Ryan Lagasse

I care about understanding AI systems mechanistically - not just observing that they work, but knowing why. I'm especially drawn to approaches that scale, like autointerpretability: understanding models from the inside without a human in the loop for every feature. My theory of change is that we need negative alignment taxes - safety properties that make a system more capable, not less, so labs adopt them because they work rather than out of goodwill. I believe progress there is one of the most direct paths to AI systems we can actually trust. My north star is to help AGI go well.