Reflections On A Day of AI Governance Research

Researching for the first time

A dork researching some AI

My first hurdle was finding the right door, turns out the AI Governance Research Sprint was not being held in a yoga studio, which in retrospect makes a lot of sense.

My second challenge was walking 5 flights of stairs on a hot summer day only to discover there was an elevator. From here, the day got more successful.

The day started with a warm welcome from Belle, who organised this wonderful event. I was one of the first research sprinters to arrive, so I casually chatted with the AI Governance heroes of Arcadia Impact and waited for more sprinters to join us.

Within 30 minutes we had a room of about 20 enthusiastic sprinters ready to do a deep dive into some AI Governance.

I was very nervous when I arrived, as I wasn’t sure what I could contribute to the AI Governance space. My background is mainly as a database jockey, not a policy wonk, but I thought at the very least, I could fetch snacks.

Everyone I met turned out to be incredibly kind, enthusiastic and also pretty nervous in many cases, so that made me feel better. I made my way through the room learning how this group of smart folks had discovered the event and what they hoped to gain from it. Most did not have a research idea, one person had an extremely strong preference for researching how bioweapons may be used, she was very interesting to talk to, but I had my heart set on looking into something to do with microchips or transformative AI. These ideas were on my mind due to me doing the pre-reading on the tube, as I am never one to skip homework, but also I am one to never do it before the last minute.

Transformative AI was on my mind because it promises to change everything we know and the chips were on my mind, because I knew basically nothing about them and wanted to know more. I found a co-conspirator who also knew nothing about chips and also wanted to know more, so we formed a team and got to work.

When I joined the event, I imagined teams of 6 policy professionals all knowing exactly what to do and one me waiting to be told what to do. Being in a team of 2 meant I really had to pull my weight and also meant it was much easier to have my voice heard, so this turned out to be a much better structure, albeit more difficult.

The first few hours my teammate David and I were reading about microchips, their supply chain and the international tensions around them. Arcadia Impact gave us some pre-reading materials including previous papers that had been done on the work and even linked us to their own course material, making getting up to speed much easier. Now that we were verified experts (kidding), we had to figure out what question we wanted to answer, there were 4 more hours to go to find a problem, solve that problem, present the problem and provide a nice summary of the problem in under 400 words. We had a bunch of vague ideas around determining what was going on with AI chip security, but we weren’t sure what this meant or what the question would be.

Eventually, I suggested we figure out the amount of resources being spent on AI chip security, so money and people. This turned out to be a really hard question to answer, we were really struggling to find sources that offered this information. Our resident project helper Callum came to ask us about our project. I was a little embarrassed to say we had a question but couldn’t answer it very well. But Callum had an idea that if we are struggling to find some information, that could mean that it is a question that needs further research! So our work product became a research proposal to find out how many resources are currently being done to enable AI chip security.

This insight made our research mission much more targeted, now we just needed to explain why we thought it was an important question, what we had found so far, then how we suggested this be done.

We did manage to pull together a presentation and 400i-sh word summary and present a sensible idea.

When presentation time came, I was pretty mentally exhausted, yet satisfied. I haven’t dedicated 6 hours to thinking, learning and researching one specific topic since university. It was a really gratifying experience.

Whilst waiting for the presentations, I got chatting to the judges, whilst they were accomplished, they were so friendly, welcoming and genuinely curious in what we were doing that my nerves were put at ease.

Then the presentations started.

My co-sprinters had achieved so much in 6 hours!

One team had managed to prompt AI to make bioweapons, compare them to real scientific methods and determine that it was a dangerously specific level of instruction, that should be better controlled. There was a team that mapped the timeline and effectiveness of advocacy in AI safety and ethics. A team mapped out how transformative AI might impact us and another team came up with actionable advice on how to improve a UN AI organisation. Really impressive stuff and we all ended up working on completely different topics, so I learned so much.

Once the presentations were done, we had some dinner and discussed how the projects went and got to know each other some more. I met some wonderful people, learned so much and worked much harder than I have for a long time. I think my main takeaway was that a group of sufficiently motivated people can achieve so much more than seems expected in 6 hours.

Thank you so much to Arcadia Impact for organising and facilitating the amazing event.