Why I'm starting this

There's too much going on right now. I read papers, watch videos, look at what startups are shipping, scroll LinkedIn and Twitter and see people posting things I feel like I should already know about. It's a lot. I get overwhelmed by how much there is to catch up on.

Reading all of it gave me ideas about everything, which is sort of the problem. But over time, working on research with a friend, I found something I actually want to focus on.

At first I thought AI safety wasn't that important. I genuinely thought it was about saving people, like physical safety. Sounds kind of stupid now. Then I read more and realized it's basically everything I was already looking at anyway. The benchmarks are that. Alignment is that. Evaluation is that. Once I saw it that way it got a lot more interesting.

Where I'm coming from

In school I did research on AI for Search and Rescue. We trained models on historical data to predict what situation a missing person might be in, and where you might find them. That kind of thing. It was interesting research.

What interests me more now comes from using these models constantly. I know where they're bad. I've seen the Anthropic alignment work, so I know agents can go rogue. I haven't run into it myself, but I understand it's a real problem and I want to dig into it.

I've been building things with AI, mostly workflows that automate what people were doing by hand. That's what most AI startups are, and I've been experimenting with the same thing.

Some of that was with Livin, a product I worked on that helps college students find housing and roommates. It had an AI search feature where you could describe what you wanted and get listings back that matched. The UI is old now. I was looking at replacing it with a voice agent you just talk to, and it searches for you.

Building that is where it clicked for me. Whatever agent we built had a lot of issues, and I tried vibe coding the fixes, and that didn't really work. I didn't know why the agent was behaving the way it was, so I was just hoping a prompt would land on the fix. That's when alignment stopped being abstract. The research behind the problem turned out to be more interesting than the product I was trying to ship.

So I've been going that direction. I joined a causal AI research group. I've been learning about structural causal models, POMDPs, causal inference. I'm reading The Book of Why by Judea Pearl. Reading papers. Reading a lot.

Why I'm writing this

Mostly to track what I'm doing. If someone else finds it interesting, good, but I don't really expect that. The main thing is having a place to put the progress down so I can see it.

I've spent almost all my time on the engineering side. CS major at a polytechnic school, where you mostly learn by building things, which is why I got into startups and product work in the first place. Now I want the theory and the research part I skipped.

So this will be whatever's going on. Research progress, what I learned, papers I found interesting.

I've also been thinking about doing YouTube shorts where I read a paper in fifteen minutes and then give a one minute presentation on it. The point is to force myself to go fast instead of sinking too much time into any one thing. That would be for casual papers, not the ones I'm actually researching.

I don't know the cadence yet. Maybe every day, maybe whenever something interesting happens. I'm just trying to start.

If you see something you think I'd find interesting, reach out. There's no comment section yet, so email works.

I've been trying to lean on AI a lot while keeping my brain alive. We'll see how that goes. Hopefully I keep this thing alive too.