Why I’m interested in AI safety

Most people didn’t know it was coming. Some knew but didn’t consider it urgent. Few anticipated how rapidly things would accelerate. I was in that second group—aware but not alarmed. Curiosity, and a compulsive habit of watching every recommended video, led me to Rob Miles. Rob Miles makes videos about AI safety. AI safety isn’t about fear. It’s about understanding and steering. These powerful systems we’re building could outpace our ability to control them.

I don’t remember exactly which video I watched first, but I remember the one that stuck with me: mesa-optimizers. The concept had an immediate intellectual appeal. Here was a problem at the intersection of computer science, philosophy, and society, with relatively few researchers working on it. But I didn’t jump in then. I was pursuing a biology degree in France, followed by a Master’s in bioinformatics. I didn’t see a clear path to AI work—I was merely an interested observer. Even as I learned about machine learning and deep learning during my Master’s, I didn’t make the connection back to AI safety. It felt too distant, too theoretical. It was only with the release of ChatGPT that things shifted. Here was a powerful AI model that was actually disrupting people’s livelihoods. This time, I jumped.

We’re building systems we don’t fully understand. That’s the core problem. Many researchers are working on interpretability—trying to open the black box. Back during my Master’s, I was already pestering people at my lab and at parties about how AI models aren’t explainable, so it made sense that I’d explore this area first. Related to it is the alignment problem: ensuring these systems do what we actually want. Today’s alignment failures aren’t just technical glitches—they’re previews of tomorrow’s catastrophic risks. Gradual disempowerment is a concrete scenario that illustrates the stakes.

I hear the rebuttals. “We could always shut them down”—but that ignores deployment realities and economic incentives. “They’re just machines”—but that misunderstands emergent capabilities. “I prefer being an optimist”—but optimism without action is just hope.

I believe technical solutions are necessary, though not sufficient. We can’t control what we don’t understand, so before steering, we need understanding. Interpretability approaches—understanding internal representations—are one path I could take. I’m also interested in questions around eliciting latent knowledge and hierarchical agency, though I haven’t read enough about these topics to commit yet.

The challenge, for me, is that AI safety requires deep technical foundations I’m still building. It’s one reason I emerged from a long educational and professional hiatus to study data science, and then AI. I’m currently enrolled in a bachelor’s program and planning to pursue graduate studies in AI. My goal is to help people understand and forecast AI behavior. I’ve been influenced by diverse voices and organizations pushing different approaches, and I recognize that the field needs multiple approaches to succeed.

We’re at a critical juncture. Powerful AI systems are here; more powerful ones are coming. This isn’t about pessimism—it’s about responsibility. Working in this area promises to be both technically fascinating and genuinely important. I’m committed to ensuring AI systems remain beneficial as they grow more capable, and I’ll document my journey as I go.