facilitating this program, and we got great feedback from our participants.
What are we doing currently?
Ben is currently leading our research lab, where we are evaluating LLM behaviours which may impact human agency. In basic terms, this looks like identifying traits which may be harmful to human agency (such as persuasiveness) and identifying the degree to which the trait is present in a given frontier LLM (such as Claude 3.5-sonnet). Alongside myself and Ben, our research team consists of Daniel Samuelson, an ML engineer and entrepreneur, Ethan Forster, an AI Ethics PhD candidate, and Jacy Reece Anthis, founder of the Sentience Institute and a visiting scholar at the Center for Human-Compatible AI, who is acting as our research supervisor. We are also proud to say that our lab is being incubated by Apart Labs, with Jacob Haimes joining us to advise on the project.
We are also running a regular biweekly reading group for the core members of our community. These are a great way to advance your knowledge of technical AI safety and to get more involved in AI safety research. The subject of the reading group is generally a technical paper, and we have a lot of fun teasing out the details and thinking about the implications for the future of the field. Notably, a couple of weeks ago we discussed Towards Guaranteed Safe AI (TGSAI): A Framework for Ensuring Robust and Reliable AI Systems.
What we plan to do in the future.
We have some exciting collaborations on the horizon, such as running another round of AISF fellowships. We are excited to offer this course to our community, as it is a great way to get established in AI Safety, figure out what the field is really about, where it is likely to go, and how to get involved.
We are also excited to announce that we will be collaborating with Condor Camp to host a retreat near the end of this year. This will be a week-long intensive dive into technical AI safety and AI governance aimed at gifted young students or early career professionals who might want a career in AI Safety and will feature close mentorship from top experts in the field. It will be an all expense paid trip to a great location in Cape Town, where participants will be able to really get a taste for the field and how to orient themselves within it. If either or both of these programs interest you, you can fill out the interest form attached here.
Concluding
Thanks for reading our newsletter, hopefully by now you know a little more about what we are about and what we intend to do. Please drop me an email if you have any feedback, comments, or suggestions for our organisation or for future newsletters.
I hope you enjoyed reading this first edition, and that you’ll stick around for future posts like this!
Kind regards,
Leo Hyams
