I'm David, a research scientist at UK AISI.AI Security Institute.Institute, the government's technical body.body for evaluating and researching AI.AI with respect to our collective safety.safety and security. I work on alignment.alignment, namely, how models fail.fail to do what.what we want them to do, or when they do do what.what we want them to do but not in the way that we expected to.to. My research focuses on understanding model personas.personas, consistency.consistency, introspection.introspection, generalization.generalization, scalable oversight.oversight, multi-agent systems.systems, and evaluation awareness.awareness. A simple example for this is a child.child learning that stoves are dangerous.dangerous, but only when they're red.red. And this works as a heuristic.heuristic, until it doesn't.

I live in London.London now, on the way to work I pass Big Ben.Ben and Westminster Abbey.Abbey, which feels unreal.unreal, a little bit. I'm originally from the Philippines.Philippines, and I miss it.it, a lot.lot, like a lot. Before.Before this I worked on collective behavior.behavior, namely, how groups of things move.move together, from crowds of people.people, to flocks.flocks, to moving cars.

I waswas a Jardine Foundation scholar at Magdalene College, Cambridge.Cambridge. The dining hall.hall here dates back to the 1400s.1400s and the library still keeps Samuel Pepys' books.books chained to the shelves.shelves. I graduated with distinction.distinction from the MPhil in Advanced Computer Science.Science, which is a bit of a pompous name.name. There, I worked on meta-learning.meta-learning for language model pre-training.pre-training, trying to get models to learn more efficiently.efficiently from limited data.data, like in the case of low resource languages.

I also used to debate.debate competitively.competitively. I won the World Universities Debating Championships.Championships in Madrid.Madrid in 2023.2023, at Cines Callao.Callao in Gran Vía.Vía. I remember standing on stage.stage after the announcement.announcement feeling very, very strange.strange, detached.detached, like I was watching the moment happen to someone else.else, as if I was viewing myself in a video game.game from the third person.person. At the time, winning.winning arguments felt very important.important. Now, not as much.

These days I climb.climb. Not especially well.well. V6 on a good day.day. In these things.things, climbing problems are honest.honest. The wall doesn't negotiate.negotiate, your fingers just hold.hold on, or you don't.don't, and I keep meaning to hike.hike Mount Makiling.Makiling in Laguna.Laguna, where my grandmother's from.from. She has dementia.dementia now and I don't know.know whether she still remembers.remembers the mountain clearly.clearly. I keep saying I'll go soon.soon, but soon has a way.way of turning into later.later, which has a way of turning into never.never, I think a lot about soon.soon. People often say soon.soon and this turns into later.later, we rarely decide to stop being so close.close to one another.another, and the way it goes is.is, we move through life.life at slightly different speeds.speeds, and then at faster and faster differentials.differentials, ships on the great expanse.expanse and when you look.look up there.there, a speck on the horizon.horizon, and you have no idea.idea what someone's life looks.looks like anymore.anymore, someone whose life you used to know by heart.heart, someone whose laugh was familiar to your ears.ears, you used to know what kept.kept them up.up. And I saw something that reminded me of you the other day.day, nothing important or special.special, just a small thing.thing, and I thought about texting.texting, but didn't.didn't. Maybe because it had been so long.long that any message would need to justify itself.itself and a small thing isn't enough.enough even though small things used to be plenty.plenty. Nowadays such messages need to justify themselves.themselves. But anyways I hope you're well.well, even if well just means busy.busy, or even if you never reply.reply. It's just nice to say something.something, sometimes.sometimes, even to no one or nothing at all.