Xander Davies, a graduate of Computer Science. Originally from Washington, DC, Xander is passionate about AI Safety and is the founder of the HAIST (Harvard AI Safety Team). He previously led the REMIX interpretability residency at Redwood Research and conducted research on the generalization behavior of neural networks at the University of Cambridge. Post graduation, he is continuing his work on alignment and interpretability at Redwood research. 

HTR:

What does being an innovator mean to you?

Xander:

When I got to Harvard, I had a sort of reverence for the real world. I looked around and thought, wow, there are all these brilliant, driven people working on all the major challenges, taking advantage of all the biggest opportunities. But being an innovator, in my opinion, means second guessing that assumption. It means acknowledging that, actually, there are many big problems and opportunities in the world that people aren’t fully addressing, and that it can be on you to recognize some of those opportunities and to work towards solving some of those open problems.

HTR:

As an innovator, is there someone that has inspired you the most? 

Xander:

There’s some recency bias at play, but Geoffrey Hinton, the 2018 Turing Award winner for his work on deep learning and often considered the godfather of AI, has been really inspiring to me. He recently quit his job at Google to openly discuss the risks of AI. It’s often emotional and difficult as an innovator to confront the negative side effects of your work, and I find his story very inspiring.

HTR:

In just one year, you started HAIST, one of the largest student organizations at Harvard. What inspired you to take on the role of starting this organization? And what are some key challenges and successes you’ve faced throughout the process?

Xander:

As a sophomore, I began doing machine learning research. During this time, and especially during my junior year, I began encountering arguments that the development of increasingly powerful AI systems might pose serious risks to humanity.  Wanting to engage with this properly, I decided to start a reading group on relevant research papers. We’d meet up once a week, I’d print out readings for everyone, bring some food, and we would just read for an hour in silence and then discuss what we’d read for an hour. It was a rewarding and fun format, and I would have been happy to do it even if no one else showed up.

Over the summer before senior year, I did AI safety research with David Kruger’s lab at Cambridge University. During that time, I engaged more with the AI safety research field and community, and I came back excited about transforming the reading group into a stronger  club. With encouragement and help from many great people, that reading group became HAIST.

Harvard does an amazing job gathering a lot of talented people who want to make a difference. I think exposing these people to important problems, especially those which aren’t being discussed enough, can really make a difference. 

HTR:

Were there any challenges along the way? And how would you describe the success of the organization today?

Xander:

Overall, I’ve been really happy with HAIST’s success. It coincided nicely with AI having its moment, which was lucky. I’ve been very inspired by the magnitude of interest in AI safety from the student body. We’d send out emails about AI safety opportunities and get hundreds of applications, which dramatically exceeded our expectations. I think this is a testament to Harvard students being interested in ways to combine technical depth with societal impact, which I think is perfect for AI safety work.

I’m proud that we’ve succeeded in creating a research-oriented vibe at HAIST, where we’re a serious place for people to think about hard problems and work together to solve them. Of course, it’s challenging to run a big group; there are many decisions which feel both very unclear and very high-stakes. But ultimately you just have to make a decision, and sometimes you have to pretend that you’re confident about that decision!

HTR:

Fantastic. Let’s delve deeper into your work in AI research, particularly in the areas of interpretability and safety research. Are there any experiences or highlights that stand out to you? 

Xander:

Interpretability is a research field aimed at understanding how AI systems perform as well as they do, by turning the computation performed by AI systems into human-understandable concepts. 

It’s counterintuitive that there would need to be a whole research field with the goal of understanding systems we ourselves built! But the way we build these AI systems is by automatically shifting around billions of ‘weights’, which you can think of as little knobs in a huge machine, to get better at some task. This process works really well, but leaves us with billions of numbers and no sense of, at a conceptual level, what those numbers mean.

It’s like if I had a long list of all the synapses in your brain, and then you asked me what you were thinking about—I’d have no clue how to use that low-level information to answer a higher-level conceptual question.

It’s really alarming to me that we are building incredibly capable systems, where even the builders of those systems have no clue what’s going on inside. I think interpretability is a very exciting field! 

HTR:

What keeps you motivated to tackle this issue? How do you maintain a sense of purpose and direction towards this each day?

Xander:

I feel privileged that AI safety, as a field, is both fascinating to work on and simultaneously perhaps one of the most important challenges humanity currently faces.. It makes it easy to stay driven.

HTR:

What hobbies or interests outside of your work have unexpectedly benefited you during your time at Harvard?

Xander:

I was a Harvard writing center tutor for a while, and I really enjoy writing. Initially, that felt disconnected from the AI work I was doing, but being comfortable writing has actually been very valuable.

HTR:

Which classes have you found most valuable?

Xander:

I found my Expos class, “The Art of Shock,” incredibly useful for improving my writing. I also thoroughly enjoyed MIT’s 6.840 (Theory of Computation), which I thought was beautiful and very well-taught. This semester, I took a course on interpretability, which was naturally aligned with my interests. Those would be my top three picks.

HTR:

Thinking towards the future, who do you want to become when you grow up?

Xander:

Who do I want to be? Work-wise, I want to be driven, thoughtful, and focused on improving the world.

HTR:

What are your immediate plans post-graduation and how do you see yourself continuing to work in the fields that currently interest you?

Xander:

Right after graduation, I’ll likely be spending a few months as a resident at Redwood Research, an AI safety non-profit. As I mentioned before, I’m passionate about AI safety due to its combination of importance and personal interest, and I’m looking forward to spending more time working in this area.

HTR:

Looking into the next 10-20 years, what makes you most optimistic about innovation in general, particularly within the space of forward-thinking technologies? Conversely, what are some elements that continue to worry you?

Xander:

I’ll start with my concerns. Innovation often brings with it a drive to “move fast and break things,” to make something cool and useful as quickly as possible. Often, this attitude is inspiring and productive. Sometimes, though, that attitude leads to recklessness and short-sightedness, like in the case of Thomas Midgley Jr., who was a renowned innovator now known for inventing leaded gasoline and kicking off the commercial use of chlorofluorocarbons.

I think the “move fast and break things” mentality is particularly worrying when it comes to AI development, where accidents could be extinction-level events. Building super-human intelligence requires a caution and rigor which is somewhat at odds with the usual spirit of innovation. It’s important to figure out how to preserve an innovative culture around AI development and safety research, while maintaining the proper level of caution.

On the other hand, I am optimistic that AI will be a powerful tool for innovators, making it cheaper and more accessible for small groups to create really powerful things. I’m especially excited about using AI to reduce the cost and improve the quality of health-care. Eventually, AI might be able to make us much more economically productive. If done safely and equitably, that’s a very exciting prospect.

HTR:

If you could travel back to your freshman year, what advice would you give your younger self about navigating Harvard and finding a purpose?

Xander:

My advice would be to appreciate the value of extracurricular activities and be mindful not to invest too much time in classes you don’t enjoy. It’s really important to decide whether you’ll take a class seriously or not. If you choose not to take a class seriously, then really don’t devote excessive time to that class. It’s worth engaging deeply in some classes, and I’ve thoroughly enjoyed a few. But investing too much time in school might mean missing out on some really valuable extracurriculars.