Marcus Williams

Meet Marcus

Marcus completed his Master’s degree in engineering physics and machine learning (ML) at Lund University in 2023. His initial interest in ML was driven by the goal of curing aging, recognizing its potential to extend the human lifespan by thousands of years. However, he soon shifted to focus on AI alignment, believing that Artificial Super Intelligence (ASI) was imminent and could either lead to human extinction or solve aging along with other global challenges.

Post-graduation, Marcus collaborated on a theoretical reinforcement learning (RL) project at the AI Safety Hub Oxford. The team's paper, "On the Expressivity of Objective-Specification Formalisms in Reinforcement Learning," was accepted at the International Conference on Learning Representations (ICLR) 2024. Subsequently, he received a Long-Term Future Fund (LTFF) grant for a six-month independent project titled "Multi-Objective RL from AI Feedback". In 2024, Marcus enrolled in MATS, where he is working with Micah Carroll on “Emergence of harmful behavior during RL on user feedback in LLMs” a project investigating how sycophancy, deception, and manipulation can arise naturally when optimizing for user satisfaction.

Interactions with EA and how EA has affected Marcus’ career

While he may not have been familiar with the term initially, the core EA principle of maximizing expected utility significantly influenced Marcus decision to work on curing aging and later, AI alignment. He valued finding EA Lund and being able to discuss such topics with like-minded people, as well as receiving support and input on grant applications from EA Sweden.

Challenges with a non-conventional career path

Marcus decided to not apply for any “conventional” employment after university, and decided to work unpaid on the AI safety project at AI Safety Hub Oxford. This meant a lot of uncertainty about the future and a lot of people telling him this was a terrible mistake. In addition, Marcus worked on his Multi-Objective RL project alone for 6 months without any collaborators or mentors, which turned out to be quite isolating and less productive than if he’d found help or support.

Why Marcus thinks AI alignment is important to work on

Marcus believes ASI is on the horizon, carrying both the risk of human extinction and the potential to resolve all global issues, ushering in an era of unprecedented well-being for all life. While the challenge of alignment is significant, he thinks its difficulty has been somewhat overstated and is more tractable than commonly perceived.

Marcus’ tips for people interested in a career in AI Safety research

1. Engage in direct research: The entry barrier for contributing to the field is lower than you might expect. I believe a dedicated and reasonably competent individual can reach the research frontier within a year.
2. It’s easier to find opportunities if you’re willing to work unpaid. While the bar for getting accepted to paid programs such as MATS, ARENA and LASR is very high, the bar for doing unpaid alternatives such as SPAR or AI Safety Camp is much lower. If possible, consider doing unpaid programs to gain experience.
3. Find collaborators and mentors. You will be much more productive and motivated if you can find good collaborators and mentors.
4. Don’t let university get in the way of your education. In my opinion, Swedish universities are not doing a good job, especially in the area of ML. The things I learnt were highly out of date and most of the professors did a bad job teaching. That being said, I still think university could be a good idea (if you get CSN) as it gives you a stable income for many years, allowing you to learn and do research projects on the side without the uncertainty and hassle of applying to jobs/grants.
5. Work that you enjoy doesn’t feel like work. I always thought he hated work. A few years ago, I quit a software development internship after a week as it was unbearable. Something which MATS taught me was that if your work is exciting/interesting, feels impactful and if you have great collaborators, then working can be great. Now I work 70 hours a week but am the happiest I have ever been.
6. Don’t have your 80k career advising call too early. While it’s easier to have multiple calls now, I felt like he had my career advising call too early, when I barely knew which alignment research directions I was most interested in. If I had it later, I think I would have gotten more out of it.